Neuroradiology (2020) 62:771–790

https://doi.org/10.1007/s00234-020-02403-1

REVIEW

Radiomics in gliomas: clinical implications of computational

modeling and fractal-based analysis
Kevin Jang 1,2 & Carlo Russo 3,4 & Antonio Di Ieva 3,4

Received: 17 January 2020 / Accepted: 10 March 2020 / Published online: 6 April 2020
# Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Radiomics is an emerging field that involves extraction and quantification of features from medical images. These data can be
mined through computational analysis and models to identify predictive image biomarkers that characterize intra-tumoral
dynamics throughout the course of treatment. This is particularly difficult in gliomas, where heterogeneity has been well
established at a molecular level as well as visually in conventional imaging. Thus, acquiring clinically useful features remains
difficult due to temporal variations in tumor dynamics. Identifying surrogate biomarkers through radiomics may provide a non-
invasive means of characterizing biologic activities of gliomas. We present an extensive literature review of radiomics-based
analysis, with a particular focus on computational modeling, machine learning, and fractal-based analysis in improving differ-
ential diagnosis and predicting clinical outcomes. Novel strategies in extracting quantitative features, segmentation methods, and
their clinical applications are producing promising results. Moreover, we provide a detailed summary of the morphometric
parameters that have so far been proposed as a means of quantifying imaging characteristics of gliomas. Newly emerging
radiomic techniques via machine learning and fractal-based analyses holds considerable potential for improving diagnostic
and prognostic accuracy of gliomas.

Key points
• Radiomic features can be mined through computational analysis to produce quantitative imaging biomarkers that characterize
intra-tumoral dynamics throughout the course of treatment.
• Surrogate image biomarkers identified through radiomics could enable a non-invasive means of characterizing biologic
activities of gliomas.
• With novel analytic algorithms, quantification of morphological or sub-regional tumor features to predict survival outcomes is
producing promising results.
• Quantifying intra-tumoral heterogeneity may improve grading and molecular sub-classifications of gliomas.
• Computational fractal-based analysis of gliomas allows geometrical evaluation of tumor irregularities and complexity, leading
to novel techniques for tumor segmentation, grading, and therapeutic monitoring.

Keywords Radiomics . Glioma . Computational modeling . Machine learning . Fractal analysis

* Antonio Di Ieva Introduction

[email protected]

1
Discipline of Surgery, Faculty of Medicine and Health, The With the increased research and development of computerized
University of Sydney, Sydney, New South Wales, Australia methods to analyze radiological images comes a new frontier
2
Nepean Hospital, Sydney, New South Wales, Australia for neuroimaging. Beyond qualitative interpretations,
3
methods that incorporate quantitative analyses are changing
Computational NeuroSurgery (CNS) Lab, Department of Clinical
Medicine, Faculty of Medicine and Health Sciences, Macquarie the way we interpret images. Radiomics is an emerging field
University, Sydney, Australia that involves extraction and quantification of features from
4
Department of Clinical Medicine, Faculty of Medicine and Health medical images [1, 2]. These data reflect underlying patholog-
Sciences, Neurosurgery Unit, Macquarie University, Level 2, Suite ical processes and in neuro-oncology can help to improve the
201, 2 Technology Place, Sydney, New South Wales 2109, Australia understanding of the biology and treatment of brain tumors.
772 Neuroradiology (2020) 62:771–790

