An Overview of Deep Learning in Medical Imaging Focusing On MRI

REVIEW
An overview of deep learning in medical imaging focusing on

MRI
Alexander Selvikvåg Lundervold a,b,∗ , Arvid Lundervold a,c,d
a
Mohn Medical Imaging and Visualization Centre (MMIV), Haukeland University Hospital, Norway
b
Department of Computing, Mathematics and Physics, Western Norway University of Applied Sciences, Norway
c
Neuroinformatics and Image Analysis Laboratory, Department of Biomedicine, University of Bergen, Norway
d
Department of Health and Functioning, Western Norway University of Applied Sciences, Norway
Received 2 October 2018; accepted 21 November 2018
Abstract
What has happened in machine learning lately, and what does it mean for the future of medical image analysis? Machine
learning has witnessed a tremendous amount of attention over the last few years. The current boom started around 2009
when so-called deep artificial neural networks began outperforming other established models on a number of important
benchmarks. Deep neural networks are now the state-of-the-art machine learning models across a variety of areas, from
image analysis to natural language processing, and widely deployed in academia and industry. These developments have
a huge potential for medical imaging technology, medical data analysis, medical diagnostics and healthcare in general,
slowly being realized. We provide a short overview of recent advances and some associated challenges in machine learning
applied to medical image processing and image analysis. As this has become a very broad and fast expanding field we will
not survey the entire landscape of applications, but put particular focus on deep learning in MRI.
Our aim is threefold: (i) give a brief introduction to deep learning with pointers to core references; (ii) indicate how deep
learning has been applied to the entire MRI processing chain, from acquisition to image retrieval, from segmentation to
disease prediction; (iii) provide a starting point for people interested in experimenting and perhaps contributing to the field
of deep learning for medical imaging by pointing out good educational resources, state-of-the-art open-source code, and
interesting sources of data and problems related medical imaging.
Keywords: Machine learning, Deep learning, Medical imaging, MRI
1 Introduction
user-friendly software frameworks, and an explosion of the
Machine learning has seen some dramatic developments available compute power, enabling the use of neural networks
recently, leading to a lot of interest from industry, academia that are deeper than ever before. These models nowadays form
and popular culture. These are driven by breakthroughs in the state-of-the-art approach to a wide variety of problems in
artificial neural networks, often termed deep learning, a set computer vision, language modeling and robotics.
of techniques and algorithms that enable computers to dis- Deep learning rose to its prominent position in computer
cover complicated patterns in large data sets. Feeding the vision when neural networks started outperforming other
breakthroughs is the increased access to data (“big data”), methods on several high-profile image analysis benchmarks.
∗ Corresponding author: Alexander Selvikvåg Lundervold, Mohn Medical Imaging and Visualization Centre (MMIV), Haukeland University Hospital,
Norway.
E-mail addresses: [email protected] (A.S. Lundervold),
[email protected] (A. Lundervold).
Z Med Phys 29 (2019) 102–127

https://doi.org/10.1016/j.zemedi.2018.11.002
www.elsevier.com/locate/zemedi
A.S. Lundervold, A. Lundervold / Z Med Phys 29 (2019) 102–127 103
Most famously on the ImageNet Large-Scale Visual Recog- The use of machine learning in general and deep learning in
nition Challenge (ILSVRC) in 2012 [1] when a deep learning particular within healthcare is still in its infancy, but there are
model (a convolutional neural network) halved the second best several strong initiatives across academia, and multiple large
error rate on the image classification task. Enabling comput- companies are pursuing healthcare projects based on machine
ers to recognize objects in natural images was until recently learning. Not only medical technology companies, but also for
thought to be a very difficult task, but by now convolutional example Google Brain [28–30],4 DeepMind [31],5 Microsoft
neural networks have surpassed even human performance on [32,33]6 and IBM [34].7 There is also a plethora of small and
the ILSVRC, and reached a level where the ILSVRC classi- medium-sized businesses in the field.8
fication task is essentially solved (i.e. with error rate close to
the Bayes rate). Deep learning techniques have become the de 2 Machine learning, artificial neural
facto standard for a wide variety of computer vision problems. networks, deep learning
They are, however, not limited to image processing and analy-
sis but are outperforming other approaches in areas like natural In machine learning one develops and studies methods that
language processing [2–4], speech recognition and synthesis give computers the ability to solve problems by learning from
[5,6],1 and in the analysis of unstructured, tabular-type data experiences. The goal is to create mathematical models that
using entity embeddings [7,8].2 can be trained to produce useful outputs when fed input data.
The sudden progress and wide scope of deep learning, Machine learning models are provided experiences in the form
and the resulting surge of attention and multi-billion dollar of training data, and are tuned to produce accurate predictions
investment, has led to a virtuous cycle of improvements and for the training data by an optimization algorithm. The main
investments in the entire field of machine learning. It is now goal of the models are to be able to generalize their learned
one of the hottest areas of study world-wide [14], and people expertise, and deliver correct predictions for new, unseen data.
with competence in machine learning are highly sought-after A model’s generalization ability is typically estimated during
by both industry and academia.3 training using a separate data set, the validation set, and used
Healthcare providers generate and capture enormous as feedback for further tuning of the model. After several iter-
amounts of data containing extremely valuable signals and ations of training and tuning, the final model is evaluated on
information, at a pace far surpassing what “traditional” meth- a test set, used to simulate how the model will perform when
ods of analysis can process. Machine learning therefore faced with new, unseen data.
quickly enters the picture, as it is one of the best ways to There are several kinds of machine learning, loosely catego-
integrate, analyze and make predictions based on large, het- rized according to how the models utilize its input data during
erogeneous data sets (cf. health informatics [15]). Healthcare training. In reinforcement learning one constructs agents that
applications of deep learning range from one-dimensional learn from their environments through trial and error while
biosignal analysis [16] and the prediction of medical events, optimizing some objective function. A famous recent appli-
e.g. seizures [17] and cardiac arrests [18], to computer-aided cation of reinforcement learning is AlphaGo and AlphaZero
detection [19] and diagnosis [20] supporting clinical decision [35], the Go-playing machine learning systems developed by
making and survival analysis [21], to drug discovery [22] and DeepMind. In unsupervised learning the computer is tasked
as an aid in therapy selection and pharmacogenomics [23], to with uncovering patterns in the data without our guidance.
increased operational efficiency [24], stratified care delivery Clustering is a prime example. Most of today’s machine learn-
[25], and analysis of electronic health records [26,27]. ing systems belong to the class of supervised learning. Here,
the computer is given a set of already labeled or annotated
data, and asked to produce correct labels on new, previously
unseen data sets based on the rules discovered in the labeled
1 Try it out here: https://deepmind.com/blog/wavenet-generative-model-
data set. From a set of input-output examples, the whole
raw-audio.
model is trained to perform specific data-processing tasks.
2 As a perhaps unsurprising side-note, these modern deep learning methods Image annotation using human-labeled data, e.g. classify-
have also entered the field of physics. Among other things, they are tasked ing skin lesions according to malignancy [36] or discovering
with learning physics from raw data when no good mathematical models
are available. For example in the analysis of gravitational waves where deep
learning has been used for classification [9], anomaly detection [10] and
denoising [11], using methods that are highly transferable across domains 4 https://ai.google/research/teams/brain/healthcare-biosciences.
(think EEG and fMRI). They are also part of mathematical model and machine 5 https://deepmind.com/applied/deepmind-health/.
learning hybrids [12,13], formed to reduce computational costs by having the 6 https://www.microsoft.com/en-us/research/research-area/medical-health-
mathematical model train a machine learning model to perform its job, or to genomics.
improve the fit with observations in settings where the mathematical model 7 https://www.research.ibm.com/healthcare-and-life-sciences.
can’t incorporate all details (think noise). 8 Aidoc, Arterys, Ayasdi, Babylon Healthcare Services, BenevolentAI,
3 See e.g. https://economicgraph.linkedin.com/research/LinkedIns-2017- Enlitic, EnvoiAI, H2O, IDx, MaxQ AI, Mirada Medical, Viz.ai, Zebra Med-
US-Emerging-Jobs-Report for a study focused on the US job market. ical Vision, and many more.
104 A.S. Lundervold, A. Lundervold / Z Med Phys 29 (2019) 102–127
cardiovascular risk factors from retinal fundus photographs artificial neural networks form one of the dominant methods
[37], are two examples of the multitude of medical imaging in machine learning, and the most intensively studied. This
related problems attacked using supervised learning. change is thanks to the growth of big data, powerful pro-
Machine learning has a long history and is split into many cessors for parallel computations (in particular, GPUs), some
sub-fields, of which deep learning is the one currently receiv- important tweaks to the algorithms used to construct and train
ing the bulk of attention. the networks, and the development of easy-to-use software
There are many excellent, openly available overviews and frameworks. The surge of interest in ANNs leads to an incred-
surveys of deep learning. For short general introductions to ible pace of developments, which also drives other parts of
deep learning, see [38,39]. For an in-depth coverage, consult machine learning with it.
the freely available book [40].9 For a broad overview of deep The freely available books [40,49] are two of the many
learning applied to medical imaging, see [41]. We will only excellent sources to learn more about artificial neural net-
mention some bare essentials of the field, hoping that these works. We’ll only give a brief indication of how they
will serve as useful pointers to the areas that are currently the are constructed and trained. The basic form of artificial
most influential in medical imaging. neural networks,12 the feedforward neural networks, are
parametrized mathematical functions y = f(x; θ) that maps an
2.1 Artificial neural networks input x to an output y by feeding it through a number of
nonlinear transformations: f(x) = (fn ◦···◦ f1 )(x). Here each
Artificial neural networks (ANNs) is one of the most famous component fk , called a network layer, consists of a simple
machine learning models, introduced already in the 1950s, and linear transformation of the previous component’s output,
actively studied since [40,Chapter 1.2].10 followed by a nonlinear function: fk = σk (θkT fk−1 ). The non-
Roughly, a neural network consists of a number of con- linear functions σ k are typically sigmoid functions or ReLUs,
nected computational units, called neurons, arranged in layers. as discussed below, and the θ k are matrices of numbers, called
There’s an input layer where data enters the network, followed the model’s weights. During the training phase, the network is
by one or more hidden layers transforming the data as it flows fed training data and tasked with making predictions at the out-
through, before ending at an output layer that produces the put layer that match the known labels, each component of the
neural network’s predictions. The network is trained to output network producing an expedient representation of its input. It
useful predictions by identifying patterns in a set of labeled has to learn how to best utilize the intermediate representations
training data, fed through the network while the outputs are to form a complex hierarchical representation of the data, end-
compared with the actual labels by an objective function. Dur- ing in correct predictions at the output layer. Training a neural
ing training the network’s parameters – the strength of each network means changing its weights to optimize the outputs
neuron – is tuned until the patterns identified by the network of the network. This is done using an optimization algorithm,
result in good predictions for the training data. Once the pat- called gradient descent, on a function measuring the correct-
terns are learned, the network can be used to make predictions ness of the outputs, called a cost function or loss function.
on new, unseen data, i.e. generalize to new data. The basic ideas behind training neural networks are simple:
It has long been known that ANNs are very flexible, able as training data is fed through the network, compute the gra-
to model and solve complicated problems, but also that they dient of the loss function with respect to every weight using
are difficult and very computationally expensive to train.11 the chain rule, and reduce the loss by changing these weights
This has lowered their practical utility and led people to, until using gradient descent. But one quickly meets huge compu-
recently, focus on other machine learning models. But by now, tational challenges when faced with complicated networks
with thousands or millions of parameters and an exponential
number of paths between the nodes and the network output.
9 https://www.deeplearningbook.org/. The techniques designed to overcome these challenges gets
10 The loose connection between artificial neural networks and neural net- quite complicated. See [40,Chapter 8,50,Chapters 3 and 4]
works in the brain is often mentioned, but quite over-blown considering the for detailed descriptions of the techniques and practical issues
complexity of biological neural networks. However, there is some interesting involved in training neural networks.
recent work connecting neuroscience and artificial neural networks, indicat-
Artificial neural networks are often depicted as a network
ing an increase in the cross-fertilization between the two fields [42–44].
11 According to the famous universal approximation theorem for artificial of nodes, as in Fig. 1.13
neural networks [45–48], ANNs are mathematically able to approximate any
continuous function defined on compact subspaces of Rn , using finitely many
neurons. There are some restrictions on the activation functions, but these can
12 These are basic when compared to for example recurrent neural networks,
be relaxed (allowing for ReLUs for example) by restricting the function space.
This is an existence theorem and successfully training a neural network to whose architectures are more involved.
13 As we shall see, modern architectures are often significantly more compli-
approximate a given function is another matter entirely. However, the theorem
does suggest that neural networks are reasonable to study and develop fur- cated than captured by the illustration and equations above, with connections
ther, at least as an engineering endeavour aimed at realizing their theoretical between non-consecutive layers, input fed in also at later layers, multiple
powers. outputs, and much more.
Figure 1. Artificial neural networks are built from simple linear functions followed by nonlinearities. One of the simplest class of neural
network is the multilayer perceptron, or feedforward neural network, originating from the work of Rosenblatt in the 1950s [51]. It is based
on simple computational units, called neurons, organized in layers. Writing i for the ith layer and j for the jth unit of that layer, the output of
the jth unit at the ith layer is z(i) (i)T
j = θj x. Here x consists of the outputs from the previous layer after they are fed through a simple nonlinear
function called an activation function, typically a sigmoid function σ(z) = 1/(1 + e−z ) or a rectified linear unit ReLU(z) = max(0, z) or small
variations thereof. Each layer therefore computes a weighted sum of the all the outputs from the neurons in the previous layers, followed
by a nonlinearity. These are called the layer activations. Each layer activation is fed to the next layer in the network, which performs the
same calculation, until you reach the output layer, where the network’s predictions are produced. In the end, you obtain a hierarchical
representation of the input data, where the earlier features tend to be very general, getting increasingly specific towards the output. By
feeding the network training data, propagated through the layers, the network is trained to perform useful tasks. A training data point (or,
typically, a small batch of training points) is fed to the network, the outputs and local derivatives at each node are recorded, and the difference
between the output prediction and the true label is measured by an objective function, such as mean absolute error (L1), mean squared error
(L2), cross-entropy loss, or Dice loss, depending on the application. The derivative of the objective function with respect to the output is
calculated and used as a feedback signal. The discrepancy is propagated backwards through the network and all the weights are updated to
reduce the error. This is achieved using backward propagation [52–54], which calculates the gradient of the objective function with respect
to the weights in each node using the chain rule together with dynamic programming, and gradient descent [55], an optimization algorithm
tasked with improving the weights.
2.2 Deep learning In medical imaging the interest in deep learning is mostly
triggered by convolutional neural networks (CNNs) [56],14 a
Traditionally, machine learning models are trained to per- powerful way to learn useful representations of images and
form useful tasks based on manually designed features other structured data. Before it became possible to use CNNs
extracted from the raw data, or features learned by other efficiently, these features typically had to be engineered by
simple machine learning models. In deep learning, the com- hand, or created by less powerful machine learning models.
puters learn useful representations and features automatically, Once it became possible to use features learned directly from
directly from the raw data, bypassing this manual and difficult the data, many of the handcrafted image features were typ-
step. By far the most common models in deep learning are var- ically left by the wayside as they turned out to be almost
ious variants of artificial neural networks, but there are others. worthless compared to feature detectors found by CNNs.15
The main common characteristic of deep learning methods is There are some strong preferences embedded in CNNs based
their focus on feature learning: automatically learning repre-
sentations of data. This is the primary difference between deep
learning approaches and more “classical” machine learning. 14 Interestingly, CNNs was applied in medical image analysis already in the
Discovering features and performing a task is merged into
early 90s, e.g. [57], but with limited success.
one problem, and therefore both improved during the same 15 However, combining hand-engineered features with CNN features is a
training process. See [38,40] for general overviews of the very reasonable approach when low amounts of training data makes it difficult
field. to learn good features automatically.
on how they are constructed, which helps us understand why approximate almost any nonlinear function [47,48].18 The
they are so powerful. Let us therefore take a look at the building activation functions are generally the very simple recti-
blocks of CNNs. fied linear units, or ReLUs, defined as ReLU(z) = max(0,
z), or variants like leaky ReLUs or parametric ReLUs.19
See [59,60] for more information about these and other
2.3 Building blocks of convolutional neural networks activation functions. Feeding the feature maps through an
activation function produces new tensors, typically also
When applying neural networks to images one can in prin- called feature maps.
ciple use the simple feedforward neural networks discussed (iii) Pooling: Each feature map produced by feeding the data
above. However, having connections from all nodes of one through one or more convolutional layer is then typically
layer to all nodes in the next is extremely inefficient. A care- pooled in a pooling layer. Pooling operations take small
ful pruning of the connections based on domain knowledge, grid regions as input and produce single numbers for each
i.e. the structure of images, leads to much better performance. region. The number is usually computed by using the max
A CNN is a particular kind of artificial neural network aimed at function (max-pooling) or the average function (average
preserving spatial relationships in the data, with very few con- pooling). Since a small shift of the input image results in
nections between the layers. The input to a CNN is arranged small changes in the activation maps, the pooling layers
in a grid structure and then fed through layers that preserve gives the CNN some translational invariance.
these relationships, each layer operation operating on a small A different way of getting the downsampling effect
region of the previous layer (Fig. 2). CNNs are able to form of pooling is to use convolutions with increased stride
highly efficient representation of the input data,16 well-suited lengths. Removing the pooling layers simplifies the
for image-oriented tasks. A CNN has multiple layers of convo- network architecture without necessarily sacrificing per-
lutions and activations, often interspersed with pooling layers, formance [61].
and is trained using backpropagation and gradient descent as Other common elements in many modern CNNs include
for standard artificial neural networks. See Section 2.1. In (iv) Dropout regularization: A simple idea that gave a huge
addition, CNNs typically have fully connected layers at the boost in the performance of CNNs. By averaging several
end, which compute the final outputs.17 models in an ensemble one tend to get better performance
than when using single models. Dropout [62] is an aver-
aging technique based on stochastic sampling of neural
(i) Convolutional layers: In the convolutional layers the acti-
networks.20 By randomly removing neurons during train-
vations from the previous layers are convolved with a set
ing one ends up using slightly different networks for each
of small parameterized filters, frequently of size 3 × 3,
(j,i) batch of training data, and the weights of the trained
collected in a tensor W , where j is the filter number
network are tuned based on optimization of multiple vari-
and i is the layer number. By having each filter share the
ations of the network.21
exact same weights across the whole input domain, i.e.
(v) Batch normalization: These layers are typically placed
translational equivariance at each layer, one achieves a
after activation layers, producing normalized activation
drastic reduction in the number of weights that need to be
maps by subtracting the mean and dividing by the standard
learned. The motivation for this weight-sharing is that fea-
deviation for each training batch. Including batch normal-
tures appearing in one part of the image likely also appear
ization layers forces the network to periodically change
in other parts. If you have a filter capable of detecting
its activations to zero mean and unit standard deviation
horizontal lines, say, then it can be used to detect them
as the training batch hits these layers, which works as a
wherever they appear. Applying all the convolutional fil-
regularizer for the network, speeds up training, and makes
ters at all locations of the input to a convolutional layer
it less dependent on careful parameter initialization [66].
produces a tensor of feature maps.
(ii) Activation layer: The feature maps from a convolu-
tional layer are fed through nonlinear activation functions.
This makes it possible for the entire neural network to 18 A neural network with only linear activations would only be able to
perform linear approximation. Adding further layers wouldn’t improve its
expressiveness.
19 Other options include exponential linear units (ELUs), and the now rarely
used sigmoid or tanh activation functions.

