Active Deep Learning For Medical Imaging Segmentation

Active Deep Learning for Medical Imaging
Segmentation
Marc Gorriz Axel Carlier

Xavier Giro-i-Nieto Emmanuel Faure
Universitat Politecnica de Catalunya University of Toulouse
Barcelona, Catalonia/Spain Toulouse, France
[email protected] [email protected]
Abstract
We propose a novel Active Learning framework capable to train effectively a

convolutional neural network for semantic segmentation of medical imaging, with
a limited amount of training labeled data. Our contribution is a practical Cost-
Effective Active Learning approach using Dropout at test time as Monte Carlo
sampling to model the pixel-wise uncertainty and to analyze the image information
to improve the training performance.
1 Motivation
The emergence of deep learning paradigm working through neural networks followed by the recent
advances in computational power have enabled the development of new intelligent diagnostics
based on computer vision. These diagnostics are capable to analyze images, performing accurate
segmentations, in order to detect the lesion areas and to make final decisions about the patient’s health
as the best of clinical eyes. Nevertheless, the use of deep convolutional neural networks (DCNN) may
be a handicap due to the human and logistic costs to get large labeled datasets to train the models
avoiding over-fitting. Active Learning (AL) is an established way to reduce this labeling workload in
order to select in an iterative way, the most informative examples from a subset of unlabeled instances,
to be manual labeled. It has been shown that using this methodology can train deep networks faster
and with fewer training samples than traditional semi-supervised learning methods.
2 Proposed methodology
Different from common Active Learning approaches, a Cost-Effective methodology [4] proposes
to automatically select and pseudo-annotate good unlabeled samples using the output network
predictions. Feeding all the unlabeled samples into the CNN, there are selected two kinds of
samples: the minority with low prediction confidence, called most informative/uncertain samples,
to be manually labeled by the user (oracle); and the majority with high prediction confidence, to
be automatically labeled with no human cost. As one can see that, these two kinds of samples are
complementary to each other for representing different confidence levels of the current model on the
unlabeled dataset. This procedure is often known as complementary sample selection.
The uncertainty for the model to predict a given input image could be a good criterion to rank all
the unlabeled data. In [2] it is explored the possibility to use Monte Carlo Dropout to approximate
the network weights distribution and hence study their behavior for specific inputs. It works by
randomly deactivating network activations during the training in order to prevent overfitting [1],
but being applied on test time will allow us to compute the pixel-wise sample uncertainty. Being
Ix a image pixel, we can estimate the uncertainty of its predicted label Iey computing the variance
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
of T different predictions on the same pixel by the effect of Dropout through the network weights.
The precision of pixel-wise uncertainty maps will depend on the number of step predictions and
the Dropout probability pd . Hight pd means hight variation of network weights making difficult a
consistent result with a finite number of step predictions. As was shown in [2] the ideal pd value is
0.5 and a maximum precision will be obtained when T ! 1. In order to integrate the method to
the CEAL complementary sample selection, we need to transform the pixel-wise uncertainty to a
numerical score in order to make a based-ranking. The easiest way is achieved by adding all the pixel
values from the uncertainty map, getting highest scores at the most doubtful samples. Depending on
the nature of the database could be needed a normalization to avoid a size-correlation.
Once all the overall uncertain values are computed, we can visualize the correlation between the
predictions goodness and their overall uncertain, by using the database ground truth. In Figure 1 we
can distinguish four regions according to the sample nature: (1) No-detected samples, although they
have low uncertainty are the most informative candidates to be manually annotated first. (2) Hight
uncertain samples, other possible candidates to be annotated by the oracle. (3) Certain samples, the
most common ones, perfect candidates to be selected as a pseudo-labels. (4) Random amount of
samples that will difficult the complementary sample selection, located at the central region.
Nevertheless, in the real world an hypothetic ground truth for the unlabeled samples can not be
available to generate that representation and a histogram could be used to obtain a lossy approximation
by projecting all the information to the uncertainty axis, despite being a poor criterion for discerning
between overlapped samples in the third and fourth regions. It can be demonstrated that selecting
randomly K samples from all the regions to be manually annotated at the initial CEAL iterations,
could be a good practice to reduce the concentration of these bad instances that could interfere at
the pseudo-labeling assignation procedure. With all the above, the complementary data selection
approach will plan a strategy to select sequentially the correct proportion of samples for each region
in each iteration, to improve gradually the network performance, without falling into over-fitting.
Sample selection per iteration

User annotations Samples
No-detections 10
Most uncertain 10
Random 15
Pseudo annotations Progressive
Figure 1: Left: Regions uncertainty representation. Right: CEAL approach.
3 Results and Future work
The dataset used is publicly available and contains 2000 RGB dermoscopy images manually annotated
by medical experts, by manual tracing the lesion boundaries in the form of a binary mask (International
Skin Imaging Collaboration). Before starting the interactive learning process, the training sets are
initialized, randomly selecting 200 samples for the test 600 samples for the labeled amount DL to
initialize the network weights and deleting the other labels to initialize the unlabeled set DU . We
trained the U-Net [3]: a convolutional neural network architecture specifically designed to solve
Biomedical Image Segmentation problems, using the CEAL approach defined in Figure 1, obtaining
a Dice Coefficient of 74 % after 9 active iterations with 2 training epochs per time.
The results are satisfactory but not as good as desired, still being samples in the central region that
interfere with the system improvement. Nonetheless, it remains an open door for future works
to continue researching new and more adapted solutions to achieve better starting points for the
pseudo-labels, with the desire to see their potential power in this field.
2
References
[1] Y. Gal and Z. Ghahramani. Bayesian Convolutional Neural Networks with Bernoulli Approximate Variational
Inference. 2015.
[2] Alex Kendall, Vijay Badrinarayanan, and Roberto Cipolla. Bayesian segnet: Model uncertainty in deep
convolutional encoder-decoder architectures for scene understanding. volume abs/1511.02680, 2015.
[3] P. Fischer O. Ronneberger and T. Brox. U-net: Convolutional networks for biomedical image segmentation.
In In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages
234–241, 2015.
[4] Keze Wang, Dongyu Zhang, Ya Li, Ruimao Zhang, and Liang Lin. Cost-effective active learning for deep
image classification. volume abs/1701.03551, 2017.

Active Deep Learning For Medical Imaging Segmentation

Uploaded by

Copyright:

Available Formats

Active Deep Learning For Medical Imaging Segmentation

Uploaded by

Document Information

Original Description:

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

Active Deep Learning For Medical Imaging Segmentation

Uploaded by

Copyright:

Available Formats

Active Deep Learning for Medical Imaging

Marc Gorriz Axel Carlier

We propose a novel Active Learning framework capable to train effectively a

Sample selection per iteration

Figure 1: Left: Regions uncertainty representation. Right: CEAL approach.

3 Results and Future work

You might also like