Radiomic features can be mined through computational anal- errors and suboptimal treatment planning [14]. Radiomics
ysis to produce quantitative imaging biomarkers that charac- could overcome this limitation by spatially mapping areas of
terize intra-tumoral dynamics throughout the course of treat- distinct genetic features of the entire tumor [21]. Image-based
ment [3]. This may allow earlier detection of therapy response quantification of molecular heterogeneity could serve as a
and subsequent tailoring of treatment to individual patients powerful tool in identifying features that are predictive of
[2–5]. response or resistance to therapy [21, 22]. Surrogate imaging
Glioma is the most common primary brain tumor in adults, biomarkers identified through radiomics could enable a non-
representing over 80% of all diagnosed brain tumors [6]. invasive means of characterizing biologic activities of
High-grade gliomas (WHO grades III–IV) grow rapidly, infil- gliomas.
trating the brain parenchyma irregularly while creating exten- In this review, we explore the various features of
sive microvascular networks. Glioblastoma (GBM) carries the radiomics-based analysis with a particular focus on computa-
worst prognosis of the high-grade gliomas, with a median tional modeling, machine learning, and fractal-based analysis
survival of 12–15 months despite surgery followed by adjunct in improving differential diagnosis and prediction of clinical
chemotherapy and radiotherapy [6]. However, survival and outcomes and responsiveness to therapy. In doing so, we pro-
response to chemotherapy are incredibly heterogeneous vide a summary of the morphometric parameters that have so
among these patients. This stands in contrast to low-grade far been proposed as a means of quantifying imaging charac-
gliomas (WHO grades I–II), which are far less aggressive. teristics of gliomas.
Traditionally, the management of gliomas was based on their
histopathological grade determined by neuropathologists. In
recent years, however, specific molecular alterations have Extracting quantitative image features
been recognized as more prognostic than histological classifi-
cations [6]. This has led to an increased understanding of the Radiomics involves computational methods to reproducibly
tumor’s genomics, proteomics, and epigenetics as well, and extract objective, quantitative data from radiologic images
the clinical relevance of these molecular features to therapeu- [23]. The features are extracted from a defined region of in-
tic response and outcome [7, 8]. Modern molecular assess- terest (ROI) that includes the whole tumor or specific regions
ments include isocitrate dehydrogenase (IDH) mutations [9, within it. Morphometric parameters are used to quantify visual
10], co-deletion of chromosome arms 1p and 19q (1p/19q) characteristics at different scales from ROIs, enabling voxel-
[11, 12], O 6 - methylguanine–DNA methyltransferase based analysis of tumor volumetric shapes and visual dynam-
(MGMT) promoter hypermethylation status [13] and ATRX ics. A standard model of radiomics analysis can be seen in
mutations [14], among others. More recently, gene Fig. 1. Several approaches to extract radiomic features have
expression–based molecular characterization of glioma, in- been described in the literature, demonstrating accurate
cluding epidermal growth factor receptor (EGFR) amplifica- methods to capture tumor shape and textural information
tion [15] and CpG island methylator phenotype (CIMP) status [24]. Quantitative features can be categorized into the follow-
have emerged as predictive biomarkers of treatment outcome ing subgroups: shape features, first-order, second-order, and
and response [16]. Both molecular and histological classifica- higher-order statistics features. First-order statistics features
tion, however, require tissue analysis through stereotactic bi- describe the distribution of individual voxels regardless of
opsies or resection. These methods are limited by its invasive- spatial relationships (i.e., histogram-based properties) [25].
ness, sampling errors, and interpreter variability [17]. Second-order statistics features, generally described as “tex-
Moreover, there is substantial homogeneity and overlap of tural” features, are the statistical inter-relationships between
imaging features for distinguishing different grades or sub- neighboring voxels [25]. This quantifies the spatial distribu-
types of gliomas. tion of voxel intensities, and thus of intra-tumoral heterogene-
Intra-tumoral heterogeneity is an important hallmark of ity. Higher-order statistics features are extracted by applying
malignancy associated with poor prognosis. Increased hetero- filters or mathematical transforms to images; for instance, to
geneity has been linked to increased tumor adaptability, identify repetitive or non-repetitive patterns, suppressing
resulting in higher proliferative capacity and survivability noise or highlighting details [25]. An overview of these pa-
leading to higher risk for treatment failure [18]. This is of rameters can be found in Table 1. For their mathematical def-
paramount importance in gliomas, where heterogeneity has initions and detailed explanations, it is suggested to consult
been well established at a molecular level as well as visually the specific literature [24–43].
in conventional imaging [19, 20]. Intra-tumoral heterogeneity Although the field of radiomics is progressing at an accel-
interferes with both molecular and histopathological assess- erating rate, it is important to appreciate the historical origins
ments as the analysis of the whole tumor can be challenging. of this complex process. Most of the aforementioned statistical
Clonal heterogeneity and genetic variations may be underrep- features are neither original nor innovative descriptors [44].
resented in histopathologic sampling leading to sampling Indeed, the use of basic morphometric features to quantify
Neuroradiology (2020) 62:771–790 773