16 It is interesting to compare this with the biological vision systems and 20 The idea of dropout is also used for other machine learning models, as in
their receptive fields of variable size (volumes in visual space) of neurons at the DART technique for regression trees [63].
different hierarchical levels. 21 In addition to increased model performance, dropout can also be used to
17 Lately, so-called fully-convolution CNNs have become popular, in which produce robust uncertainty measures in neural networks. By leaving dropout
average pooling across the whole input after the final activation layer replaces turned on also during inference one effectively performs variational inference
the fully-connected layers, significantly reducing the total number of weights [58,64,65]. This relates standard deep neural networks to Bayesian neural
in the network. networks, synthesized in the field of Bayesian deep learning.
Figure 2. Building blocks of a typical CNN. A slight modification of a figure in [58], courtesy of the author.
In the design of new and improved CNN architectures, these learning in medical imaging, but merely sketch some of the
components are combined in increasingly complicated and landscape before going into a more systematic exposition of
interconnected ways, or even replaced by other more conve- deep learning in MRI.
nient operations. When architecting a CNN for a particular Convolutional neural networks can be used for efficiency
task there are multiple factors to consider, including under- improvement in radiology practices through protocol determi-
standing the task to be solved and the requirements to be met, nation based on short-text classification [86]. They can also
figuring out how to best feed the data to the network, and be used to reduce the gadolinium dose in contrast-enhanced
optimally utilizing one’s budget for computation and mem- brain MRI by an order of magnitude [87] without significant
ory consumption. In the early days of modern deep learning reduction in image quality. Deep learning is applied in radio-
one tended to use very simple combinations of the building therapy [88], in PET-MRI attenuation correction [89,90], in
blocks, as in Lenet [56] and AlexNet [1]. Later network archi- radiomics [91,92] (see [93] for a review of radiomics related
tectures are much more complex, each generation building on to radiooncology and medical physics), and for theranostics
ideas and insights from previous architectures, resulting in in neurosurgical imaging, combining confocal laser endomi-
updates to the state-of-the-art. Table 1 contains a short list of croscopy with deep learning models for automatic detection
some famous CNN architectures, illustrating how the building of intraoperative CLE images on-the-fly [94].
blocks can be combined and how the field moves along. Another important application area is advanced deformable
These neural networks are typically implemented in one or image registration, enabling quantitative analysis across dif-
more of a small number of software frameworks that domi- ferent physical imaging modalities and across time.22 For
nates machine learning research, all built on top of NVIDIA’s example elastic registration between 3D MRI and transrec-
CUDA platform and the cuDNN library. Today’s deep learning tal ultrasound for guiding targeted prostate biopsy [95];
methods are almost exclusively implemented in either Tensor- deformable registration for brain MRI where a “cue-aware
Flow, a framework originating from Google Research, Keras, deep regression network” learns from a given set of train-
a deep learning library originally built by Fraņcois Chollet ing images the displacement vector associated with a pair
and recently incorporated in TensorFlow, or Pytorch, a frame- of reference-subject patches [96]; fast deformable image
work associated with Facebook Research. There are very few registration of brain MR image pairs by patch-wise pre-
exceptions (YOLO built using the Darknet framework [85] diction of the Large Deformation Diffeomorphic Metric
is one of the rare ones). All the main frameworks are open Mapping model [97]23 ; unsupervised convolutional neural
source and under active development. network-based algorithm for deformable image registra-
tion of cone-beam CT to CT using a deep convolutional
inverse graphics network [98]; deep learning-based 2D/3D
3 Deep learning, medical imaging and MRI
Deep learning methods are increasingly used to improve
clinical practice, and the list of examples is long, growing 22 E.g. test-retest examinations, or motion correction in dynamic imaging.
daily. We will not attempt a comprehensive overview of deep 23 Available at https://github.com/rkwitt/quicksilver.
Table 1
A far from exhaustive, non-chronological, list of CNN architectures and some high-level descriptions.
AlexNet [1] The network that launched the current deep learning boom by winning the 2012 ILSVRC competition by a huge margin.
Notable features include the use of RELUs, dropout regularization, splitting the computations on multiple GPUs, and
using data augmentation during training. ZFNet [67], a relatively minor modification of AlexNet, won the 2013
ILSVRC competition.
VGG [68] Popularized the idea of using smaller filter kernels and therefore deeper networks (up to 19 layers for VGG19, compared
to 7 for AlexNet and ZFNet), and training the deeper networks using pre-training on shallower versions.
GoogLeNet [69] Promoted the idea of stacking the layers in CNNs more creatively, as networks in networks, building on the idea of [70].
Inside a relatively standard architecture (called the stem), GoogLeNet contains multiple inception modules, in which
multiple different filter sizes are applied to the input and their results concatenated. This multi-scale processing allows
the module to extract features at different levels of detail simultaneously. GoogLeNet also popularized the idea of not
using fully-connected layers at the end, but rather global average pooling, significantly reducing the number of model
parameters. It won the 2014 ILSVRC competition.
ResNet [71] Introduced skip connections, which makes it possible to train much deeper networks. A 152 layer deep ResNet won the
2015 ILSVRC competition, and the authors also successfully trained a version with 1001 layers. Having skip
connections in addition to the standard pathway gives the network the option to simply copy the activations from layer to
layer (more precisely, from ResNet block to ResNet block), preserving information as data goes through the layers.
Some features are best constructed in shallow networks, while others require more depth. The skip connections facilitate
both at the same time, increasing the network’s flexibility when fed input data. As the skip connections make the
network learn residuals, ResNets perform a kind of boosting.
Highway nets [72] Another way to increase depth based on gating units, an idea from Long Short Term Memory (LSTM) recurrent
networks, enabling optimization of the skip connections in the network. The gates can be trained to find useful
combinations of the identity function (as in ResNets) and the standard nonlinearity through which to feed its input.
DenseNet [73] Builds on the ideas of ResNet, but instead of adding the activations produced by one layer to later layers, they are simply
concatenated together. The original inputs in addition to the activations from previous layers are therefore kept at each
layer (again, more precisely, between blocks of layers), preserving some kind of global state. This encourages feature
reuse and lowers the number of parameters for a given depth. DenseNets are therefore particularly well-suited for
smaller data sets (outperforming others on e.g. Cifar-10 and Cifar-100).
ResNext [74] Builds on ResNet and GoogLeNet by using inception modules between skip connections.
SENets [75] Squeeze-and-Excitation Networks, which won the ILSVRC 2017 competition, builds on ResNext but adds trainable
parameters that the network can use to weigh each feature map, where earlier networks simply added them up. These
SE-blocks allows the network to model the channel and spatial information separately, increasing the model capacity.
SE-blocks can easily be added to any CNN model, with negligible increase in computational costs.
NASNet [76] A CNN architecture designed by a neural network, beating all the previous human-designed networks at the ILSVRC
competition. It was created using AutoML,a Google Brain’s reinforcement learning approach to architecture design
[77]. A controller network (a recurrent neural network) proposes architectures aimed to perform at a specific level for a
particular task, and by trial and error learns to propose better and better models. NASNet was based on Cifar-10, and has
relatively modest computational demands, but still outperformed the previous state-of-the-art on ILSVRC data.
YOLO [78] Introduced a new, simplified way to do simultaneous object detection and classification in images. It uses a single CNN
operating directly on the image and outputting bounding boxes and class probabilities. It incorporates several elements
from the above networks, including inception modules and pretraining a smaller version of the network. It’s fast enough
to enable real-time processing.b YOLO makes it easy to trade accuracy for speed by reducing the model size.
YOLOv3-tiny was able to process images at over 200 frames per second on a standard benchmark data set, while still
producing reasonable predictions.
GANs [79] A generative adversarial network consists of two neural networks pitted against each other. The generative network G is
tasked with creating samples that the discriminative network D is supposed to classify as coming from the generative
network or the training data. The networks are trained simultaneously, where G aims to maximize the probability that D
makes a mistake while D aims for high classification accuracy.
Siamese nets [80] An old idea (e.g. [81]) that’s recently been shown to enable one-shot learning, i.e. learning from a single example. A
Siamese network consists of two identical neural networks, both the architecture and the weights, attached at the end.
They are trained together to differentiate pairs of inputs. Once trained, the features of the networks can be used to
perform one-shot learning without retraining.
U-net [82] A very popular and successful network for segmentation in 2D images. When fed an input image, it is first downsampled
through a “traditional” CNN, before being upsampled using transpose convolutions until it reaches its original size. In
addition, based on the ideas of ResNet, there are skip connections that concatenates features from the downsampling to
the upsampling paths. It is a fully-convolutional network, using the ideas first introduced in [83].
V-net [84] A three-dimensional version of U-net with volumetric convolutions and skip connections as in ResNet.
a https://cloud.google.com/automl.
b You can watch YOLO in action here https://youtu.be/VOC3huqHrss.
registration frame-work for registration of preoperative 3D measurement techniques is more recent, but already impres-
data and intraoperative 2D X-ray images in image-guided sive. From MR image acquisition and signal processing in
therapy [99]; real-time prostate segmentation during targeted MR fingerprinting, to denoising and super-resolution, and into
prostate biopsy, utilizing temporal information in the series of image synthesis.
ultrasound images [100].
This is just a tiny sliver of the many applications of 3.1.1 Data acquisition and image reconstruction
deep learning to central problems in medical imaging.
There are several thorough reviews and overviews of the Research on CNN and RNN-based image reconstruc-
field to consult for more information, across modalities tion methods is rapidly increasing, pioneered by Yang
and organs, and with different points of view and level of et al. [132] at NIPS 2016 and Wang et al. [133] at
technical details. For example the comprehensive review ISBI 2016. Recent applications addresses e.g. convolutional
[101],24 covering both medicine and biology and spanning recurrent neural networks for dynamic MR image reconstruc-
from imaging applications in healthcare to protein-protein tion [134], reconstructing good quality cardiac MR images
interaction and uncertainty quantification; key concepts of from highly undersampled complex-valued k-space data by
deep learning for clinical radiologists [102–111], includ- learning spatio-temporal dependencies, outperforming 3D
ing radiomics and imaging genomics (radiogenomics) [112], CNN approaches and compressed sensing-based dynamic
and toolkits and libraries for deep learning [113]; deep MRI reconstruction algorithms in computational complexity,
learning in neuroimaging and neuroradiology [114]; brain reconstruction accuracy and speed for different undersam-
segmentation [115]; stroke imaging [116,117]; neuropsy- pling rates. Schlemper et al. [135] created a deep cascade
chiatric disorders [118]; breast cancer [119,120]; chest of concatenated CNNs for dynamic MR image reconstruction,
imaging [121]; imaging in oncology [122–124]; medical making use of data augmentation, both rigid and elastic defor-
ultrasound [125,126]; and more technical surveys of deep mations, to increase the variation of the examples seen by the
learning in medical image analysis [41,127–129]. Finally, network and reduce overfitting.25 Using variational networks
for those who like to be hands-on, there are many instruc- for single-shot fast spin-echo MRI with variable density sam-
tive introductory deep learning tutorials available online. pling, Chen et al. [136] enabled real-time (200 ms per section)
For example [130], with accompanying code available image reconstruction, outperforming conventional parallel
at https://github.com/paras42/Hello World Deep Learning, imaging and compressed sensing reconstruction. In [137], the
where you’ll be guided through the construction of a authors explored the potential for transfer learning (pretrained
system that can differentiate a chest X-ray from an models) and assessed the generalization of learned image
abdominal X-ray using the Keras/TensorFlow frame- reconstruction regarding image contrast, SNR, sampling pat-
work through a Jupyter Notebook. Other nice tutorials tern and image content, using a variational network and true
are http://bit.ly/adltktutorial, based on the Deep Learn- measurement k-space data from patient knee MRI record-
ing Toolkit (DLTK) [131], and https://github.com/usuyama/ ings and synthetic k-space data generated from images in the
pydata-medical-image, based on the Microsoft Cognitive Berkeley Segmentation Data Set and Benchmarks. Employ-
Toolkit (CNTK). ing least-squares generative adversarial networks (GANs) that
Let’s now turn to the field of MRI, in which deep learning learns texture details and suppresses high-frequency noise,
has seen applications at each step of entire workflows. From [138] created a novel compressed sensing framework that
acquisition to image retrieval, from segmentation to disease can produce diagnostic quality reconstructions “on the fly”
prediction. We divide this into two parts: (i) the signal pro- (30 ms).26 A unified framework for image reconstruction
cessing chain close to the physics of MRI, including image [139], called automated transform by manifold approxima-
restoration and multimodal image registration (Fig. 3), and (ii) tion (AUTOMAP) consisting of a feedforward deep neural
the use of deep learning in MR image segmentation, disease network with fully connected layers followed by a sparse
detection, disease prediction and systems based on images and convolutional autoencoder, formulate image reconstruction
text data (reports), addressing a few selected organs such as generically as a data-driven supervised learning task that gen-
the brain, the kidney, the prostate and the spine (Fig. 4). erates a mapping between the sensor and the image domain
based on an appropriate collection of training data (e.g. MRI
3.1 From image acquisition to image registration examinations collected from the Human Connectome Project,
transformed to the k-space sensor domain).
Deep learning in MRI has typically been focused on
segmentation and classification of reconstructed magni-
tude images. Its penetration into the lower levels of MRI 25 Code available at https://github.com/js3611/Deep-MRI-Reconstruction.
26 In their GAN setting, a generator network is used to map undersampled
data to a realistic-looking image with high measurement fidelity, while a dis-
24 A continuous collaborative manuscript (https://greenelab.github.io/ criminator network is trained jointly to score the quality of the reconstructed
deep-review) with >500 references. image.
Figure 3. Deep learning in the MR signal processing chain, from image acquisition (in complex-valued k-space) and image reconstruction,
to image restoration (e.g. denoising) and image registration. The rightmost column illustrates coregistration of multimodal brain MRI.
sMRI = structural 3D T1-weighted MRI, dMRI = diffusion weighted MRI (stack of slices in blue superimposed on sMRI), fMRI = functional
BOLD MRI (in red).
Figure 4. Deep learning for MR image analysis in selected organs, partly from ongoing work at MMIV.
There are also other approaches and reports on deep was trained on synthetically generated data containing simple
learning in MR image reconstruction, e.g. [140–143], a fun- geometric shapes such as cubes, rectangles and spheres. After
damental field rapidly progressing. training their model on synthetic data it was able to generalize
to real-world clinical brain MRI data, computing susceptibil-
3.1.2 Quantitative parameters – QSM and MR ity maps within seconds end-to-end. The authors conclude that
fingerprinting their method, combined with fast imaging sequences, could
make QSM feasible in standard clinical practice.
Another area that is developing within deep learning for Magnetic resonance fingerprinting (MRF) was introduced
MRI is the estimation of quantitative tissue parameters from a little more than five years ago [149], and has been called
recorded complex-valued data. For example within quanti- “a promising new approach to obtain standardized imaging
tative susceptibility mapping, and in the exciting field of biomarkers from MRI” by the European Society of Radiology
magnetic resonance fingerprinting. [150]. It uses a pseudo-randomized acquisition that causes the
Quantitative susceptibility mapping (QSM) is a growing signals from different tissues to have a unique signal evolution
field of research in MRI, aiming to noninvasively estimate (“fingerprint”) that is a function of the multiple material prop-
the magnetic susceptibility of biological tissue [144,145]. The erties being investigated. Mapping the signals back to known
technique is based on solving the difficult, ill-posed inverse tissue parameters (T1, T2 and proton density) is then a rather
problem of determining the magnetic susceptibility from local difficult inverse problem. MRF is closely related to the idea of
magnetic fields. Recently Yoon et al. [146] constructed a compressed sensing [151] in MRI [152] in that MRF under-
three-dimensional CNN, named QSMnet and based on the samples data in k-space producing aliasing artifacts in the
U-Net architecture, able to generate high quality suscepti- reconstructed images that can be suppressed by compressed
bility source maps from single orientation data. The authors sensing.27 It can be regarded as a quantitative multiparamet-
generated training data by using the gold-standard for QSM: ric MRI analysis, and with recent acquisition schemes using a
the so-called COSMOS method [147]. The data was based single-shot spiral trajectory with undersampling, whole-brain
on 60 scans from 12 healthy volunteers. The resulting model
both simplified and improved the state-of-the-art for QSM.
Rasmussen and coworkers [148] took a different approach. 27 See [153–157] for recent perspectives and developments connecting deep
They also used a U-Net-based convolutional neural network learning-based reconstruction methods to the more general research field of
to perform field-to-source inversion, called DeepQSM, but it inverse problems.
coverage of T1 , T2 and proton density maps can be acquired at In the same context, Fang et al. [167] used a deep learning
1.2 × 1.2 × 3 mm3 voxel resolution in less than 5 min [158]. method to extract tissue properties from highly undersampled
The processing of MRF after acquisition usually involves 2D MRF-FISP data in brain imaging, where 2300 time points
using various pattern recognition algorithms that try to match were acquired from each measurement and each time point
the fingerprints to a predefined dictionary of predicted signal consisted of data from one spiral readout only. The real and
evolutions,28 created using the Bloch equations [149,163]. imaginary parts of the complex signal were separated into two
Recently, deep learning methodology has been applied to channels. They used MRF signal from a patch of 32 × 32 pix-
MR fingerprinting. Cohen et al. [164] reformulated the MRF els to incorporate correlated information between neighboring
reconstruction problem as learning an optimal function that pixels. In their work they designed a standard three-layer CNN
maps the recorded signal magnitudes to the corresponding with T1 and T2 as output.
tissue parameter values, trained on a sparse set of dictionary Virtue et al. [168] investigated a different approach to
entries. To achieve this they fed voxel-wise MRI data acquired MRF. By generating 100,000 synthetic MRI signals using
with an MRF sequence (MRF-EPI, 25 frames in ∼3 s; or a Bloch equation simulator they were able to train feedfor-
MRF-FISP, 600 frames in ∼7.5 s) to a four-layer neural net- ward deep neural networks to map new MRI signals to the
work consisting of two hidden layers with 300 × 300 fully tissue parameters directly, producing approximate solutions
connected nodes and two nodes in the output layer, consid- to the inverse mapping problem of MRF. In their work they
ering only T1 and T2 parametric maps. The network, called designed a new complex activation function, the complex
MRF Deep RecOnstruction NEtwork (DRONE), was trained cardioid, that was used to construct a complex-valued feedfor-
by an adaptive moment estimation stochastic gradient descent ward neural network. This three-layer network outperformed
algorithm with a mean squared error loss function. Their dic- both the standard MRF techniques based on dictionary match-
tionary consisted of ∼70,000 entries (product of discretized ing, and also the analogous real neural network operating on
T1 and T2 values) and training the network to convergence the real and imaginary components separately. This suggested
with this dictionary (∼10 MB for MRF-EPI and ∼300 MB that complex-valued networks are better suited at uncovering
for MRF-FISP) required 10 to 70 min using an NVIDIA K80 information in complex data.29
GPU with 2 GB memory. They found their reconstruction
time (10 to 70 ms per slice) to be 300 to 5000 times faster 3.1.3 Image restoration (denoising, artifact detection)
than conventional dictionary-matching techniques, using both
well-characterized calibrated ISMRM/NIST phantoms and Estimation of noise and image denoising in MRI has been an
in vivo human brains. important field of research for many years [171,172], employ-
A similar deep learning approach to predict quantitative ing a plethora of methods. For example Bayesian Markov
parameter values (T1 and T2 ) from MRF time series was taken random field models [173], rough set theory [174], higher-
by Hoppe et al. [165]. In their experiments they used 2D MRF- order singular value decomposition [175], wavelets [176],
FISP data with variable TR (12–15 ms), flip angles (5–74◦ ) independent component analysis [177], or higher order PDEs
and 3000 repetitions, recorded on a MAGNETOM 3T Skyra. [178].
A high resolution dictionary was simulated to generate a large Recently, deep learning approaches have been introduced
collection of training and testing data, using tissues T1 and to denoising. In their work on learning implicit brain MRI
T2 relaxation time ranges as present in normal brain at 3T manifolds using deep neural networks, Bermudez et al. [179]
(e.g. [166]) resulting in ∼1.2 × 105 time series. In contrast implemented an autoencoder with skip connections for image
to [164], their deep neural network architecture was inspired denoising, testing their approach with adding various levels
from the domain of speech recognition due to the similarity of Gaussian noise to more than 500 T1-weighted brain MR
of the two tasks. The architecture with the smallest average images from healthy controls in the Baltimore Longitudinal
error for validation data was a standard convolutional neural Study of Aging. Their autoencoder network outperformed the
network consisting of an input layer of 3000 nodes (number of current FSL SUSAN denoising software according to peak
samples in the recorded time series), four hidden layers, and signal-to-noise ratios. Benou et al. [180] addressed spatio-
an output layers with two nodes (T1 and T2 ). Matching one temporal denoising of dynamic contrast-enhanced MRI of the
time series was about 100 times faster than the conventional brain with bolus injection of contrast agent (CA), proposing a
[149] matching method and with very small mean absolute novel approach using ensembles of deep neural networks for
deviations from ground truth values. noise reduction. Each DNN was trained on a different range
of SNRs and types of CA concentration time curves (denoted
“pathology experts”, “healthy experts”, “vessel experts”) to
28 A dictionary of time series for every possible combination of param-
eters like (discretized) T1 and T2 relaxation times, spin-density (M0 ), B0 ,