Fig. 1 Flowchart illustrates the standard radiomics workflow and the use manually or through automated methods and rendered in three dimen-
of radiomics in clinical decision making. In this example, feature sions (3D). Quantitative features are extracted and placed in a database
extractions are being performed on a SWI sequence of a segmented along with other clinical and genomic data. The database is then mined to
brain tumor. The process begins with acquisition of high-quality images. identify imaging features with the highest diagnostic, prognostic, or pre-
A region of interest (ROI) is identified on these images that contain the dictive value for outcomes of interest
whole tumor or sub-regions within the tumor. These are segmented either

image properties, as well as the use of filters and mathematical Despite these recent advances in computational techniques,
transforms, can be traced back several decades in the field of extraction methods are inherently limited as they distil a com-
image analysis and engineering [44]. In clinical medicine, one plex dataset of over a million voxels per magnetic resonance
of the first studies to correlate imaging findings with histology (MR) imaging sequence down to a relatively small number of
was published in 1988 [45]. With no advanced computational quantitative features. In addition, many textural analysis studies
techniques, Earnest et al. were able to demonstrate that en- still remain on a two-dimensional level (extracting features
hancing regions on images of astrocytomas overlapped with from a single slice). For a radiomic feature to be reliable, it must
areas of neovascularity and cell proliferation as determined fulfill two factors. First, the feature must be able to capture
through biopsy [45]. Thus, the novelty of radiomics relies on distinct patterns that are associated with improved differential
the –omics suffix, a term originally used for molecular biology diagnosis and clinical outcomes. And second, the feature must
disciplines. The term is now used for scientific fields that be stable under various image acquisition parameters [46].
generate complex high dimensional data from a single source. Although MR imaging technology has greatly improved in
A key advantage of -omics data is that these data can be mined the past decades, linking radiomic features to underlying tissue
and extended to generate hypotheses. In clinical imaging, the dynamics has not been fully explored. Indeed, acquiring useful
aim of radiomics is to initially capture as much data as possi- radiomic features becomes increasingly difficult with temporal
ble and then to use downstream database mining to identify variations in blood flow and tumor dynamics [47].
image features with the highest diagnostic and prognostic In recent years, there has been much effort to develop bio-
value. logically inspired radiomic features. These features build on
biologic hypotheses and can be used to define tissue-level data
774 Neuroradiology (2020) 62:771–790

Table 1 Parameters used in radiomics-based analysis of gliomas

Parameters and main references Definition

First-order texture statistics

Entropy [24] Measures the inherent randomness in the gray level intensities of an image or ROI.
Uniformity [24] Measures the homogeneity of gray level intensities within an image or ROI.
Second- and higher order texture statistics
Gray-level co-occurrence matrix [26] Examines the spatial distribution of gray level intensities within an image through a 2D gray
tone histogram.
Angular second movement [24] Measures the textural uniformity of an image (also referred to as homogeneity).
Captures the two-dimensional complexity of the edge of the tumor abnormalities.
Inverse difference moment [24] Measures local image homogeneity as it assumes larger values for smaller gray tone
differences in pair elements.
Contrast [24] Measures spatial tone frequency of an image as the difference between the highest and lowest
values of a contiguous set of pixels.
Correlation [24] Measure of gray tone linear dependencies in the image.
Bounding ellipsoid volume ratio [27] Ratio of the tumor volume to the volume of the smallest ellipsoid that entirely encapsulates
the tumor. Captures the three-dimensional complexity of tumors.
Semi-axis diameter ratios [28] Ratios of the minor semi-axis length to the longest bounding ellipsoid semi-axis diameter.
Captures the three-dimensional complexity of tumors.
Margin fluctuation [27, 28] Captures the two-dimensional complexity of the edge of the tumor abnormalities.
Standard deviation of the difference between the ordered radial distances of the tumor edge
from the centroid to all the boundary points, smoothed with an averaging filter of length
equal to 10% of tumor boundary.
Mean intensity [29] Average intensity of the pixel values within the ROI.
Mean of positive pixel values [29] Average pixel values of only the positive pixel values within the ROI.
Standard deviation (SD) [29] Quantification of the variance from the mean value
(high SD indicating wide variation of pixel values).
Kurtosis [29] Peakedness (or pointedness) of the histogram of pixel values.
Positive kurtosis = more peaked distribution
Negative kurtosis = flatter distribution
Skewness [29] Quantifies asymmetry of the histogram.
Negative skewness = longer tail on left side of histogram
Positive skewness = longer tail on right
Gray-level run matrix (GLRL) [30] Number of contiguous voxels that have the same gray level value.
Characterizes the gray level run lengths of different gray level intensities in any direction.
Short runs emphasis (SRE) [30] Measures distributions of short runs. Higher values indicate fine textures.
Long runs emphasis (LRE) [30] Measures distribution of long runs. Higher values indicate course textures.
Gray level non-uniformity (GLN) [30] Measures the distribution of runs over the gray values. Low value when runs are equally
distributed along gray levels. Lower value indicates higher similarity in intensity values.
Run length non-uniformity (RLN) [30] Measures distribution of runs over run lengths. Low value when runs are equally distributed
over run lengths.
Run percentage (RP) [30] Measures the fraction of the number of realized runs and the maximum number of potential
runs.
Highly uniform ROI volumes produce a low run percentage.
Neighborhood gray tone difference matrix [31] One dimensional matrix where each gray level entry is the summation of the differences
between all the pixels with gray level value and the average gray level value of its
neighborhood.
Coarseness [31] Quantitative measure of local uniformity.
Busyness [31] Rapid intensity changes of neighborhoods in a given ROI.
Complexity [31] Quantifies the complexity of the spatial information present in an image.
Texture strength [31] Characterizing the visual esthetics of an image.
Local binary pattern (LBP) [32] Quantifies local pixel structures through a binary coding scheme.
Measures tumor microenvironment.
Scale-invariant feature transform (SIFT) [33, 34] Detects distributed key points with radius on tumor images.
Measures tumor spatial characteristics.
Histogram of oriented gradients (HOG) [35] Computes block-wise histogram gradients with multiple orientations.
Neuroradiology (2020) 62:771–790 775

Table 1 (continued)

Parameters and main references Definition

Measures tumor microenvironment.

Fractal
Fractal dimension (box-counting and A non-integer number between 0 and 2, in a two-dimensional space, or 0 and 3, in a
sand-box algorithms) [36–38] three-dimensional volume, that quantifies the space-filling properties of irregularly
shaped objects.
Outline box dimension [24] Evaluates the irregularity in shape of the image. (i.e., how much it deviates from classic
geometric figures)
Lacunarity [39] Pixel distribution of an image at different box sizes and at various grid orientations.
Describes the degree of non-homogeneity within an image.
Spatial filtering
Median filter [40] Reduces sparse noise. Sets each pixel in ROI equal to the median pixel value of its specified
neighborhood.
Entropy filter [41] Accentuates edges by brightening pixels which have dissimilar neighbors.
Sets each pixel in the ROI equal to the entropy (measure of disorder) of the pixel values
in its specified neighborhood.
Laplacian of Gaussian (LoG) filter [41] Laplacian filter is a derivative filter used to find areas of rapid change (edges) in an image.
Images are first smoothed using Gaussian filter before applying the Laplacian.