off-resonance (f), and also voxel-wise cerebral blood volume (CBV), mean 29 Complex-valued deep learning is also getting some attention in a broader
vessel radius (R), blood oxygen saturation (SO2 ) and T2∗ [159–161], and community of researchers, and has been shown to lead to improved models.
more, e.g. MFR-ASL [162]. See e.g. [169,170] and the references therein.
generate a reconstruction hypothesis from noisy input by weighted images [195]. Recently deep learning approaches
using a classification DNN to select the most likely hypoth- has been introduced, e.g. generating super-resolution single
esis and provide a “clean output” curve. Training data was (no reference information) and multi-contrast (applying a
generated synthetically using a three-parameter Tofts phar- high-resolution image of another modality as reference) brain
macokinetic (PK) model and noise realizations. To improve MR images using CNNs [196]; constructing superresolution
this model, accounting for spatial dependencies of PK phar- brain MRI by a CNN stacked by multi-scale fusion units [197];
macokinetics, they used concatenated noisy time curves from and super-resolution musculoskeletal MRI (“DeepResolve”)
first-order neighbourhood pixels in their expert DNNs and [198]. In DeepResolve thin (0.7 mm) slices in knee images
ensemble hypothesis DNN, collecting neighboring recon- (DESS) from 124 patients included in the Osteoarthritis Ini-
structions before a boosting procedure produced the final tiative were used for training and 17 patients for testing, with
clean output for the pixel of interest. They tested their trained a 10 s inference time per 3D (344 × 344 × 160) volume. The
ensemble model on 33 patients from two different DCE- resulting images were evaluated both quantitatively (MSE,
MRI databases with either stroke or recurrent glioblastoma PSNR, and the perceptual window-based structural similarity
(RIDER NEURO30 ), acquired at different sites, with differ- SSIM32 index) and qualitatively by expert radiologists.
ent imaging protocols, and with different scanner vendors and
field strengths. The qualitative and quantitative (MSE) denois- 3.1.5 Image synthesis
ing results were better than spatiotemporal Beltrami, moving
Image synthesis in MRI have traditionally been seen as a
average, the dynamic Non Local Means method [181], and
method to derive new parametric images or new tissue contrast
stacked denoising autoencoders [182]. The run-time compar-
from a collection of MR acquisition performed at the same
isons were also in favor of the proposed sDNN. In this context
imaging session, i.e. “an intensity transformation applied to
of DCE-MRI, it’s tempting to speculate whether deep neural
a given set of input images to generate a new image with a
network approaches could be used for direct estimation of
specific tissue contrast” [199]. Another avenue of MRI syn-
tracer-kinetic parameter maps from highly undersampled (k,
thesis is related to quantitative imaging and the development
t)-space data in dynamic recordings [183,184], a powerful
and use of physical phantoms, imaging calibration/standard
way to by-pass 4D DCE-MRI reconstruction altogether and
test objects with specific material properties. This is done
map sensor data directly to spatially resolved pharmacokinetic
in order to assess the performance of an MRI scanner or to
parameters, e.g. Ktrans , vp , ve in the extended Tofts model or
assess imaging biomarkers reliably with application-specific
parameters in other classic models [185]. A related approach
phantoms such as a structural brain imaging phantom, DCE-
in the domain of diffusion MRI, by-passing the model-fitting
MRI perfusion phantom, diffusion phantom, flow phantom,
steps and computing voxelwise scalar tissue properties (e.g.
breast phantom or a proton-density fat fraction phantom [200].
radial kurtosis, fiber orientation dispersion index) directly
The in silico modeling of MR images with certain underlying
from the subsampled DWIs was taken by Golkov et al. [186]
properties, e.g. [201,202], or model-based generation of large
in their proposed “q-space deep learning” family of methods.
databases of (cardiac) images from real healthy cases [203]
Deep learning methods has also been applied to MR artifact
is also part of this endeavour. In this context, deep learning
detection, e.g. poor quality spectra in MRSI [187]; detection
approaches have accelerated research and the amount of costly
and removal of ghosting artifacts in MR spectroscopy [188];
training data.
and automated reference-free detection of patient motion arti-
The last couple of years have seen impressive results for
facts in MRI [189].
photo-realistic image synthesis using deep learning tech-
3.1.4 Image super-resolution niques, especially generative adversarial networks (GANs,
introduced by Goodfellow et al. in 2014 [79]), e.g. [204–206].
Image super-resolution, reconstructing a higher-resolution These can also be used for biological image synthesis
image or image sequence from the observed low-resolution [207,208] and text-to-image synthesis [209–211].33 Recently,
image [190], is an exciting application of deep learning a group of researchers from NVIDIA, MGH & BWH Center
methods.31 for Clinical Data Science in Boston, and the Mayo Clinic in
Super-resolution for MRI have been around for almost 10 Rochester [212] designed a clever approach to generate syn-
years [191,192] and can be used to improve the trade-off thetic abnormal MRI images with brain tumors by training a
between resolution, SNR, and acquisition time [193], gener- GAN based on pix2pix34 using two publicly available data sets
ate 7T-like MR images on 3T MRI scanners [194], or obtain of brain MRI (ADNI and the BRATS’15 Challenge, and later
super-resolution T1 maps from a set of low resolution T1 also the Ischemic Stroke Lesion Segmentation ISLES’2018
30 https://wiki.cancerimagingarchive.net/display/Public/RIDER+NEURO+ 32 http://www.cns.nyu.edu/∼lcv/ssim.
MRI. 33 See here https://github.com/xinario/awesome-gan-for-medical-imaging
31 See http://course.fast.ai/lessons/lesson14.html for an instructive introduc- for a list of interesting applications of GAN in medical imaging.
tion to super-resolution. 34 https://phillipi.github.io/pix2pix.
Challenge). This approach is highly interesting as medical and applications37 [228–237], including “previous genera-
imaging datasets are often imbalanced, with few patholog- tion” artificial neural networks [238].
ical findings, limiting the training of deep learning models. Recently, deep learning methods have been applied to
Such generative models for image synthesis serve as a form image registration in order to improve accuracy and speed
of data augmentation, and also as an anonymization tool. (e.g. Section 3.4 in [41]). For example: deformable image
The authors achieved comparable tumor segmentation results registration [97,239]; model-to-image registration [240,241];
when trained on the synthetic data rather than on real patient MRI-based attenuation correction for PET [242,243];
data. A related approach to brain tumor segmentation using PET/MRI dose calculation [244]; unsupervised end-to-end
coarse-to-fine GANs was taken by Mok and Chung [213]. learning for deformable registration of 2D CT/MR images
Guibas et al. [214] used a two-stage pipeline for generating [245]; an unsupervised learning model for deformable, pair-
synthetic medical images from a pair of GANs, addressing wise 3D medical image registration by Balakrishnan et al.
retinal fundus images, and provided an online repository (Syn- [246]38 ; and a deep learning framework for unsupervised
thMed) for synthetic medical images. Kitchen and Seah [215] affine and deformable image registration [247].
used GANs to synthetize realistic prostate lesions in T2 , ADC,
Ktrans resembling the SPIE-AAPM-NCI ProstateX Challenge 3.2 From image segmentation to diagnosis and
201635 training data. prediction
Other applications are unsupervised synthesis of T1-
weighted brain MRI using a GAN [179]; image synthesis We leave the lower-level applications of deep learning
with context-aware GANs [216]; synthesis of patient- in MRI to consider higher-level (down-stream) applications
specific transmission image for PET attenuation correction in such as fast and accurate image segmentation, disease pre-
PET/MR imaging of the brain using a CNN [217]; pseudo-CT diction in selected organs (brain, kidney, prostate, and spine)
synthesis for pelvis PET/MR attenuation correction using a and content-based image retrieval, typically applied to recon-
Dixon-VIBE Deep Learning (DIVIDE) network [218]; image structed magnitude images. We have chosen to focus our
synthesis with GANs for tissue recognition [219]; synthetic overview on deep learning applications close to the MR
data augmentation using a GAN for improved liver lesion clas- physics and will be brief in the present section, even if
sification [220]; and deep MR to CT synthesis using unpaired the following applications are very interesting and clinically
data [221]. important.
3.2.1 Image segmentation

3.1.6 Image registration
Image segmentation, the holy grail of quantitative image
Image registration36 is an increasingly important field analysis, is the process of partitioning an image into multiple
within MR image processing and analysis as more comple- regions that share similar attributes, enabling localization and
mentary and multiparametric tissue information are collected quantification.39 It has an almost 50 years long history, and
in space and time within shorter acquisition times, at higher has become the biggest target for deep learning approaches
spatial (and temporal) resolutions, often longitudinally, and in medical imaging. The multispectral tissue classification
across patient groups, larger cohorts, or atlases. Tradition- report by Vannier et al. in 1985 [248], using statistical pat-
ally one has divided the tasks of image registration into tern recognition techniques (and satellite image processing
dichotomies: intra vs. inter-modality, intra vs. inter-subject, software from NASA), represented one of the most semi-
rigid vs. deformable, geometry-based vs. intensity-based, and nal works leading up to today’s machine learning in medical
prospective vs. retrospective image registration. Mathemati- imaging segmentation. In this early era, we also had the oppor-
cally, registration is a challenging mix of geometry (spatial tunity to contribute with supervised and unsupervised machine
transformations), analysis (similarity measures), optimiza- learning approaches for MR image segmentation and tissue
tion strategies, and numerical schemes. In prospective motion classification [249–252]. An impressive range of segmenta-
correction, real-time MR physics is also an important part tion methods and approaches have been reported (especially
of the picture [223,224]. A wide range of methodological for brain segmentation) and reviewed, e.g. [253–261]. MR
approaches have been developed and tested for various organs image segmentation using deep learning approaches, typically
CNNs, are now penetrating the whole field of applications.
35 37 And different hardware e.g. GPUs [225–227] as image registration is

https://www.aapm.org/GrandChallenge/PROSTATEx-2.
36 Image registration can be defined as “the determination of a one-to-one often computationally time consuming.
38 With code available at https://github.com/voxelmorph/voxelmorph.
mapping between the coordinates in one space and those in another, such that
39 Segmentation is also crucial for functional imaging, enabling tissue phys-
points in the two spaces that correspond to the same anatomical point are
mapped to each other” (C.R Maurer [222], 1993). iology quantification with preservation of anatomical specificity.
For example acute ischemic lesion segmentation in DWI textual paragraphs [313–316], and also to identify findings in
[262]; brain tumor segmentation [263]; segmentation of the radiology reports [317–319].
striatum [264]; segmentation of organs-at-risks in head and
neck CT images [265]; and fully automated segmentation
of polycystic kidneys [266]; deformable segmentation of 4 Open science and reproducible research in
the prostate [267]; and spine segmentation with 3D mul- machine learning for medical imaging
tiscale CNNs [268]. See [41,101] for more comprehensive
lists. Machine learning is moving at a breakneck speed, too fast
for the standard peer-review process to keep up. Many of the
3.2.2 Diagnosis and prediction most celebrated and impactful papers in machine learning
over the past few years are only available as preprints, or
A presumably complete list of papers up to 2017 using published in conference proceedings long after their results
deep learning techniques for brain image analysis is pro- are well-known and incorporated in the research of others.
vided as Table 1 in Litjens at al. [41]. In the Table 2 we Bypassing peer-review has some downsides, of course, but
add some more recent work on organ-specific deep learning these are somewhat mitigated by researchers’ willingness to
using MRI, restricting ourselves to brain, kidney, prostate and share code and data.41
spine. Most of the main new ideas and methods are posted to the
arXiv preprint server,42 and the accompanying code shared
on the GitHub platform.43 The data sets used are often openly
3.3 Content-based image retrieval
available through various repositories. This, in addition to the
many excellent online educational resources,44 makes it easy
The objective of content-based image retrieval (CBIR)
to get started in the field. Select a problem you find inter-
in radiology is to provide medical cases similar to a given
esting based on openly available data, a method described in
image in order to assist radiologists in the decision-making
a preprint, and an implementation uploaded to GitHub. This
process. It typically involves large case databases, clever
forms a good starting point for an interesting machine learning
image representations and lesion annotations, and algorithms
project.
that are able to quickly and reliably match and retrieve
Another interesting aspect about modern machine learning
the most similar images and their annotations in the case
and data science is the prevalence of competitions, with the
database. CBIR has been an active area of research in med-
annual ImageNet ILSVRC competition as the main driver of
ical imaging for many years, addressing a wide range of
progress in deep learning for computer vision since 2012.
applications, imaging modalities, organs, and methodologi-
Each competition typically draws large number of partici-
cal approaches, e.g. [297–303], and at a larger scale outside
pants, and the top results often push the state-of-the art to
the medical field using deep learning techniques, e.g. at
a new level. In addition to inspiring new ideas, competitions
Microsoft, Apple, Facebook, and Google (reverse image
also provide natural entry points to modern machine learn-
search40 ), and others. See e.g. [304–308] and the code repos-
ing. It is interesting to note how deep learning-based models
itories https://github.com/topics/image-retrieval. One of the
are completely dominating the leaderboards of essentially all
first application of deep learning for CBIR in the medical
image-based competitions. Other machine learning models,
domain came in 2015 when Sklan et al. [309] trained a CNN
or non-machine learning-based techniques, have largely been
to perform CBIR with more than one million random MR and
outclassed.
CT images, with disappointing results (true positive rate of
What’s true about the openness of machine learning in gen-
20%) on their independent test set of 2100 labeled images.
eral is increasingly true also for the sub-field of machine
Medical CBIR is now, however, dominated by deep learning
learning for medical image analysis. We’ve listed a few
algorithms [310–312]. As an example, by retrieving medical
examples of openly available implementations, data sets and
cases similar to a given image, Pizarro et al. [278] developed
challenges in Tables 3–5.
a CNN for automatically inferring the contrast of MRI scans
based on the image intensity of multiple slices. Recently,
deep learning methods have also been used for automated
generation of radiology reports, typically incorporating long- 41 In the spirit of sharing and open science, we’ve created a GitHub
short-term-memory (LSTM) network models to generate the repository to accompany our article, available at https://github.com/
MMIV-ML/DLMI2018.
42 http://arxiv.org.
43 https://github.com.
40 See “search by image” https://images.google.com/https://developers. 44 For example http://www.fast.ai, https://www.deeplearning.ai, http://
google.com/custom-search, and also https://tineye.com, indexing more than cs231n.stanford.edu, https://developers.google.com/machine-learning/crash-

30 billion images. course.
Table 2
A short list of deep learning applications per organ, task, reference and description.
Brain
Brain extraction [269] A 3D CNN for skull stripping
Functional connectomes [270] Transfer learning approach to enhance deep neural network classification of brain functional connectomes
[271] Multisite diagnostic classification of schizophrenia using discriminant deep learning with functional connectivity
MRI
Structural connectomes [272] A convolutional neural network-based approach (https://github.com/MIC-DKFZ/TractSeg) that directly
segments tracts in the field of fiber orientation distribution function (fODF) peaks without using tractography,
image registration or parcellation. Tested on 105 subjects from the Human Connectome Project
Brain age [273] Chronological age prediction from raw brain T1-MRI data, also testing the heritability of brain-predicted age
using a sample of 62 monozygotic and dizygotic twins
Alzheimer’s disease [274] Landmark-based deep multi-instance learning evaluated on 1526 subjects from three public datasets (ADNI-1,
ADNI-2, MIRIAD)
[275] Identify different stages of AD
[276] Multimodal and multiscale deep neural networks for the early diagnosis of AD using structural MR and
FDG-PET images
Vascular lesions [277] Evaluation of a deep learning approach for the segmentation of brain tissues and white matter hyperintensities of
presumed vascular origin in MRI
Identification of MRI [278] Using deep learning algorithms to automatically identify the brain MRI contrast, with implications for managing
contrast large databases
Meningioma [279] Fully automated detection and segmentation of meningiomas using deep learning on routine multiparametric
MRI
Glioma [280] Glioblastoma segmentation using heterogeneous MRI data from clinical routine
[281] Deep learning for segmentation of brain tumors and impact of cross-institutional training and testing
[282] Automatic semantic segmentation of brain gliomas from MRI using a deep cascaded neural network
[283] AdaptAhead optimization algorithm for learning deep CNN applied to MRI segmentation of glioblastomas
(BRATS)
Multiple sclerosis [284] Deep learning of joint myelin and T1w MRI features in normal-appearing brain tissue to distinguish between
multiple sclerosis patients and healthy controls
Kidney
Abdominal organs [285] CNNs to improve abdominal organ segmentation, including left kidney, right kidney, liver, spleen, and stomach
in T2 -weighted MR images
Cyst segmentation [266] An artificial multi-observer deep neural network for fully automated segmentation of polycystic kidneys
Renal transplant [286] A deep-learning-based classifier with stacked non-negative constrained autoencoders to distinguish between
rejected and non-rejected renal transplants in DWI recordings
Prostate
Cancer (PCa) [287] Proposed a method for end-to-end prostate segmentation by integrating holistically (image-to-image) nested
edge detection with fully convolutional networks. Their nested networks automatically learn a hierarchical
representation that can improve prostate boundary detection. Obtained very good results (Dice coefficient, 5-fold
cross validation) on MRI scans from 250 patients
[288] Computer-aided diagnosis with a CNN, deciding ‘cancer’ ‘no cancer’ trained on data from 301 patients with a
prostate-specific antigen level of <20 ng/mL who underwent MRI and extended systematic prostate biopsy with
or without MRI-targeted biopsy
[289] Automatic approach based on deep CNN, inspired from VGG, to classify PCa and noncancerous tissues with
multiparametric MRI using data from the PROSTATEx database
[290] Deep CNN and a non-deep learning using feature detection (the scale-invariant feature transform and the
bag-of-words model, a representative method for image recognition and analysis) were used to distinguish
pathologically confirmed PCa patients from prostate benign conditions patients with prostatitis or prostate
benign hyperplasia in a collection of 172 patients with more than 2500 morphologic 2D T2 -w MR images
[291] Designed a system which can concurrently identify the presence of PCa in an image and localize lesions based
on deep CNN features (co-trained CNNs consisting of two parallel convolutional networks for ADC and T2 -w
images respectively) and a single-stage SVM classifier for automated detection of PCa in multiparametric MRI.
Evaluated on a dataset of 160 patients
[292] Designed and tested multimodel CNNs, using clinical data from 364 patients with a total of 463 PCa lesions and
450 identified noncancerous image patches. Carefully investigated three critical factors which could greatly
affect the performance of their multimodal CNNs but had not been carefully studied previously: (1) Given
limited training data, how can these be augmented in sufficient numbers and variety for fine-tuning deep CNN
networks for PCa diagnosis? (2) How can multimodal mp-MRI information be effectively combined in CNNs?
(3) What is the impact of different CNN architectures on the accuracy of PCa diagnosis?
Table 2 (Continued)
Spine
Vertebrae labeling [293] Designed a CNN for detection and labeling of vertebrae in MR images with clinical annotations as training data
Intervertebral disc [268] 3D multi-scale fully connected CNNs with random modality voxel dropout learning for intervertebral disc
localization localization and segmentation from multi-modality MR images
Disc-level labeling, spinal [294] CNN model denoted DeepSPINE, having a U-Net architecture combined with a spine-curve fitting method for
stenosis grading automated lumbar vertebral segmentation, disc-level designation, and spinal stenosis grading with a natural
language processing scheme
Lumbal neural forminal [295] Addressed the challenge of automated pathogenesis-based diagnosis, simultaneously localizing and grading
stenosis (LNFS) multiple spinal structures (neural foramina, vertebrae, intervertebral discs) for diagnosing LNFS and discover
pathogenic factors. Proposed a deep multiscale multitask learning network integrating a multiscale multi-output
learning and a multitask regression learning into a fully convolutional network where (i) a DMML-Net merges
semantic representations to reinforce the salience of numerous target organs (ii) a DMML-Net extends
multiscale convolutional layers as multiple output layers to boost the scale-invariance for various organs, and
(iii) a DMML-Net joins the multitask regression module and the multitask loss module to combine the mutual
benefit between tasks
Spondylitis vs [296] CNN model for differentiating between tuberculous and pyogenic spondylitis in MR images. Compared their
tuberculosis CNN performance with that of three skilled radiologists using spine MRIs from 80 patients
Metastasis [290] A multi-resolution approach for spinal metastasis detection using deep Siamese neural networks comprising
three identical subnetworks for multi-resolution analysis and detection. Detection performance was evaluated on
a set of 26 cases using a free-response receiver operating characteristic analysis (observer is free to mark and rate
as many suspicious regions as are considered clinically reportable)
Table 3
A short list of openly available code for ML in medical imaging.
Summary Reference Implementation
NiftyNet. An open source convolutional neural networks platform [320,321] http://niftynet.io

for medical image analysis and image-guided therapy
DLTK. State of the art reference implementations for deep learning [131] https://github.com/DLTK/DLTK
on medical images
DeepMedic [322] https://github.com/Kamnitsask/deepmedic
U-Net: Convolutional Networks for Biomedical Image [323] https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net
Segmentation
V-net [84] https://github.com/faustomilletari/VNet
SegNet: A Deep Convolutional EncoderDecoder Architecture for [324] https://mi.eng.cam.ac.uk/projects/segnet
Robust Semantic Pixel-Wise Labelling
Brain lesion synthesis using GANs [212] https://github.com/khcs/brain-synthesis-lesion-segmentation
GANCS: Compressed Sensing MRI based on Deep Generative [325] https://github.com/gongenhao/GANCS
Adversarial Network
Deep MRI Reconstruction [135] https://github.com/js3611/Deep-MRI-Reconstruction
Graph Convolutional Networks for brain analysis in populations, [326] https://github.com/parisots/population-gcn
combining imaging and non-imaging data
5 Challenges, limitations and future amounts of hype and excitement.45 Others are more domain-
perspectives specific.
In deep learning for standard computer vision tasks, like
It is clear that deep neural networks are very useful when object recognition and localization, powerful models and a
one is tasked with producing accurate decisions based on set of best practices have been developed over the last few
complicated data sets. But they come with some significant years. The pace of development is still incredibly high, but
challenges and limitations that you either have to accept or certain things seem to be settled, at least momentarily. Using
try to overcome. Some are general: from technical challenges
related to the lack of mathematical and theoretical underpin-
nings of many central deep learning models and techniques, 45 Lipton: Machine Learning: The Opportunity and the Opportunists
and the resulting difficulty in deciding exactly what it is that
https://www.technologyreview.com/video/612109, Jordan: Artificial Intel-
makes one model better than another, to societal challenges ligence – The Revolution Hasn’t Happened Yet https://medium.com/
related to maximization and spread of the technological ben- @mijordan3/artificial-intelligence-the-revolution-hasnt-happened-yet-
efits [327,328] and the problems related to the tremendous 5e1d5812e1e7.
Table 4
A short list of medical imaging data sets and repositories.
Name Summary Link
OpenNeuro An open platform for sharing neuroimaging data under the public domain license. https://openneuro.orga
Contains brain images from 168 studies (4,718 participants) with various imaging
modalities and acquisition protocols.
UK Biobank Health data from half a million participants. Contains MRI images from 15,000 http://www.ukbiobank.ac.uk/
participants, aiming to reach 100,000.
TCIA The cancer imaging archive hosts a large archive of medical images of cancer http://www.cancerimagingarchive.net
accessible for public download. Currently contains images from 14,355 patients
across 77 collections.
ABIDE The autism brain imaging data exchange. Contains 1114 datasets from 521 individuals http://fcon 1000.projects.nitrc.org/indi/abide
with Autism Spectrum Disorder and 593 controls.
ADNI The Alzheimer’s disease neuroimaging initiative. Contains image data from almost http://adni.loni.usc.edu/
2000 participants (controls, early MCI, MCI, late MCI, AD)
a Data can be downloaded from the AWS S3 Bucket https://registry.opendata.aws/openneuro.
Table 5
A short list of medical imaging competitions.
Name Summary Link
Grand-Challenges Grand challenges in biomedical image https://grand-challenge.org/

analysis. Hosts and lists a large number of
competitions
RSNA Pneumonia Detection Challenge Automatically locate lung opacities on chest https://www.kaggle.com/c/rsna-pneumonia-detection-challenge
radiographs
HVSMR 2016 Segment the blood pool and myocardium http://segchd.csail.mit.edu/
from a 3D cardio-vascular magnetic
resonance image
ISLES 2018 Ischemic Stroke Lesion Segmentation 2018. http://www.isles-challenge.org/
The goal is to segment stroke lesions based
on acute CT perfusion data.
BraTS 2018 Multimodal Brain Tumor Segmentation. The http://www.med.upenn.edu/sbia/brats2018.html
goal is to segment brain tumors in
multimodal MRI scans.
CAMELYON17 The goal is to develop algorithms for https://camelyon17.grand-challenge.org/Home
automated detection and classification of
breast cancer metastases in whole-slide
images of histological lymph node sections.
ISIC 2018 Skin Lesion Analysis Towards Melanoma https://challenge2018.isic-archive.com/
Detection
Kaggle’s 2018 Data Science Bowl Spot Nuclei. Speed Cures. https://www.kaggle.com/c/data-science-bowl-2018
Kaggle’s 2017 Data Science Bowl Turning Machine Intelligence Against Lung https://www.kaggle.com/c/data-science-bowl-2017
Cancer
Kaggle’s 2016 Data Science Bowl Transforming How We Diagnose Heart https://www.kaggle.com/c/second-annual-data-science-bowl
Disease
MURA Determine whether a bone X-ray is normal https://stanfordmlgroup.github.io/competitions/mura/
or abnormal
the basic building blocks described above, placed according memory and compute consumption when using CNNs with
to the ideas behind, say, ResNet and SENet, will easily result higher-dimensional image data, challenges that researchers
in close to state-of-the-art performance on two-dimensional are trying various approaches to deal with (treating 3D as
object detection, image classification and segmentation stacks of 2Ds, patch- or segment-based training and inference,
tasks. downscaling, etc.). It is clear that the ideas behind state-of-the-
However, the story for deep learning in medical imag- art two-dimensional CNNs can be lifted to three dimensions,
ing is not quite as settled. One issue is that medical images but also that adding a third spatial dimension results in addi-
are often three-dimensional, and three-dimensional convo- tional constraints. Other important challenges are related to
lutional neural networks are as well-developed as their 2D data, trust, interpretability, workflow integration, and regula-
counterparts. One quickly meet challenges associated to tions, as discussed below.
5.1 Data weights from this network to a network designed for the task at
hand. For two-dimensional images one will almost always use
This is a crucially important obstacle for deep neural net- a network that has been pre-trained on the ImageNet data set.
works, especially in medical data analysis. When deploying The basic features in the earlier layers of the neural network
deep neural networks, or any other machine learning model, found from this data set typically retain their usefulness in any
one is instantly faced with challenges related to data access, other image-related task (or are at least form a better starting
privacy issues, data protection, and more. point than random initialization of the weights, which is the
As privacy and data protection is often a requirement when alternative). Starting from weights tuned on a larger training
dealing with medical data, new techniques for training models data set can also make the network more robust. Focusing
without exposing the underlying training data to the user of the weight updates during training on later layers requires
the model are necessary. It is not enough to merely restrict less data than having to do significant updates throughout the
access to the training set used to construct the model, as it entire network. One can also do interorgan transfer learning
is easy to use the model itself to discover details about the in 3D, an idea we have used for kidney segmentation, where
training set [329]. Even hiding the model and only exposing pre-training a network to do brain segmentation decreased the
a prediction interface would still leave it open to attack, for number of annotated kidneys needed to achieve good segmen-
example in the form of model-inversion [330] and member- tation performance [338]. The idea of pre-training networks
ship attacks [331]. Most current work on deep learning for is not restricted to images. Pre-training entire models has
medical data analysis use either open, anonymized data sets recently been demonstrated to greatly impact the performance
(as those in Table 4), or locally obtained anonymized research of natural language processing systems [2–4].
data, making these issues less relevant. However, the general Another widely used technique is augmenting the training
deep learning community are focusing a lot of attention on data set by applying various transformations that preserves the
the issue of privacy, and new techniques and frameworks for labels, as in rotations, scalings and intensity shifts of images,
federated learning [332]46 and differential privacy [333–335] or more advanced data augmentation techniques like anatomi-
are rapidly improving. There are a few examples of these ideas cally sound deformations, or other data set specific operations
entering the medical machine learning community, as in [336] (for example in our work on kidney segmentation from DCE-
where the distribution of deep learning models among several MRI, where we used image registration to propagate labels
medical institutions was investigated, but then without consid- through a time course of images [339]). Data synthesis, as in
ering the above privacy issues. As machine learning systems [212], is another interesting approach.
in medicine grows to larger scales, perhaps even including In short, as expert annotators are expensive, or simply not
computations and learning on the “edge”, federated learning available, spending large computational resources to expand
and differential privacy will likely become the focus of much your labeled training data set, e.g. indirectly through transfer
research in our community. learning or directly through data augmentation, is typically
If you are able to surmount these obstacles, you will be worthwhile. But whatever you do, the way current deep neu-
confronted with deep neural networks’ insatiable appetite for ral networks are constructed and trained results in significant
training data. These are very inefficient models, requiring data size requirements. There are new ways of construct-
large number of training samples before they can produce ing more data-efficient deep neural networks on the horizon,
anything remotely useful, and labeled training data is typi- for example by encoding more domain-specific elements in
cally both expensive and difficult to produce. In addition, the the neural network structure as in the capsule systems of
training data has to be representative of the data the network [340,341], which adds viewpoint invariance. It is also possi-
will meet in the future. If the training samples are from a data ble to add attention mechanisms to neural networks [342,343],
distribution that is very different from the one met in the real enabling them to focus their resources on the most informative
world, then the network’s generalization performance will be components of each layer input.
lower than expected. See [337] for a recent exploration of However, the networks that are most frequently used, and
this issue. Considering the large difference between the high- with the best raw performance, remain the data-hungry stan-
quality images one typically work with when doing research dard deep neural networks.
and the messiness of the real, clinical world, this can be a major
obstacle when putting deep learning systems into production. 5.2 Interpretability, trust and safety
Luckily there are ways to alleviate these problems some-
what. A widely used technique is transfer learning, also called As deep neural networks relies on complicated inter-
fine-tuning or pre-training: first you train a network to perform connected hierarchical representations of the training data
a task where there is an abundance of data, and then you copy to produce its predictions, interpreting these predictions
becomes very difficult. This is the “black box” problem
of deep neural networks [344]. They are capable of pro-
46 See for example https://ai.googleblog.com/2017/04/federated-learning- ducing extremely accurate predictions, but how can you
collaborative.html. trust predictions based on features you cannot understand?
Considerable effort goes into developing new ways to deal learning with other techniques, or (iii) no deep learning com-
with this problem, including DARPA launching a whole pro- ponent at all.
gram “Explainable AI ”47 dedicated to this issue, and lots of Beyond the application of machine learning in medi-
research going into enhancing interpretability [345,346], and cal imaging, we believe that the attention in the medical
finding new ways to measure sensitivity and visualize features community can also be leveraged to strengthen the general
[67,347–350]. computational mindset among medical researchers and practi-
Another way to increase their trustworthiness is to make tioners, mainstreaming the field of computational medicine.49
them produce robust uncertainty estimates in addition to Once there are enough high-impact software-systems based
predictions. The field of Bayesian Deep Learning aims to on mathematics, computer science, physics and engineering
combine deep learning and Bayesian approaches to uncer- entering the daily workflow in the clinic, the acceptance for
tainty. The ideas date back to the early 90s [351–353], but other such systems will likely grow. The access to bio-sensors
the field has recently seen renewed interest from the machine and (edge) computing on wearable devices for monitoring dis-
learning community at large, as new ways of computing uncer- ease or lifestyle, plus an ecosystem of machine learning and
tainty estimates from state of the art deep learning models have other computational medicine-based technologies, will then
been developed [58,64,354]. In addition to producing valuable likely facilitate the transition to a new medical paradigm that
measures that function as uncertainty measures [65,355,356], is predictive, preventive, personalized, and participatory – P4
these techniques can also lessen deep neural networks suscep- medicine [359].50
tibility to adversarial attacks [354,357].
5.3 Workflow integration, regulations