variation in tumors. This provides an opportunity to study there is no consensus whether investigators should seek the
spatial variations and biologic evolution of tumors [46, 47]. ground truth or reproducibility of segmentation [1]. Manual
For instance, spatial distances were used to quantitatively segmentation by expert clinicians is often considered the
measure glioblastoma heterogeneity [48]. Variations in spatial ground truth, despite the high inter-reader variability. This
distances within a defined tumor sub-region were associated method is also time consuming, labor intensive, and not al-
with distinct prognostic information. In another study, spatial ways feasible for radiomics-based analysis which often re-
heterogeneity and early temporal changes in regions of high quires very large datasets [2]. To this end, automated and
and low perfusion in gliomas were predictive of physiologic semi-automated segmentation methods are being explored to
responses to radiation therapy [49]. Zhou et al. further pro- minimize manual input and improve consistency and repro-
posed a novel concept of extracting quantitative features from ducibility [68]. Whether reproducibility of segmentation out-
distinct tumor sub-regions, such as by their local contrast en- weighs the ground truth remains a controversial topic.
hancement, areas of edema, and cellularity in MR imaging Ultimately, however, the validity of any given radiomic fea-
[50]. A more recent study also revealed that T2-FLAIR and ture as a biomarker would have to be assessed on its ability to
ADC sequences were inversely proportional to cell density predict the outcome of interest—i.e., a given molecular, geno-
[51]. These novel radiomic features offer an opportunity to mic, or clinical endpoint, rather than the “ground truth” as it
quantitatively analyze the tumor environment, unlike the qual- pertains to the method of segmentation.
itative semantic features commonly used by radiologists to Automated and semi-automated segmentation methods
describe lesions. have been introduced for various imaging modalities and an-
atomical regions. Both require maximum automaticity with
minimal operator interaction, accuracy, time efficiency, and
The importance of image segmentation boundary reproducibility [2]. Certain algorithms use region-
growing methods that require an operator to set seed points,
Segmentation methods can be broadly categorized into the thresholds, and iteration termination conditions within the
following: threshold-based, region-based, edge-based, de- ROI [54]. Although these approaches are effective for rela-
formable model, machine- and deep learning–based and tively homogenous lesions, intensive user correction is often
model-based [52]. A review of these methods can be found required for lesions that are heterogenous. For instance, glio-
in Table 2. mas reveal infiltrative growth with lack of clear boundary and
Segmentation is a crucial step of the radiomics process for fixed growth pattern. Their complex pathological processes
many reasons. First, data are extracted from the segmented can be seen as complex changes in brightness and texture on
volumes. Moreover, different segmentation methods yield MR images. Distinct tissues may have similar gray levels,
very different geometrical parameters. In gliomas, this is par- which makes accurate and reproducible segmentation of glio-
ticularly challenging due to their irregular borders. Currently, mas challenging [68].
776 Neuroradiology (2020) 62:771–790