Acknowledgements
We thank Renate Grüner for useful discussions. The anony-
Another stumbling block for successful incorporation of
mous reviewers gave us excellent constructive feedback that
deep learning methods is workflow integration. It is possible to
led to several improvements throughout the article. Our work
end up developing clever machine learning system for clinical
was financially supported by the Bergen Research Founda-
use that turn out to be practically useless for actual clini-
tion through the project “Computational medical imaging and
cians. Attempting to augment already established procedures
machine learning – methods, infrastructure and applications”.
necessitates knowledge of the entire workflow. Involving the
end-user in the process of creating and evaluating systems can
make this a little less of an issue, and can also increase the end References
users’ trust in the systems,48 as you can establish a feedback
loop during the development process. But still, even if there is [1] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with
interest on the “ground floor” and one is able to get prototype deep convolutional neural networks. In: Pereira F, Burges CJC, Bottou
L, Weinberger KQ, editors. Advances in neural information processing
systems into the hands of clinicians, there are many higher- systems 25. Curran Associates, Inc.; 2012. p. 1097–105.
ups to convince and regulatory, ethical and legal hurdles to [2] Peters M, Neumann M, Iyyer M, Gardner M, Clark C, Lee K, et al.
overcome. Deep contextualized word representations. In: Proceedings of the 2018
conference of the North American Chapter of the Association for Com-
putational Linguistics: Human Language Technologies (long papers),
5.4 Perspectives and future expectations vol. 1. 2018. p. 2227–37.
[3] Howard J, Ruder S. Universal language model fine-tuning for text
Deep learning in medical data analysis is here to stay. Even classification. In: Proceedings of the 56th annual meeting of the Asso-
though there are many challenges associated to the intro- ciation for Computational Linguistics (volume 1: long papers). 2018.
duction of deep learning in clinical settings, the methods p. 328–39.
[4] Radford A, Narasimhan K, Salimans T, Sutskever I. Improving lan-
produce results that are too valuable to discard. This is illus- guage understanding by generative pre-training; 2018.
trated by the tremendous amounts of high-impact publications [5] Xiong W, Wu L, Alleva F, Droppo J, Huang X, Stolcke A. The
in top-journals dealing with deep learning in medical imag- Microsoft 2017 conversational speech recognition system. In: Proc.
ing (for example [16,20,29,31,39,89,136,139,161,272,284], speech and signal processing (ICASSP) 2018 IEEE int. conf. acoustics.
2018. p. 5934–8.
all published in 2018). As machine learning researchers and
[6] van den Oord A, Dieleman S, Zen H, Simonyan K, Vinyals O,
practitioners gain more experience, it will become easier to Graves A, et al. WaveNet: a generative model for raw audio; 2016,
classify problems according to what solution approach is the arXiv:1609.03499v2.
most reasonable: (i) best approached using deep learning tech- [7] Guo C, Berkhahn F. Entity embeddings of categorical variables; 2016,
niques end-to-end, (ii) best tackled by a combination of deep arXiv:1604.06737.
47 https://www.darpa.mil/program/explainable-artificial-intelligence. 49 In-line with the ideas of the convergence of disciplines and the “future of
48 This is the approach we have taken at our MMIV center https://mmiv.no, health”, as described in [358].
located inside the Department of Radiology. 50 http://p4mi.org.
[8] De Brébisson A, Simon É, Auvolat A, Vincent P, Bengio Y. Artificial [30] Poplin R, Chang P-C, Alexander D, Schwartz S, Colthurst T, Ku A,
neural networks applied to taxi destination prediction; 2015, arXiv et al. A universal SNP and small-indel variant caller using deep neural
preprint arXiv:1508.00021. networks. Nat Biotechnol 2018.
[9] George D, Huerta E. Deep learning for real-time gravitational wave [31] De Fauw J, Ledsam JR, Romera-Paredes B, Nikolov S, Tomasev N,
detection and parameter estimation: results with advanced LIGO data. Blackwell S, et al. Clinically applicable deep learning for diagnosis
Phys Lett B 2018;778:64–70. and referral in retinal disease. Nat Med 2018;24:1342.
[10] George D, Shen H, Huerta E. Classification and unsupervised clus- [32] Qin Y, Kamnitsas K, Ancha S, Nanavati J, Cottrell G, Criminisi A,
tering of LIGO data with deep transfer learning. Phys Rev D et al. Autofocus layer for semantic segmentation, arXiv:1805.08403
2018;97:101501. (2018).
[11] Shen H, George D, Huerta E, Zhao Z. Denoising gravitational [33] Kamnitsas K, Baumgartner C, Ledig C, Newcombe V, Simpson J, Kane
waves using deep learning with recurrent denoising autoencoders, A, et al. Unsupervised domain adaptation in brain lesion segmentation
arXiv:1711.09919 (2017). with adversarial networks. In: International conference on information
[12] Raissi M, Karniadakis GE. Hidden physics models: machine learn- processing in medical imaging. 2017. p. 597–609.
ing of nonlinear partial differential equations. J Comput Phys [34] Xiao C, Choi E, Sun J. Opportunities and challenges in developing
2018;357:125–41. deep learning models using electronic health records data: a systematic
[13] Karpatne A, Atluri G, Faghmous JH, Steinbach M, Banerjee A, Gan- review. J Am Med Inform Assoc 2018.
guly A, et al. Theory-guided data science: a new paradigm for scientific [35] Silver D, Schrittwieser J, Simonyan K, Antonoglou I, Huang A, Guez
discovery from data. IEEE Trans Knowl Data Eng 2017;29:2318–31. A, et al. Mastering the game of Go without human knowledge. Nature
[14] Gartner. Top strategic technology trends for 2018; 2018. 2017;550:354.
[15] Ravi D, Wong C, Deligianni F, Berthelot M, Andreu-Perez J, Lo B, [36] Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al.
et al. Deep learning for health informatics. IEEE J Biomed Health Dermatologist-level classification of skin cancer with deep neural net-
Inform 2017;21:4–21. works. Nature 2017;542:115–8.
[16] Ganapathy N, Swaminathan R, Deserno TM. Deep learning on [37] Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado
1-D biosignals: a taxonomy-based survey. Yearbook Med Inform GS, et al. Prediction of cardiovascular risk factors from retinal fundus
2018;27:98–109. photographs via deep learning. Nat Biomed Eng 2018;2:158.
[17] Kuhlmann L, Lehnertz K, Richardson MP, Schelter B, Zaveri HP. [38] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:
Seizure prediction – ready for a new era. Nat Rev Neurol 2018. 436.
[18] Kwon J-M, Lee Y, Lee Y, Lee S, Park J. An algorithm based on deep [39] Hinton G. Deep learning a technology with the potential to transform
learning for predicting in-hospital cardiac arrest. J Am Heart Assoc health care; 2018. p. 1–2.
2018;7. [40] Goodfellow I, Bengio Y, Courville A. Deep learning. MIT Press; 2016.
[19] Shin H-C, Roth HR, Gao M, Lu L, Xu Z, Nogues I, et al. Deep www.deeplearningbook.org.
convolutional neural networks for computer-aided detection: CNN [41] Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian
architectures, dataset characteristics and transfer learning. IEEE Trans M, et al. A survey on deep learning in medical image analysis. Med
Med Imaging 2016;35:1285–98. Image Anal 2017;42:60–88.
[20] Kermany DS, Goldbaum M, Cai W, Valentim CCS, Liang H, Bax- [42] Marblestone AH, Wayne G, Kording KP. Toward an integration of
ter SL, et al. Identifying medical diagnoses and treatable diseases by deep learning and neuroscience. Front Comput Neurosci 2016;10:94.
image-based deep learning. Cell 2018;172:1122–31, e9. [43] Hassabis D, Kumaran D, Summerfield C, Botvinick M. Neuroscience-
[21] Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. inspired artificial intelligence. Neuron 2017;95:245–58.
DeepSurv: personalized treatment recommender system using a Cox [44] Banino A, Barry C, Uria B, Blundell C, Lillicrap T, Mirowski P, et al.
proportional hazards deep neural network. BMC Med Res Methodol Vector-based navigation using grid-like representations in artificial
2018;18:24. agents. Nature 2018;557:429.
[22] Jimènez J, Škalič M, Martínez-Rosell G, De Fabritiis G. [45] Cybenko G. Approximation by superpositions of a sigmoidal function.
KDEEP: protein-ligand absolute binding affinity prediction via 3D- Math Control Signals Syst 1989;2:303–14.
Convolutional Neural Networks. J Chem Inf Model 2018;58:287–96. [46] Hornik K, Stinchcombe M, White H. Multilayer feedforward networks
[23] Kalinin AA, Higgins GA, Reamaroon N, Soroushmehr S, Allyn-Feuer are universal approximators. Neural Netw 1989;2:359–66.
A, Dinov ID, et al. Deep learning in pharmacogenomics: from gene reg- [47] Leshno M, Lin VY, Pinkus A, Schocken S. Multilayer feedforward
ulation to patient stratification. Pharmacogenomics 2018;19:629–50. networks with a nonpolynomial activation function can approximate
[24] Jiang S, Chin K-S, Tsui KL. A universal deep learning approach any function. Neural Netw 1993;6:861–7.
for modeling the flow of patients under different severities. Comput [48] Sonoda S, Murata N. Neural network with unbounded activa-
Methods Programs Biomed 2018;154:191–203. tion functions is universal approximator. Appl Comput Harm Anal
[25] Vranas KC, Jopling JK, Sweeney TE, Ramsey MC, Milstein AS, 2017;43:233–68.
Slatore CG, et al. Identifying distinct subgroups of ICU patients: a [49] Nielsen MA. Neural networks and deep learning. Determination Press;
machine learning approach. Crit Care Med 2017;45:1607–15. 2015.
[26] Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, et al. [50] Aggarwal CC. Neural networks and deep learning. Springer; 2018.
Scalable and accurate deep learning with electronic health records. [51] Rosenblatt F. The perceptron: a probabilistic model for information
NPJ Dig Med 2018;1:18. storage and organization in the brain. Psychol Rev 1958;65:386.
[27] Shickel B, Tighe PJ, Bihorac A, Rashidi P. Deep EHR: a survey [52] Linnainmaa S. The representation of the cumulative rounding error
of recent advances in deep learning techniques for electronic health of an algorithm as a Taylor expansion of the local rounding errors
record (EHR) analysis. IEEE J Biomed Health Inform 2017. [Master’s thesis]. Univ. Helsinki; 1970. p. 6–7 [in Finnish].
[28] Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy [53] Werbos P. Beyond regression: new tools for prediction and analysis
A, et al. Development and validation of a deep learning algorithm for in the behavioral sciences [Ph.D. dissertation]. Harvard University;
detection of diabetic retinopathy in retinal fundus photographs. JAMA 1974.
2016;316:2402–10. [54] Rumelhart DE, Hinton GE, Williams RJ. Learning representations by
[29] Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell M, Cor- back-propagating errors. Nature 1986;323:533.
rado G, et al. Predicting cardiovascular risk factors in retinal fundus [55] Cauchy A. Méthode générale pour la résolution des systemes déqua-
photographs using deep learning. Nat Biomed Eng 2018. tions simultanées. C R Sci Paris 1847;25:536–8.
[56] LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning M, Cortes C, Lawrence ND, Weinberger KQ, editors. Advances in
applied to document recognition. Proc IEEE 1998;86:2278–324. neural information processing systems 27. Curran Associates, Inc.;
[57] Lo SC, Freedman MT, Lin JS, Mun SK. Automatic lung nodule 2014. p. 2672–80.
detection using profile matching and back-propagation neural network [80] Koch G, Zemel R, Salakhutdinov R. Siamese neural networks for one-
techniques. J Digit Imaging 1993;6:48–54. shot image recognition. In: ICML deep learning workshop, vol. 2.
[58] Murray S. An exploratory analysis of multi-class uncertainty approx- 2015.
imation in Bayesian convolution neural networks [Master’s thesis]. [81] Bromley J, Guyon I, LeCun Y, Säckinger E, Shah R. Signature verifi-
University of Bergen; 2018. cation using a “Siamese” time delay neural network. In: Advances in
[59] Clevert D-A, Unterthiner T, Hochreiter S. Fast and accurate deep net- neural information processing systems. 1993. p. 737–44.
work learning by exponential linear units (ELUS), arXiv:1511.07289 [82] Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for
(2015). biomedical image segmentation. In: International conference on med-
[60] He K, Zhang X, Ren S, Sun J. Delving deep into rectifiers: surpassing ical image computing and computer-assisted intervention. 2015. p.
human-level performance on imagenet classification. In: Proceedings 234–41.
of the IEEE international conference on computer vision. 2015. p. [83] Long J, Shelhamer E, Darrell T. Fully convolutional networks for
1026–34. semantic segmentation. In: Proceedings of the IEEE conference on
[61] Springenberg JT, Dosovitskiy A, Brox T, Riedmiller M. Striving for computer vision and pattern recognition. 2015. p. 3431–40.
simplicity: the all convolutional net, arXiv:1412.6806 (2014). [84] Milletari F, Navab N, Ahmadi S-A. V-net: fully convolutional neural
[62] Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. networks for volumetric medical image segmentation. In: 2016 fourth
Dropout: a simple way to prevent neural networks from overfitting. J international conference on 3D Vision (3DV), IEEE. 2016. p. 565–71.
Mach Learn Res 2014;15:1929–58. [85] Redmon J. Darknet: open source neural networks in C.
[63] Rashmi K, Gilad-Bachrach R. Dart: dropouts meet multiple additive http://pjreddie.com/darknet/, 2013–2016.
regression trees. In: International conference on artificial intelligence [86] Lee YH. Efficiency improvement in a busy radiology practice: determi-
and statistics. 2015. p. 489–97. nation of musculoskeletal magnetic resonance imaging protocol using
[64] Gal Y. Uncertainty in deep learning [Ph.D. thesis]. University of Cam- deep learning convolutional neural networks. J Digit Imaging 2018.
bridge; 2016. [87] Gong E, Pauly JM, Wintermark M, Zaharchuk G. Deep learning
[65] Wickstrøm K, Kampffmeyer M, Jenssen R. Uncertainty modeling and enables reduced gadolinium dose for contrast-enhanced brain MRI.
interpretability in convolutional neural networks for polyp segmenta- J Magn Reson Imaging 2018;48:330–40.
tion. In: 2018 IEEE 28th international workshop on machine learning [88] Meyer P, Noblet V, Mazzara C, Lallement A. Survey on deep learning
for signal processing (MLSP), IEEE. 2018. p. 1–6. for radiotherapy. Comput Biol Med 2018;98:126–46.
[66] Ioffe S, Szegedy C. Batch normalization: accelerating deep network [89] Liu F, Jang H, Kijowski R, Bradshaw T, McMillan AB. Deep learn-
training by reducing internal covariate shift. In: International confer- ing MR imaging-based attenuation correction for PET/MR imaging.
ence on machine learning. 2015. p. 448–56. Radiology 2018;286:676–84.
[67] Zeiler MD, Fergus R. Visualizing and understanding convolutional net- [90] Mehranian A, Arabi H, Zaidi H. Vision 20/20: magnetic resonance
works. In: European conference on computer vision, Springer. 2014. imaging-guided attenuation correction in PET/MRI: challenges, solu-
p. 818–33. tions, and opportunities. Med Phys 2016;43:1130–55.
[68] Simonyan K, Zisserman A. Very deep convolutional networks for [91] Lao J, Chen Y, Li Z-C, Li Q, Zhang J, Liu J, et al. A deep learning-based
large-scale image recognition, arXiv:1409.1556 (2014). radiomics model for prediction of survival in glioblastoma multiforme.
[69] Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, et al. Going Sci Rep 2017;7:10353.
deeper with convolutions. In: Proceedings of the IEEE conference on [92] Oakden-Rayner L, Carneiro G, Bessen T, Nascimento JC, Bradley
computer vision and pattern recognition. 2014. p. 1–9. AP, Palmer LJ. Precision radiology: predicting longevity using feature
[70] Lin M, Chen Q, Yan S. Network in network, arXiv:1312.4400 (2013). engineering and deep learning methods in a radiomics framework. Sci
[71] He K, Zhang X, Ren S, Sun J. Deep residual learning for image recog- Rep 2017;7:1648.
nition. In: Proceedings of the IEEE conference on computer vision and [93] Peeken JC, Bernhofer M, Wiestler B, Goldberg T, Cremers D, Rost B,
pattern recognition. 2015. p. 770–8. et al. Radiomics in radiooncology – challenging the medical physicist.
[72] Srivastava RK, Greff K, Schmidhuber J. Training very deep net- Phys Med 2018;48:27–36.
works. In: Advances in neural information processing systems. 2015. [94] Izadyyazdanabadi M, Belykh E, Mooney MA, Eschbacher JM, Nakaji
p. 2377–85. P, Yang Y, et al. Prospects for theranostics in neurosurgical imag-
[73] Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely con- ing: empowering confocal laser endomicroscopy diagnostics via deep
nected convolutional networks. In: CVPR, vol. 1. 2016. p. 3. learning. Front Oncol 2018;8:240.
[74] Xie S, Girshick R, Dollár P, Tu Z, He K. Aggregated residual trans- [95] Haskins G, Kruecker J, Kruger U, Xu S, Pinto PA, Wood BJ, Yan
formations for deep neural networks. In: 2017 IEEE conference on P. Learning deep similarity metric for 3D MR-TRUS registration,
computer vision and pattern recognition (CVPR), IEEE. 2016. p. arXiv:1806.04548v1 (2018).
5987–95. [96] Cao X, Yang J, Zhang J, Wang Q, Yap P-T, Shen D. Deformable image
[75] Hu J, Shen L, Sun G. Squeeze-and-excitation networks, registration using a cue-aware deep regression network. IEEE Trans
arXiv:1709.01507 (2017). Bio-med Eng 2018;65:1900–11.
[76] Zoph B, Vasudevan V, Shlens J, Le QV. Learning transferable archi- [97] Yang X, Kwitt R, Styner M, Niethammer M. Quicksilver: Fast pre-
tectures for scalable image recognition, arXiv:1707.07012 2 (2017). dictive image registration – a deep learning approach. Neuroimage
[77] Bello I, Zoph B, Vasudevan V, Le QV. Neural optimizer search with 2017;158:378–96.
reinforcement learning. In: Precup D, Teh YW, editors. Proceedings [98] Kearney VP, Haaf S, Sudhyadhom A, Valdes G, Solberg D. An unsu-
of the 34th international conference on machine learning. Proceed- pervised convolutional neural network-based algorithm for deformable
ings of machine learning research, vol. 70. Sydney, Australia: PMLR, image registration. Phys Med Biol 2018.
International Convention Centre; 2017. p. 459–68. [99] Zheng J, Miao S, Jane Wang Z, Liao R. Pairwise domain adapta-
[78] Redmon J, Divvala S, Girshick R, Farhadi A. You only look once: uni- tion module for CNN-based 2-D/3-D registration. J Med Imaging
fied, real-time object detection. In: Proceedings of the IEEE conference (Bellingham, Wash.) 2018;5:021204.
on computer vision and pattern recognition. 2015. p. 779–88. [100] Anas EMA, Mousavi P, Abolmaesumi P. A deep learning approach for
[79] Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, real time prostate segmentation in freehand ultrasound guided biopsy.
Ozair S, et al. Generative adversarial nets. In: Ghahramani Z, Welling Med Image Anal 2018;48:107–16.
[101] Ching T, Himmelstein DS, Beaulieu-Jones BK, Kalinin A, Do BT, Way [124] Xue Y, Chen S, Qin J, Liu Y, Huang B, Chen H. Application of deep
GP, et al. Opportunities and obstacles for deep learning in biology and learning in automated analysis of molecular images in cancer: a survey.
medicine. J R Soc Interface 2018;15. Contrast Media Mol Imaging 2017;2017:9512370.
[102] Lee J-G, Jun S, Cho Y-W, Lee H, Kim GB, Seo JB, et al. Deep [125] Brattain LJ, Telfer BA, Dhyani M, Grajo JR, Samir AE. Machine learn-
learning in medical imaging: general overview. Korean J Radiol ing for medical ultrasound: status, methods, and future opportunities.
2017;18:570–84. Abdom Radiol 2018;43:786–99.
[103] Rueckert D, Glocker B, Kainz B. Learning clinically useful infor- [126] Huang Q, Zhang F, Li X. Machine learning in ultrasound
mation from images: past, present and future. Med Image Anal computer-aided diagnostic systems: a survey. BioMed Res Int
2016;33:13–8. 2018;2018:5137904.
[104] Chartrand G, Cheng PM, Vorontsov E, Drozdzal M, Turcotte S, Pal [127] Shen D, Wu G, Suk H-I. Deep learning in medical image analysis.
CJ, et al. Deep learning: a primer for radiologists. Radiographics Annu Rev Biomed Eng 2017;19:221–48.
2017;37:2113–31. [128] Suzuki K. Overview of deep learning in medical imaging. Radiol Phys
[105] Erickson BJ, Korfiatis P, Akkus Z, Kline TL. Machine learning for Technol 2017;10:257–73.
medical imaging. Radiographics 2017;37:505–15. [129] Cao C, Liu F, Tan H, Song D, Shu W, Li W, et al. Deep learning
[106] Mazurowski MA, Buda M, Saha A, Bashir MR. Deep learning in and its applications in biomedicine. Genomics Proteomics Bioinform
radiology: an overview of the concepts and a survey of the state of the 2018;16:17–32.
art, arXiv:1802.08717v1 (2018). [130] Lakhani P, Gray DL, Pett CR, Nagy P, Shih G. Hello world deep
[107] McBee MP, Awan OA, Colucci AT, Ghobadi CW, Kadom N, Kansagra learning in medical imaging. J Digit Imaging 2018.
AP, et al. Deep learning in radiology. Acad Radiol 2018. [131] Pawlowski N, Ktena SI, Lee MC, Kainz B, Rueckert D, Glocker
[108] Savadjiev P, Chong J, Dohan A, Vakalopoulou M, Reinhold C, Paragios B, et al. DLTK: state of the art reference implementations for deep
N, et al. Demystification of AI-driven medical image interpretation: learning on medical images, arXiv:1711.06853 (2017).
past, present and future. Eur Radiol 2018. [132] Yang Y, Sun J, Li H, Xu Z. Deep ADMM-Net for compressive sensing
[109] Thrall JH, Li X, Li Q, Cruz C, Do S, Dreyer K, et al. Artificial MRI. In: Lee DD, Sugiyama M, Luxburg UV, Guyon I, Garnett R,
intelligence and machine learning in radiology: opportunities, chal- editors. Advances in neural information processing systems 29. Curran
lenges, pitfalls, and criteria for success. J Am Coll Radiol: JACR Associates, Inc.; 2016. p. 10–8.
2018;15:504–8. [133] Wang S, Su Z, Ying L, Peng X, Zhu S, Liang F, et al. Accelerating
[110] Yamashita R, Nishio M, Do RKG, Togashi K. Convolutional neural magnetic resonance imaging via deep learning. In: 2016 IEEE 13th
networks: an overview and application in radiology. Insights Imaging international symposium on biomedical imaging (ISBI), IEEE. 2016.
2018. p. 514–7.
[111] Yasaka K, Akai H, Kunimatsu A, Kiryu S, Abe O. Deep learn- [134] Qin C, Hajnal JV, Rueckert D, Schlemper J, Caballero J, Price AN.
ing with convolutional neural network in radiology. Jpn J Radiol Convolutional recurrent neural networks for dynamic MR image
2018;36:257–72. reconstruction. IEEE Trans Med Imaging 2018.
[112] Giger ML. Machine learning in medical imaging. J Am Coll Radiol: [135] Schlemper J, Caballero J, Hajnal JV, Price AN, Rueckert D. A deep
JACR 2018;15:512–20. cascade of convolutional neural networks for dynamic MR image
[113] Erickson BJ, Korfiatis P, Akkus Z, Kline T, Philbrick K. Toolkits and reconstruction. IEEE Trans Med Imaging 2018;37:491–503.
libraries for deep learning. J Digit Imaging 2017;30:400–5. [136] Chen F, Taviani V, Malkiel I, Cheng JY, Tamir JI, Shaikh J, et al.
[114] Zaharchuk G, Gong E, Wintermark M, Rubin D, Langlotz CP. Deep Variable-density single-shot fast Spin-Echo MRI with deep learning
learning in neuroradiology. AJNR Am J Neuroradiol 2018. reconstruction by using variational networks. Radiology 2018:180445.
[115] Akkus Z, Galimzianova A, Hoogi A, Rubin DL, Erickson BJ. Deep [137] Knoll F, Hammernik K, Kobler E, Pock T, Recht MP, Sodickson DK.
learning for brain MRI segmentation: state of the art and future direc- Assessment of the generalization of learned image reconstruction and
tions. J Digit Imaging 2017;30:449–59. the potential for transfer learning. Magn Reson Med 2018.
[116] Lee E-J, Kim Y-H, Kim N, Kang D-W. Deep into the brain: artificial [138] Mardani M, Gong E, Cheng JY, Vasanawala SS, Zaharchuk G, Xing
intelligence in stroke imaging. J Stroke 2017;19:277–85. L, et al. Deep generative adversarial neural networks for compressive
[117] Feng R, Badgeley M, Mocco J, Oermann EK. Deep learning guided sensing (GANCS) MRI. IEEE Trans Med Imaging 2018.
stroke management: a review of clinical applications. J Neurointervent [139] Zhu B, Liu JZ, Cauley SF, Rosen BR, Rosen MS. Image reconstruction
Surg 2018;10:358–62. by domain-transform manifold learning. Nature 2018;555:487–92.
[118] Vieira S, Pinaya WHL, Mechelli A. Using deep learning to investigate [140] Eo T, Jun Y, Kim T, Jang J, Lee H-J, Hwang D. KIKI-net: cross-domain
the neuroimaging correlates of psychiatric and neurological disorders: convolutional neural networks for reconstructing undersampled mag-
methods and applications. Neurosci Biobehav Rev 2017;74:58–75. netic resonance images. Magn Reson Med 2018;80:2188–201.
[119] Burt JR, Torosdagli N, Khosravan N, RaviPrakash H, Mortazi A, Tis- [141] Han Y, Yoo J, Kim HH, Shin HJ, Sung K, Ye JC. Deep learning
savirasingham F, et al. Deep learning beyond cats and dogs: recent with domain adaptation for accelerated projection-reconstruction MR.
advances in diagnosing breast cancer with deep neural networks. Br J Magn Reson Med 2018;80:1189–205.
Radiol 2018;91:20170545. [142] Shi J, Liu Q, Wang C, Zhang Q, Ying S, Xu H. Super-resolution
[120] Samala RK, Chan H-P, Hadjiiski LM, Helvie MA, Cha KH, Richter reconstruction of MR image with a novel residual learning network
CD. Multi-task transfer learning deep convolutional neural network: algorithm. Phys Med Biol 2018;63:085011.
application to computer-aided diagnosis of breast cancer on mammo- [143] Yang G, Yu S, Dong H, Slabaugh G, Dragotti PL, Ye X, et al.
grams. Phys Med Biol 2017;62:8894–908. DAGAN: deep de-aliasing generative adversarial networks for fast
[121] van Ginneken B. Fifty years of computer analysis in chest imag- compressed sensing MRI reconstruction. IEEE Trans Med Imaging
ing: rule-based, machine learning, deep learning. Radiol Phys Technol 2018;37:1310–21.
2017;10:23–32. [144] Deistung A, Schäfer A, Schweser F, Biedermann U, Turner R,
[122] Morin O, Vallires M, Jochems A, Woodruff HC, Valdes G, Braunstein Reichenbach JR. Toward in vivo histology: a comparison of quan-
SE, et al. A deep look into the future of quantitative imaging in oncol- titative susceptibility mapping (QSM) with magnitude-, phase-,
ogy: a statement of working principles and proposal for change. Int J and r2-imaging at ultra-high magnetic field strength. Neuroimage
Radiat Oncol Biol Phys 2018. 2013;65:299–314.
[123] Parmar C, Barry JD, Hosny A, Quackenbush J, Aerts JWL. Data anal- [145] Deistung A, Schweser F, Reichenbach JR. Overview of quantitative
ysis strategies in medical imaging. Clin Cancer Res 2018;24:3492–9. susceptibility mapping. NMR Biomed 2017;30:e3569.
[146] Yoon J, Gong E, Chatnuntawech I, Bilgic B, Lee J, Jung W, et al. Quan- [168] Virtue P, Yu SX, Lustig M. Better than real: complex-valued neural
titative susceptibility mapping using deep neural network: QSMnet. nets for MRI fingerprinting. In: Proc. IEEE int. conf. image processing
Neuroimage 2018;179:199–206. (ICIP). 2017. p. 3953–7.
[147] Liu T, Spincemaille P, De Rochefort L, Kressler B, Wang Y. [169] Tygert M, Bruna J, Chintala S, LeCun Y, Piantino S, Szlam A. A
Calculation of susceptibility through multiple orientation sampling mathematical motivation for complex-valued convolutional networks.
(COSMOS): a method for conditioning the inverse problem from mea- Neural Comput 2016;28:815–25.
sured magnetic field map to susceptibility source image in MRI. Magn [170] Trabelsi C, Bilaniuk O, Zhang Y, Serdyuk D, Subramanian S, Santos
Reson Med 2009;61:196–204. JF, et al. Deep complex networks, arXiv:1705.09792 (2017).
[148] Rasmussen KGB, Kristensen MJ, Blendal RG, Ostergaard LR, [171] Sijbers J, den Dekker AJ, Van Audekerke J, Verhoye M, Van Dyck
Plocharski M, O’Brien K, et al. DeepQSM-using deep learning to D. Estimation of the noise in magnitude MR images. Magn Reson
solve the dipole inversion for MRI susceptibility mapping. Biorxiv Imaging 1998;16:87–90.
2018:278036. [172] McVeigh ER, Henkelman RM, Bronskill MJ. Noise and filtration in
[149] Ma D, Gulani V, Seiberlich N, Liu K, Sunshine JL, Duerk JL, et al. magnetic resonance imaging. Med Phys 1985;12:586–91.
Magnetic resonance fingerprinting. Nature 2013;495:187–92. [173] Baselice F, Ferraioli G, Pascazio V, Sorriso A. Bayesian MRI denoising
[150] E. S. of Radiology (ESR). Magnetic resonance fingerprinting – a in complex domain. Magn Reson Imaging 2017;38:112–22.
promising new approach to obtain standardized imaging biomarkers [174] Phophalia A, Mitra SK. 3d MR image denoising using rough set and
from MRI. Insights Imaging 2015;6:163–5. kernel PCA method. Magn Reson Imaging 2017;36:135–45.
[151] Donoho DL. Compressed sensing. IEEE Trans Inf Theory [175] Zhang X, Xu Z, Jia N, Yang W, Feng Q, Chen W, et al. Denoising of
2006;52:1289–306. 3D magnetic resonance images by using higher-order singular value
[152] Lustig M, Donoho D, Pauly JM. Sparse MRI: the application decomposition. Med Image Anal 2015;19:75–86.
of compressed sensing for rapid MR imaging. Magn Reson Med [176] Van De Ville D, Seghier ML, Lazeyras F, Blu T, Unser M.
2007;58:1182–95. Wspm: wavelet-based statistical parametric mapping. Neuroimage
[153] McCann MT, Jin KH, Unser M. Convolutional neural networks for 2007;37:1205–17.
inverse problems in imaging: a review. IEEE Signal Process Mag [177] Salimi-Khorshidi G, Douaud G, Beckmann CF, Glasser MF, Griffanti
2017;34:85–95. L, Smith SM. Automatic denoising of functional MRI data: combining
[154] Shah V, Hegde C. Solving linear inverse problems using GAN priors: independent component analysis and hierarchical fusion of classifiers.
an algorithm with provable guarantees, arXiv:1802.08406 (2018). Neuroimage 2014;90:449–68.
[155] Lucas A, Iliadis M, Molina R, Katsaggelos AK. Using deep neural [178] Lysaker M, Lundervold A, Tai X-C. Noise removal using fourth-order
networks for inverse problems in imaging: beyond analytical methods. partial differential equation with applications to medical magnetic
IEEE Signal Process Mag 2018;35:20–36. resonance images in space and time. IEEE Trans Image Process
[156] Aggarwal HK, Mani MP, Jacob M. Modl: model based deep learning 2003;12:1579–90.
architecture for inverse problems. IEEE Trans Med Imaging 2018. [179] Bermudez C, Plassard AJ, Davis TL, Newton AT, Resnick SM, Land-
[157] Li H, Schwab J, Antholzer S, Haltmeier M. Nett: solving inverse man BA. Learning implicit brain MRI manifolds with deep learning.
problems with deep neural networks, arXiv:1803.00092 (2018). Proc SPIE 2018;10574.
[158] Ma D, Jiang Y, Chen Y, McGivney D, Mehta B, Gulani V, et al. Fast 3D [180] Benou A, Veksler R, Friedman A, Riklin Raviv T. Ensemble of
magnetic resonance fingerprinting for a whole-brain coverage. Magn expert deep neural networks for spatiotemporal denoising of contrast-
Reson Med 2018;79:2190–7. enhanced MRI sequences. Med Image Anal 2017;42:145–59.
[159] Christen T, Pannetier NA, Ni WW, Qiu D, Moseley ME, Schuff N, [181] Gal Y, Mehnert AJH, Bradley AP, McMahon K, Kennedy D, Crozier S.
et al. MR vascular fingerprinting: a new approach to compute cerebral Denoising of dynamic contrast-enhanced MR images using dynamic
blood volume, mean vessel radius, and oxygenation maps in the human non-local means. IEEE Trans Med Imaging 2010;29:302–10.
brain. Neuroimage 2014;89:262–70. [182] Vincent P, Larochelle H, Lajoie I, Bengio Y, Manzagol P-A. Stacked
[160] Lemasson B, Pannetier N, Coquery N, Boisserand LSB, Collomb N, denoising autoencoders: learning useful representations in a deep net-
Schuff N, et al. MR vascular fingerprinting in stroke and brain tumors work with a local denoising criterion. J Mach Learn Res (JMLR)
models. Sci Rep 2016;6:37071. 2010;11:3371–408.
[161] Rieger B, Akakaya M, Pariente JC, Llufriu S, Martinez-Heras E, [183] Dikaios N, Arridge S, Hamy V, Punwani S, Atkinson D. Direct para-
Weingrtner S, et al. Time efficient whole-brain coverage with MR metric reconstruction from undersampled (k,t)-space data in dynamic
fingerprinting using slice-interleaved echo-planar imaging. Sci Rep contrast enhanced MRI. Med Image Anal 2014;18:989–1001.
2018;8:6667. [184] Guo Y, Lingala SG, Zhu Y, Lebel RM, Nayak KS. Direct estima-
[162] Wright KL, Jiang Y, Ma D, Noll DC, Griswold MA, Gulani V, et al. tion of tracer-kinetic parameter maps from highly undersampled brain
Estimation of perfusion properties with MR fingerprinting arterial spin dynamic contrast enhanced MRI. Magn Reson Med 2017;78:1566–78.
labeling. Magn Reson Imaging 2018;50:68–77. [185] Sourbron SP, Buckley DL. Classic models for dynamic contrast-