Table 2 Summary of glioma segmentation methods

Segmentation Method description Advantages Disadvantages

Global and local Depend on measuring thresholds from the Conceptually simple and Inapplicable to enhancing tumor areas.
thresholding histogram of an image. computationally fast. [53]
Region based
Region-growing Begins from a single pixel or group of pixels Computationally simple. Partial volume effect limits the accuracy
(seeds). Examines neighboring pixels for Can correctly segment regions with of segmentation. [55]
similarity and are included to ROI. similar properties and generate Sensitive to noise or variation in intensity
connected region. [54] (may result in holes or
over-segmentation).
Requires manual input for seed selection.
Watershed Treats pixels as a local topography Segments multiple regions Over-segmentation.
(elevation). The algorithm floods basin simultaneously.
until it reaches the watershed lines, Produces a complete contour of an
producing a complete contour of images. image (does not require contour
joining). [56]
Edge based Based on identifying differences between Computationally fast Resulting edge does not completely
pixels to determine the boundaries of an Does not require prior information enclose the object. [52]
object. about image content. [52] Sensitive to image noise.
Sensitive to significant variations in
gray level values.
Machine learning based
Supervised Uses labeled training data. Can be used for different tasks by Requires patient-specific training. [52]
simply changing the training set. [52] Human variability in manually labeling
Can reduce manual engineering task by training data.
providing labeled data and
appropriate parameters for the
learning algorithm. [52]
Unsupervised Training data are automatically labeled by Completely automated system. Number of regions often needs to be
numerically grouping similar pixels. pre-specified. [52]
Tumors can be divided into multiple
regions. [52]
Tumors may not have clearly defined
textural boundaries. [52]
Fuzzy C means Unsupervised segmentation by pixel Unsupervised. Time consuming. [57]
classification. Tumor boundaries always converged. Highly sensitive to noise and
[52] heterogeneity.
Artificial neural Supervised clustering method. Extracted Able to model non-trivial distributions Difficulty gathering training samples. [58]
networks features are fed through input nodes, and non-linear dependences. [52] Slow learning phase.
mathematical operations are applied and Able to learn from historical cases and
classification is made as a final output. automatically generate new rules.
[58]
Markov random Unsupervised clustering method that Able to represent complex Difficulty selecting parameters that
fields integrates spatial information into the dependencies among data instances. control the strength of spatial
clustering process. [52] interactions. Requires algorithms that
are computationally intensive. [52]
Deep learning based (CNN)
Interconnected operating modules
Single path Unique flow of information: input data is Fast computation and conceptually Limited parameters.
processed; feature maps are mined then simple. [59, 60] Single flow of information. [59, 60]
used for predicting label in the output
layer.
Multi-path Composed of different CNNs that work in Able to extract more diverse features. Computationally intensive (data
(parallel) parallel to capture more comprehensive Verdict is validated by interconnected preparation and processing). [59]
features. [61] nodules.
Included information may provide
contextual information to network
(e.g., multi-resolution). [59]
Multi-path Different CNNs arranged in a series (cascade) Able to extract even more diverse Requires careful preparation in designing
(series) with input from previous network. This features. network. [59]
makes the overall CNN deeper. [62] Different networks may require training.
Neuroradiology (2020) 62:771–790 777

Table 2 (continued)

Segmentation Method description Advantages Disadvantages

Enables refining information at any Could show minimal improvement.

stage. [59]
Input modalities
Single For processing information from a single More adaptable to different situations. Single source of information.
modality imaging modality. (e.g., as T1 most commonly provides
datasets for tissue and sub-cortical
segmentation). [63]
Easily used for various modalities.
Multi-modality For processing different sources of Useful for gaining contrast information. More parameters required than
information from multiple imaging [63] single-modality. [59, 63]
modalities.
Patch dimension
2D CNN Considers features from a single plane (i.e., Extensible to complex network Heavily reliant on initial network design
axial, sagittal or coronal). [60] structures. [59, 60] for good results. [59, 60]
Flexible and adaptable. Excludes 3D nature of MRI.
Fast computation.
2.5D CNN (or Provided with features from the 3 anatomical Faster than 3D. Computational more complex than 2D
tri-planar) planes (i.e., axial, sagittal, and coronal), Accounts for 3D nature of MRI. CNN.
using a multi-path design. Gains implicit contextual information.
[64]
3D CNN Extracts 3D segments directly from the MRI Able to examine 3D MRI volume Expensive computational cost.
volume. [65] directly. [59] Scaling to larger features may be
Better performance than 2D. computationally intensive.
Gains implicit contextual information. May require large training data due to
large number of parameters (memory
requirements). [65]
Number of predictions at a time
CNN The traditional approach where a single patch In theory, requires far less parameters Time consuming—as a single patch yields
is processed by a network, returning a than FCN. a single classification.
single output. [59]
Fully The fully connected layers are replaced with a Quicker segmentation than CNN (some Requires more parameters to be
convolution- fully convolutional layer—allowing dense can classify a single volume in one established.
al networks pixel-wise prediction. [66] shot). [59, 66] Requires more training samples.
(FCN) May return more false positive predictions
when classifying enhancing tumors.
[59, 66]
Model-based
Parametric Defined by a set of curves of internal and Able to extract boundary features for Depends on user-guidance to place land-
deformable external forces. Internal forces smooth the the same regions. marks to steer the segmentation.
models (active curves, while external forces change the Can be used for 3D volumetric Sensitive to noise.
contour or direction of curves toward the edges of an segmentation without training data. Requires initializing the contour that is
snake) anatomical area. Able to accommodate for changing close to the ROI.
biological structures over time. [67] May converge to wrong boundaries in
heterogeneous lesions. [67]
Level sets Represents contour as the zero-level set of a Accommodates to topological changes. Expensive computational cost.
higher dimensional function, then the Applicable to volumetric segmentation.
method formulates the motion of the con- [52]
tour as the evolution of the level-set func-
tion