[163] Panda A, Mehta BB, Coppo S, Jiang Y, Ma D, Seiberlich N, et al. enhanced MRI. NMR Biomed 2013;26:1004–27.
Magnetic resonance fingerprinting-an overview. Curr Opin Biomed [186] Golkov V, Dosovitskiy A, Sperl JI, Menzel MI, Czisch M, Samann
Eng 2017;3:56–66. P, et al. q-space deep learning: twelve-fold shorter and model-free
[164] Cohen O, Zhu B, Rosen MS. MR fingerprinting deep reconstruction diffusion MRI scans. IEEE Trans Med Imaging 2016;35:1344–51.
network (DRONE). Magn Reson Med 2018;80:885–94. [187] Gurbani SS, Schreibmann E, Maudsley AA, Cordova JS, Soher BJ,
[165] Hoppe E, Krzdrfer G, Wrfl T, Wetzl J, Lugauer F, Pfeuffer J, et al. Poptani H, et al. A convolutional neural network to filter artifacts in
Deep learning for magnetic resonance fingerprinting: a new approach spectroscopic MRI. Magn Reson Med 2018;80:1765–75.
for predicting quantitative parameter values from time series. Stud [188] Kyathanahally SP, Dring A, Kreis R. Deep learning approaches for
Health Technol Inform 2017;243:202–6. detection and removal of ghosting artifacts in MR spectroscopy. Magn
[166] Bojorquez JZ, Bricq S, Acquitter C, Brunotte F, Walker PM, Lalande Reson Med 2018;80:851–63.
A. What are normal relaxation times of tissues at 3T? Magn Reson [189] Küstner T, Liebgott A, Mauch L, Martirosian P, Bamberg F, Nikolaou
Imaging 2017;35:69–80. K, et al. Automated reference-free detection of motion artifacts in
[167] Fang Z, Chen Y, Lin W, Shen D. Quantification of relaxation times in magnetic resonance images. MAGMA 2018;31:243–56.
MR fingerprinting using deep learning. In: Proceedings of the Interna- [190] Yue L, Shen H, Li J, Yuan Q, Zhang H, Zhang L. Image super-
tional Society for Magnetic Resonance in Medicine. Scientific meeting resolution: the techniques, applications, and future. Signal Process
and exhibition 25. 2017. 2016;128:389–408.
[191] Shilling RZ, Robbie TQ, Bailloeul T, Mewes K, Mersereau RM, Brum- anonymization using generative adversarial networks. In: International
mer ME. A superresolution framework for 3-D high-resolution and workshop on simulation and synthesis in medical imaging. 2018. p.
high-contrast imaging using 2-D multislice MRI. IEEE Trans Med 1–11.
Imaging 2009;28:633–44. [213] Mok TCW, Chung ACS. Learning data augmentation for brain tumor
[192] Ropele S, Ebner F, Fazekas F, Reishofer G. Super-resolution MRI segmentation with coarse-to-fine generative adversarial networks,
using microscopic spatial modulation of magnetization. Magn Reson arXiv:1805.11291 (2018).
Med 2010;64:1671–5. [214] Guibas JT, Virdi TS, Li PS. Synthetic medical images from dual gen-
[193] Plenge E, Poot DHJ, Bernsen M, Kotek G, Houston G, Wielopolski P, erative adversarial networks, arXiv:1709.01872 (2017).
et al. Super-resolution methods in MRI: can they improve the trade-off [215] Kitchen A, Seah J. Deep generative adversarial neural networks for
between resolution, signal-to-noise ratio, and acquisition time? Magn realistic prostate lesion MRI synthesis, arXiv:1708.00129 (2017).
Reson Med 2012;68:1983–93. [216] Nie D, Trullo R, Lian J, Petitjean C, Ruan S, Wang Q, et al. Medical
[194] Bahrami K, Shi F, Rekik I, Gao Y, Shen D. 7T-guided super-resolution image synthesis with context-aware generative adversarial networks,
of 3T MRI. Med Phys 2017;44:1661–77. Medical image computing and computer-assisted intervention: MIC-
[195] Van Steenkiste G, Poot DHJ, Jeurissen B, den Dekker AJ, Vanhevel CAI. In: International conference on medical image computing and
F, Parizel PM, et al. Super-resolution T1 estimation: quantitative high computer-assisted intervention 10435. 2017. p. 417–25.
resolution T1 mapping from a set of low resolution T1 -weighted images [217] Spuhler KD, Gardus J, Gao Y, DeLorenzo C, Parsey C, Huang.
with different slice orientations. Magn Reson Med 2017;77:1818–30. Synthesis of patient-specific transmission image for PET attenuation
[196] Zeng K, Zheng H, Cai C, Yang Y, Zhang K, Chen Z. Simultaneous correction for PET/MR imaging of the brain using a convolutional
single-and multi-contrast super-resolution for brain MRI images based neural network. J Nucl Med 2018.
on a convolutional neural network. Comput Biol Med 2018;99:133–41. [218] Torrado-Carvajal A, Vera-Olmos J, Izquierdo-Garcia D, Catalano OA,
[197] Liu C, Wu X, Yu X, Tang Y, Zhang J, Zhou J. Fusing multi-scale Morales MA, Margolin J, et al. Dixon-VIBE deep learning (DIVIDE)
information in convolution network for MR image super-resolution pseudo-CT synthesis for pelvis PET/MR attenuation correction. J Nucl
reconstruction. Biomed Eng Online 2018;17:114. Med 2018.
[198] Chaudhari AS, Fang Z, Kogan F, Wood J, Stevens KJ, Gibbons EK, [219] Zhang Q, Wang H, Lu H, Won D, Yoon SW. Medical image synthesis
et al. Super-resolution musculoskeletal MRI using deep learning. with generative adversarial networks for tissue recognition. In: Proc.
Magn Reson Med 2018;80:2139–54. IEEE int. conf. healthcare informatics (ICHI). 2018. p. 199–207.
[199] Jog A, Carass A, Roy S, Pham DL, Prince JL. Random forest [220] Frid-Adar M, Klang E, Amitai M, Goldberger J, Greenspan H.
regression for magnetic resonance image synthesis. Med Image Anal Synthetic data augmentation using GAN for improved liver lesion clas-
2017;35:475–88. sification. In: Proc. IEEE 15th int. symp. biomedical imaging (ISBI
[200] Keenan KE, Ainslie M, Barker AJ, Boss MA, Cecil KM, Charles C, 2018). 2018. p. 289–93.
et al. Quantitative magnetic resonance imaging phantoms: a review [221] Wolterink JM, Dinkla AM, Savenije MHF, Seevinck PR, van den
and the need for a system phantom. Magn Reson Med 2018;79:48–61. Berg CAT, Isgum I. Deep MR to CT synthesis using unpaired data,
[201] Jurczuk K, Kretowski M, Eliat P-A, Saint-Jalmes H, Bezy-Wendling arXiv:1708.01155v1 (2017).
J. In silico modeling of magnetic resonance flow imaging in complex [222] Fitzpatrick JM, Maurer Jr CR. A review of medical image registration;
vascular networks. IEEE Trans Med Imaging 2014;33:2191–209. 1993.
[202] Zhou Y, Giffard-Roisin S, De Craene M, Camarasu-Pop S, D’Hooge [223] Maclaren J, Herbst M, Speck O, Zaitsev M. Prospective motion cor-
J, Alessandrini M, et al. A framework for the generation of real- rection in brain imaging: a review. Magn Reson Med 2013;69:621–36.
istic synthetic cardiac ultrasound and magnetic resonance imaging [224] Zaitsev M, Akin B, LeVan P, Knowles BR. Prospective motion cor-
sequences from the same virtual patients. IEEE Trans Med Imaging rection in functional MRI. Neuroimage 2017;154:33–42.
2018;37:741–54. [225] Fluck O, Vetter C, Wein W, Kamen A, Preim B, Westermann R. A
[203] Duchateau N, Sermesant M, Delingette H, Ayache N. Model-based survey of medical image registration on graphics hardware. Comput
generation of large databases of cardiac images: synthesis of patho- Methods Programs Biomed 2011;104:e45–57.
logical cine MR sequences from real healthy cases. IEEE Trans Med [226] Shi L, Liu W, Zhang H, Xie Y, Wang D. A survey of GPU-based
Imaging 2018;37:755–66. medical image computing techniques. Quant Imaging Med Surg
[204] Creswell A, White T, Dumoulin V, Arulkumaran K, Sengupta B, 2012;2:188–206.
Bharath AA. Generative adversarial networks: an overview. IEEE [227] Eklund A, Dufort P, Forsberg D, LaConte SM. Medical image pro-
Signal Process Mag 2018;35:53–65. cessing on the GPU – past, present and future. Med Image Anal
[205] Hong Y, Hwang U, Yoo J, Yoon S. How generative adver- 2013;17:1073–94.
sarial networks and their variants work: an overview of GAN, [228] Maintz JB, Viergever MA. A survey of medical image registration.
arXiv:1711.05914v7 (2017). Med Image Anal 1998;2:1–36.
[206] Huang H, Yu PS, Wang C. An introduction to image synthesis with [229] Glocker B, Sotiras A, Komodakis N, Paragios N. Deformable medical
generative adversarial nets, arXiv:1803.04469v1 (2018). image registration: setting the state of the art with discrete methods.
[207] Osokin A, Chessel A, Salas REC, Vaggi F. GANs for biological image Annu Rev Biomed Eng 2011;13:219–44.
synthesis. In: Proc. IEEE int. conf. computer vision (ICCV). 2017. p. [230] Sotiras A, Davatzikos C, Paragios N. Deformable medical image reg-
2252–61. istration: a survey. IEEE Trans Med Imaging 2013;32:1153–90.
[208] Antipov G, Baccouche M, Dugelay J. Face aging with conditional gen- [231] Oliveira FPM, Tavares JMRS. Medical image registration: a review.
erative adversarial networks. In: Proc. IEEE int. conf. image processing Comput Methods Biomech Biomed Eng 2014;17:73–93.
(ICIP). 2017. p. 2089–93. [232] Saha PK, Strand R, Borgefors G. Digital topology and geom-
[209] Bodnar C. Text to image synthesis using generative adversarial net- etry in medical imaging: a survey. IEEE Trans Med Imaging
works, arXiv:1805.00676v1 (2018). 2015;34:1940–64.
[210] Dong H, Yu S, Wu C, Guo Y. Semantic image synthesis via adversarial [233] Viergever MA, Maintz JBA, Klein S, Murphy K, Staring M, Pluim
learning, arXiv:1707.06873v1 (2017). JPW. A survey of medical image registration – under review. Med
[211] Reed S, Akata Z, Yan X, Logeswaran L, Schiele B, Lee H. Generative Image Anal 2016;33:140–4.
adversarial text to image synthesis, arXiv:1605.05396v2 (2016). [234] Song G, Han J, Zhao Y, Wang Z, Du H. A review on medical
[212] Shin H-C, Tenenholtz NA, Rogers JK, Schwarz CG, Senjem ML, image registration as an optimization problem. Curr Med Imaging
Gunter JL, et al. Medical image synthesis for data augmentation and Rev 2017;13:274–83.
[235] Ferrante E, Paragios N. Slice-to-volume medical image registration: a [258] Torres HR, Queiros S, Morais P, Oliveira B, Fonseca JC, Vilaa JL.
survey. Med Image Anal 2017;39:101–23. Kidney segmentation in ultrasound, magnetic resonance and computed
[236] Keszei AP, Berkels B, Deserno TM. Survey of non-rigid registration tomography images: a systematic review. Comput Methods Programs
tools in medicine. J Digit Imaging 2017;30:102–16. Biomed 2018;157:49–67.
[237] Nag S. Image registration techniques: a survey, arXiv:1712.07540v1 [259] Bernal J, Kushibar K, Asfaw DS, Valverde S, Oliver A, Mart R,
(2017). et al. Deep convolutional neural networks for brain image analysis
[238] Jiang J, Trundle P, Ren J. Medical image analysis with artificial neural on magnetic resonance imaging: a review. Artif Intell Med 2018.
networks. Comput Med Imaging Graph 2010;34:617–31. [260] Moccia S, De Momi E, El Hadji S, Mattos LS. Blood vessel segmen-
[239] Wu G, Kim M, Wang Q, Munsell BC, Shen D. Scalable tation algorithms-review of methods, datasets and evaluation metrics.
high-performance image registration framework by unsupervised Comput Methods Programs Biomed 2018;158:71–91.
deep feature representations learning. IEEE Trans Biomed Eng [261] Makropoulos A, Counsell SJ, Rueckert D. A review on automatic fetal
2016;63:1505–16. and neonatal brain MRI segmentation. Neuroimage 2018;170:231–48.
[240] Salehi SSM, Khan S, Erdogmus D, Gholipour A. Real-time deep pose [262] Chen L, Bentley P, Rueckert D. Fully automatic acute ischemic lesion
estimation with geodesic loss for image-to-template rigid registration. segmentation in DWI using convolutional neural networks. NeuroIm-
IEEE Trans Med Imaging 2018. age Clin 2017;15:633–43.
[241] Toth D, Miao S, Kurzendorfer T, Rinaldi CA, Liao R, Mansi T, et al. [263] Havaei M, Davy A, Warde-Farley D, Biard A, Courville A, Bengio
3D/2D model-to-image registration by imitation learning for cardiac Y, et al. Brain tumor segmentation with deep neural networks. Med
procedures. Int J Comput Assist Radiol Surg 2018. Image Anal 2017;35:18–31.
[242] Han X. MR-based synthetic CT generation using a deep convolutional [264] Choi H, Jin KH. Fast and robust segmentation of the striatum
neural network method. Med Phys 2017;44:1408–19. using deep convolutional neural networks. J Neurosci Methods
[243] Liu M, Cheng D, Wang K, Wang Y, Initiative ADN. Multi-modality 2016;274:146–53.
cascaded convolutional neural networks for Alzheimer’s disease diag- [265] Ibragimov B, Xing L. Segmentation of organs-at-risks in head and
nosis. Neuroinformatics 2018;16:295–308. neck CT images using convolutional neural networks. Med Phys
[244] Xiang L, Qiao Y, Nie D, An L, Wang Q, Shen D. Deep auto-context 2017;44:547–57.
convolutional neural networks for standard-dose PET image estima- [266] Kline TL, Korfiatis P, Edwards ME, Blais JD, Czerwiec FS, Harris
tion from low-dose PET/MRI. Neurocomputing 2017;267:406–16. PC, et al. Performance of an artificial multi-observer deep neural net-
[245] Shan S, Yan W, Guo X, Chang EI-C, Fan Y, Xu Y. Unsuper- work for fully automated segmentation of polycystic kidneys. J Digit
vised end-to-end learning for deformable medical image registration, Imaging 2017;30:442–8.
arXiv:1711.08608v2 (2017). [267] Guo Y, Gao Y, Shen D. Deformable MR prostate segmentation via
[246] Balakrishnan G, Zhao A, Sabuncu MR, Guttag J, Dalca AV. An unsu- deep feature learning and sparse patch matching. IEEE Trans Med
pervised learning model for deformable medical image registration, Imaging 2016;35:1077–89.
arXiv:1802.02604v3 (2018). [268] Li X, Dou Q, Chen H, Fu C-W, Qi X, Belav DL, et al. 3D multi-scale
[247] de Vos BD, Berendsen FF, Viergever MA, Sokooti H, Staring M, Isgum FCN with random modality voxel dropout learning for intervertebral
I. A deep learning framework for unsupervised affine and deformable disc localization and segmentation from multi-modality MR images.
image registration, arXiv:1809.06130v1 (2018). Med Image Anal 2018;45:41–54.
[248] Vannier MW, Butterfield RL, Jordan D, Murphy WA, Levitt RG, Gado [269] Kleesiek J, Urban G, Hubert A, Schwarz D, Maier-Hein K, Bendszus
M. Multispectral analysis of magnetic resonance images. Radiology M, et al. Deep MRI brain extraction: a 3D convolutional neural network
1985;154:221–4. for skull stripping. Neuroimage 2016;129:460–9.
[249] Lundervold A, Moen K, Taxt T. Automatic recognition of normal [270] Li H, Parikh NA, He L. A novel transfer learning approach to enhance
and pathological tissue types in MR images. In: Proc. of the NOBIM deep neural network classification of brain functional connectomes.
conference. 1988. Front Neurosci 2018;12:491.
[250] Taxt T, Lundervold A, Fuglaas B, Lien H, Abeler V. Multispectral anal- [271] Zeng L-L, Wang H, Hu P, Yang B, Pu W, Shen H, et al. Multi-site diag-
ysis of uterine corpus tumors in magnetic resonance imaging. Magn nostic classification of schizophrenia using discriminant deep learning
Reson Med 1992;23:55–76. with functional connectivity MRI. EBioMedicine 2018;30:74–85.
[251] Taxt T, Lundervold A. Multispectral analysis of the brain using mag- [272] Wasserthal J, Neher P, Maier-Hein KH. Tract-Seg-fast and accurate
netic resonance imaging. IEEE Trans Med Imaging 1994;13:470–81. white matter tract segmentation. Neuroimage 2018;183:239–53.
[252] Lundervold A, Storvik G. Segmentation of brain parenchyma and [273] Cole JH, Poudel RPK, Tsagkrasoulis D, Caan MWA, Steves C, Spector
cerebrospinal fluid in multispectral magnetic resonance images. IEEE TD, et al. Predicting brain age with deep learning from raw imag-
Trans Med Imaging 1995;14:339–49. ing data results in a reliable and heritable biomarker. Neuroimage
[253] Cabezas M, Oliver A, Llad X, Freixenet J, Cuadra MB. A review 2017;163:115–24.
of atlas-based segmentation for magnetic resonance brain images. [274] Liu M, Zhang J, Adeli E, Shen D. Landmark-based deep multi-instance
Comput Methods Programs Biomed 2011;104:e158–77. learning for brain disease diagnosis. Med Image Anal 2018;43:157–68.
[254] Garca-Lorenzo D, Francis S, Narayanan S, Arnold DL, Collins DL. [275] Islam J, Zhang Y. Brain MRI analysis for Alzheimer’s disease diagno-
Review of automatic segmentation methods of multiple sclerosis white sis using an ensemble system of deep convolutional neural networks.
matter lesions on conventional magnetic resonance imaging. Med Brain Inform 2018;5:2.
Image Anal 2013;17:1–18. [276] Lu D, Popuri K, Ding GW, Balachandar R, Beg MF. Multimodal
[255] Smistad E, Falch TL, Bozorgi M, Elster AC, Lindseth F. Medical image and multiscale deep neural networks for the early diagnosis of
segmentation on GPUs – a comprehensive review. Med Image Anal Alzheimer’s disease using structural MR and FDG-PET images. Sci
2015;20:1–18. Rep 2018;8:5697.
[256] Bernal J, Kushibar K, Asfaw DS, Valverde S, Oliver A, Mart R, et al. [277] Moeskops P, de Bresser J, Kuijf HJ, Mendrik AM, Biessels GJ, Pluim
Deep convolutional neural networks for brain image analysis on mag- JPW, et al. Evaluation of a deep learning approach for the segmentation
netic resonance imaging: a review, arXiv:1712.03747v3 (2017). of brain tissues and white matter hyperintensities of presumed vascular
[257] Dora L, Agrawal S, Panda R, Abraham A. State-of-the-art meth- origin in MRI. NeuroImage Clin 2018;17:251–62.
ods for brain tissue segmentation: a review. IEEE Rev Biomed Eng [278] Pizarro R, Assemlal H-E, De Nigris D, Elliott C, Antel S, Arnold
2017;10:235–49. D, et al. Using deep learning algorithms to automatically identify
the brain MRI contrast: implications for managing large databases. [298] Kumar A, Kim J, Cai W, Fulham M, Feng D. Content-based medi-
Neuroinformatics 2018. cal image retrieval: a survey of applications to multidimensional and
[279] Laukamp KR, Thiele F, Shakirin G, Zopfs D, Faymonville A, Timmer multimodality data. J Digit Imaging 2013;26:1025–39.
M, et al. Fully automated detection and segmentation of meningiomas [299] Faria AV, Oishi K, Yoshida S, Hillis A, Miller I, Mori S. Content-
using deep learning on routine multiparametric MRI. Eur Radiol 2018. based image retrieval for brain MRI: an image-searching engine and
[280] Perkuhn M, Stavrinou P, Thiele F, Shakirin G, Mohan M, Garmpis D, population-based analysis to utilize past clinical data for future diag-
et al. Clinical evaluation of a multiparametric deep learning model for nosis. NeuroImage Clin 2015;7:367–76.
glioblastoma segmentation using heterogeneous magnetic resonance [300] Kumar A, Nette F, Klein K, Fulham M, Kim J. A visual analytics
imaging data from clinical routine. Invest Radiol 2018. approach using the exploration of multidimensional feature spaces for
[281] AlBadawy EA, Saha A, Mazurowski MA. Deep learning for segmenta- content-based medical image retrieval. IEEE J Biomed Health Inform
tion of brain tumors: impact of cross-institutional training and testing. 2015;19:1734–46.
Med Phys 2018;45:1150–8. [301] Bedo MVN, Pereira Dos Santos D, Ponciano-Silva M, de Azevedo-
[282] Cui S, Mao L, Jiang J, Liu C, Xiong S. Automatic semantic segmenta- Marques PM, Ferreira de Carvalho APdL, Traina C. Endowing
tion of brain gliomas from MRI images using a deep cascaded neural a content-based medical image retrieval system with percep-
network. J Healthc Eng 2018;2018:4940593. tual similarity using ensemble strategy. J Digit Imaging 2016;
[283] Hoseini F, Shahbahrami A, Bayat P. Adaptahead optimization algo- 29:22–37.
rithm for learning deep CNN applied to MRI segmentation. J Digit [302] Muramatsu C. Overview on subjective similarity of images for content-
Imaging 2018. based medical image retrieval. Radiol Phys Technol 2018.
[284] Yoo Y, Tang LYW, Brosch T, Li DKB, Kolind S, Vavasour I, et al. Deep [303] Spanier AB, Caplan N, Sosna J, Acar B, Joskowicz. A fully auto-
learning of joint myelin and T1w MRI features in normal-appearing matic end-to-end method for content-based image retrieval of CT scans
brain tissue to distinguish between multiple sclerosis patients and with similar liver lesion annotations. Int J Comput Assist Radiol Surg
healthy controls. NeuroImage Clin 2018;17:169–78. 2018;13:165–74.
[285] Bobo MF, Bao S, Huo Y, Yao Y, Virostko J, Plassard AJ, et al. Fully con- [304] Gordo A, Almazan J, Revaud J, Larlus D. End-to-end learning of deep
volutional neural networks improve abdominal organ segmentation. visual representations for image retrieval; 2017.
Proc SPIE 2018;10574. [305] Liu P, Guo J, Wu C, Cai D. Fusion of deep learning and compressed
[286] Shehata M, Khalifa F, Soliman A, Ghazal M, Taher M, Abou El-Ghar domain features for content-based image retrieval. IEEE Trans Image
A, et al. Computer-aided diagnostic system for early detection of acute Process 2017;26:5706–17.
renal transplant rejection using diffusion-weighted MRI. IEEE Trans [306] Han J, Zhang D, Cheng G, Liu N, Xu D. Advanced deep-learning
Bio-med Eng 2018. techniques for salient and category-specific object detection: a survey.
[287] Cheng R, Roth HR, Lay N, Lu L, Turkbey B, Gandler W, et al. Auto- IEEE Signal Process Mag 2018;35:84–100.
matic magnetic resonance prostate segmentation by deep learning with [307] Piplani T, Bamman D. Deepseek: content based image search &
holistically nested networks. J Med Imaging 2017;4:041302. retrieval, arXiv:1801.03406v2 (2018).
[288] Ishioka J, Matsuoka Y, Uehara S, Yasuda Y, Kijima T, Yoshida S, et al. [308] Yang J, Liang J, Shen H, Wang K, Rosin PL, Yang. Dynamic match
Computer-aided diagnosis of prostate cancer on magnetic resonance kernel with deep convolutional features for image retrieval. IEEE Trans
imaging using a convolutional neural network algorithm. BJU Int 2018. Image Process 2018;27:5288–302.
[289] Song Y, Zhang Y-D, Yan X, Liu H, Zhou M, Hu B, et al. Computer- [309] Sklan JES, Plassard AJ, Fabbri D, Landman BA. Toward content based
aided diagnosis of prostate cancer using a deep convolutional neural image retrieval with deep convolutional neural networks. Proc SPIE
network from multiparametric MRI. J Magn Reson Imaging: JMRI 2015;9417.
2018. [310] Bressan RS, Alves DHA, Valerio LM, Bugatti PH, Saito PTM. DOC-
[290] Wang X, Yang W, Weinreb J, Han J, Li Q, Kong X, et al. Searching ToR: the role of deep features in content-based mammographic image
for prostate cancer by fully automated magnetic resonance imag- retrieval. In: Proc. IEEE 31st int. symp. computer-based medical sys-
ing classification: deep learning versus non-deep learning. Sci Rep tems (CBMS). 2018. p. 158–63.
2017;7:15415. [311] Qayyum A, Anwar SM, Awais M, Majid M. Medical image retrieval
[291] Yang X, Liu C, Wang Z, Yang J, Min HL, Wang L, et al. Co-trained con- using deep convolutional neural network, arXiv:1703.08472v1 (2017).
volutional neural networks for automated detection of prostate cancer [312] Chung Y-A, Weng W-H. Learning deep representations of medical
in multi-parametric MRI. Med Image Anal 2017;42:212–27. images using Siamese CNNs with application to content-based image
[292] Le MH, Chen J, Wang L, Wang Z, Liu W, Cheng K-TT, et al. retrieval, arXiv:1711.08490v2 (2017).
Automated diagnosis of prostate cancer in multi-parametric MRI [313] Jing B, Xie P, Xing E. On the automatic generation of medical imaging
based on multimodal convolutional neural networks. Phys Med Biol reports, arXiv:1711.08195v3 (2017).
2017;62:6497–514. [314] Li CY, Liang X, Hu Z, Xing EP. Hybrid retrieval-generation rein-
[293] Forsberg D, Sjöblom E, Sunshine JL. Detection and labeling of ver- forced agent for medical image report generation, arXiv:1805.08298v1
tebrae in MR images using deep learning with clinical annotations as (2018).
training data. J Digit Imaging 2017;30:406–12. [315] Moradi M, Madani A, Gur Y, Guo Y, Syeda-Mahmood T. Bimodal
[294] Lu J-T, Pedemonte S, Bizzo B, Doyle S, Andriole KP, Michalski MH, network architectures for automatic generation of image annotation
et al. DeepSPINE: automated lumbar vertebral segmentation, disc- from text, arXiv:1809.01610v1 (2018).
level designation, and spinal stenosis grading using deep learning, [316] Zhang Y, Ding DY, Qian T, Manning CD, Langlotz CP. Learning to
arXiv:1807.10215v1 (2018). summarize radiology findings, arXiv:1809.04698v1.
[295] Han Z, Wei B, Leung S, Nachum IB, Laidley D, Li S. Auto- [317] Pons E, Braun LMM, Hunink MGM, Kors JA. Natural lan-
mated pathogenesis-based diagnosis of lumbar neural foraminal guage processing in radiology: a systematic review. Radiology
stenosis via deep multiscale multitask learning. Neuroinformatics 2016;279:329–43.
2018;16:325–37. [318] Zech J, Pain M, Titano J, Badgeley M, Schefflein J, Su A, et al. Natural
[296] Kim KH, Do W-J, Park S-H. Improving resolution of MR images with language-based machine learning models for the annotation of clinical
an adversarial network incorporating images with different contrast. radiology reports. Radiology 2018;287:570–80.
Med Phys 2018;45:3120–31. [319] Goff DJ, Loehfelm TW. Automated radiology report summarization
[297] Pilevar AH. CBMIR: content-based image retrieval algorithm for med- using an open-source natural language processing pipeline. J Digit
ical image databases. J Med Signals Sens 2011;1:12–8. Imaging 2018;31:185–92.
[320] Gibson E, Li W, Sudre C, Fidon L, Shakir DI, Wang G, et al. NiftyNet: pneumonia in chest radiographs: a cross-sectional study. PLoS Med
a deep-learning platform for medical imaging. Comput Methods Pro- 2018;15:e1002683.
grams Biomed 2018;158:113–22. [338] Lundervold A, Lundervold A, Rørvik J. Fast semi-supervised segmen-
[321] Li W, Wang G, Fidon L, Ourselin S, Cardoso MJ, Vercauteren T. On the tation of the kidneys in DCE-MRI using convolutional neural networks
compactness, efficiency, and representation of 3d convolutional net- and transfer learning; 2017.
works: brain parcellation as a pretext task. In: International conference [339] Lundervold A, Sprawka K, Lundervold A. Fast estimation of kidney
on information processing in medical imaging (IPMI). 2017. volumes and time courses in DCE-MRI using convolutional neural
[322] Kamnitsas K, Ledig C, Newcombe VF, Simpson JP, Kane AD, networks; 2018.
Menon DK, et al. Efficient multi-scale 3D CNN with fully connected [340] Hinton GE, Krizhevsky A, Wang SD. Transforming auto-encoders. In:
CRF for accurate brain lesion segmentation. Med Image Anal 2017; International conference on artificial neural networks. 2011. p. 44–51.
36:61–78. [341] Sabour S, Frosst N, Hinton GE. Dynamic routing between capsules. In:
[323] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks Advances in neural information processing systems. 2017. p. 3856–66.
for biomedical image segmentation. In: Medical image computing [342] Mnih V, Heess N, Graves A, et al. Recurrent models of visual atten-
and computer-assisted intervention (MICCAI). LNCS, vol. 9351. tion. In: Advances in neural information processing systems. 2014. p.
Springer; 2015. p. 234–41, arXiv:1505.04597 [cs.CV]. 2204–12.
[324] Badrinarayanan V, Kendall A, Cipolla R. Seg-Net: a deep convolu- [343] Xu K, Ba J, Kiros R, Cho K, Courville A, Salakhudinov R, et al. Show,
tional encoder-decoder architecture for image segmentation. IEEE attend and tell: neural image caption generation with visual attention.
Trans Pattern Anal Mach Intell 2017. In: International conference on machine learning. 2015. p. 2048–57.
[325] Mardani M, Gong E, Cheng JY, Vasanawala S, Zaharchuk G Alley M, [344] Castelvecchi D. Can we open the black box of AI? Nat News
et al. Deep generative adversarial networks for compressed sensing 2016;538:20.
automates MRI. arXiv:1706.00051 (2017). [345] Olah C, Satyanarayan A, Johnson I, Carter S, Schubert L, Ye K, et al.
[326] Parisot S, Ktena SI, Ferrante E, Lee M, Moreno RG, Glocker B, The building blocks of interpretability. Distill 2018;3:e10.
et al. Spectral graph convolutions for population-based disease pre- [346] Montavon G, Samek W, Müller K-R. Methods for interpreting and
diction. In: International conference on medical image computing and understanding deep neural networks. Digit Signal Process 2017.
computer-assisted intervention. 2017. p. 177–85. [347] Yosinski J, Clune J, Nguyen A, Fuchs T, Lipson H. Understanding neu-
[327] Marcus G. Deep learning: a critical appraisal, arXiv:1801.00631 ral networks through deep visualization. In: Deep learning workshop,
(2018). 31st international conference on machine learning. 2015.
[328] Lipton ZC, Steinhardt J. Troubling trends in machine learning schol- [348] Olah C, Mordvintsev A, Schubert L. Feature visualization. Distill
arship; 2018. 2017;2:e7.
[329] Zhang C, Bengio S, Hardt M, Recht B, Vinyals O. Understanding deep [349] Hohman FM, Kahng M, Pienta R, Chau DH. Visual analytics in deep
learning requires rethinking generalization, arXiv:1611.03530 (2016). learning: an interrogative survey for the next frontiers. IEEE Trans Vis
[330] Fredrikson M, Jha S, Ristenpart T. Model inversion attacks that exploit Comput Graph 2018.
confidence information and basic countermeasures. In: Proceedings of [350] Bach S, Binder A, Montavon G, Klauschen F, Müller K-R, Samek
the 22nd ACM SIGSAC conference on computer and communications W. On pixel-wise explanations for non-linear classifier decisions by
security, ACM. 2015. p. 1322–33. layer-wise relevance propagation. PLOS ONE 2015;10.
[331] Shokri R, Stronati M, Song C, Shmatikov V. Membership inference [351] Neal RM. Bayesian learning for neural networks [Ph.D. thesis]. Uni-
attacks against machine learning models. In: 2017 IEEE symposium versity of Toronto; 1995.
on security and privacy (SP), IEEE. 2017. p. 3–18. [352] MacKay DJ. A practical Bayesian framework for backpropagation
[332] McMahan B, Moore E, Ramage D, Hampson S, Arcas BA. networks. Neural Comput 1992;4:448–72.
Communication-efficient learning of deep networks from decentral- [353] Dayan P, Hinton GE, Neal RM, Zemel RS. The Helmholtz machine.
ized data. In: Singh A, Zhu J, editors. Proceedings of the 20th Neural Comput 1995;7:889–904.
international conference on artificial intelligence and statistics, Pro- [354] Li Y, Gal Y. Dropout inference in Bayesian neural networks with alpha-
ceedings of machine learning research, vol. 54. Fort Lauderdale, FL, divergences. In: International conference on machine learning. 2017.
USA: PMLR; 2017. p. 1273–82. p. 2052–61.
[333] Papernot N, Abadi M, Erlingsson U, Goodfellow I, Talwar K. Semi- [355] Leibig C, Allken V, Ayhan MS, Berens P, Wahl S. Leveraging uncer-
supervised knowledge transfer for deep learning from private training tainty information from deep neural networks for disease detection.
data, arXiv:1610.05755 (2016). Sci Rep 2017;7:17816.
[334] Papernot N, Song S, Mironov I, Raghunathan A, Talwar K, Erlingsson [356] Kendall A, Badrinarayanan V, Cipolla R. Bayesian segnet: model
Ú. Scalable private learning with PATE, arXiv:1802.08908 (2018). uncertainty in deep convolutional encoder-decoder architectures for
[335] McMahan HB, Ramage D, Talwar K, Zhang L. Learning differen- scene understanding, arXiv:1511.02680 (2015).
tially private recurrent language models. In: International conference [357] Feinman R, Curtin RR, Shintre S, Gardner AB. Detecting adversarial
on learning representations. 2017. samples from artifacts, arXiv:1703.00410 (2017).
[336] Chang K, Balachandar N, Lam C, Yi D, Brown J, Beers A, et al. [358] Sharp P, Hockfield S. Convergence: the future of health. Science
Distributed deep learning networks among institutions for medical 2017;355:589.
imaging. J Am Med Inform Assoc: JAMIA 2018;25:945–54. [359] Hood L, Flores M. A personal view on systems medicine and the emer-
[337] Zech JR, Badgeley MA, Liu M, Costa AB, Titano JJ, Oermann EK. gence of proactive P4 medicine: predictive, preventive, personalized
Variable generalization performance of a deep learning model to detect and participatory. New Biotechnol 2012;29:613–24.
Available online at www.sciencedirect.com
ScienceDirect

An Overview of Deep Learning in Medical Imaging Focusing On MRI

Uploaded by

Copyright:

Available Formats

An Overview of Deep Learning in Medical Imaging Focusing On MRI

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Overview of Deep Learning in Medical Imaging Focusing On MRI

Uploaded by

Copyright:

Available Formats

REVIEW

An overview of deep learning in medical imaging focusing on

Received 2 October 2018; accepted 21 November 2018

Keywords: Machine learning, Deep learning, Medical imaging, MRI

Z Med Phys 29 (2019) 102–127

used sigmoid or tanh activation functions.

eters like (discretized) T1 and T2 relaxation times, spin-density (M0 ), B0 ,

3.2.1 Image segmentation

35 37 And different hardware e.g. GPUs [225–227] as image registration is

google.com/custom-search, and also https://tineye.com, indexing more than cs231n.stanford.edu, https://developers.google.com/machine-learning/crash-

NiftyNet. An open source convolutional neural networks platform [320,321] http://niftynet.io

Grand-Challenges Grand challenges in biomedical image https://grand-challenge.org/

5.3 Workﬂow integration, regulations

Available online at www.sciencedirect.com

You might also like