Routine medical imaging techniques yield a wide variation institution, with possible variations in individual patients.
in acquisition parameters. For MR imaging, these include se- These variables may affect image noise and texture, and con-
quence-type, echo time, repetition time, number of excita- sequently the quality (and reproducibility) of the radiomic
tions, contrast-enhanced T1-weighted images, diffusion- features. In a given T1- or T2-weighted sequence, no voxel
weighted, and fluid attenuation sequences [44]. Moreover, intensity carries a fixed tissue-specific numerical value [44]. It
different manufacturers offer different reconstruction algo- is important to consider that some acquisition settings may
rithms, and reconstruction parameters are modified at each yield unstable features, which may produce different values
778 Neuroradiology (2020) 62:771–790

when extracted under identical conditions. For instance, even analyzing multimodal parameters and the vast amount of in-
when scanning the same patient in the same position using the formation we can deduct from clinical imaging to shape our
same scanner with the same sequence over multiple sessions, clinical practice.
signal intensity may change, whereas tissue contrast remains Similarly, in a recent retrospective study, Molina et al. re-
unaffected [69]. These limitations must be considered when vealed the predictive potential of 3D textural heterogeneity of
comparing radiomic features among patients as the process GBMs in post-contrast T1-weighted MR images [74].
relies on the numeric value of voxel intensity. One method Textural features were quantified as spatial distribution of
could be to perform textural analysis on features quantifying voxel intensities, allowing visualization of heterogeneity pat-
the relationship between voxel intensities, not requiring values terns within the segmented ROI of the tumor. These parame-
of individual voxel intensity, or through image compensation ters were classified as local (co-occurrence matrixes [CM]),
(normalization) before performing quantitative analysis [69], regional (run-length matrices [RLM]), or global (voxel inten-
such as the Brightness Progressive Normalization algorithm, sity histograms). High parametric values describing tumor ho-
introduced by Russo C and published first by Di Ieva et al. in mogeneity were associated with longer survival groups, while
2012 [70]. high values in heterogeneity were associated with poor sur-
vival [74]. This produced a threshold for classifying subset of
patients into long- and short-term survivors, which may ulti-
Clinical application of radiomics mately guide patient selection for surgical resection.
With the rise of machine learning in clinical medicine,
Survival prediction predicting overall survival in GBM patients has reached a
new frontier. Sanghani et al. analyzed tumor volumetric,
The survival prediction according to radiological features re- shape, and texture features from multiple MR images to pre-
mains a challenge in gliomas, above all in glioblastoma, due to dict overall survival of patients using machine learning tech-
intra-tumoral heterogeneity. A recent genomic analysis by niques [76]. Survival groups were defined as short (<
Sottoriva et al. revealed extensive intra-tumor variability at 10 months), medium (10–15 months), and long (> 15 months).
molecular, cellular, and tissue scales [71]. However, the clin- Using a support vector machine (SVM) classification for fea-
ical relevance of the spatial imaging characteristics remains ture selection, the morphological features were stratified into
enigmatic. Stratifying accurate prognosis of survival using two groups—2-class (< 400 days and > 400 days) and 3-class
radiomic spatial features pushes gliomas closer to the para- (short, medium, and long as defined above) survival group
digm of precision medicine. With novel analytic algorithms, prediction. The feature selection and prediction framework
quantification of morphological or sub-regional tumor fea- produced high accuracy for both classes in predicting overall
tures to predict survival outcomes has produced promising survival, where 2-class classification yielded 97.5% and 3-
results. Recent progress in the clinical application of class yielded 87.1%. These results testify to the power of
radiomics is summarized in Table 3. radiomics in predicting disease prognosis, thus providing in-
valuable information for tailoring treatment plans to individu-
Morphometric analysis: shape, texture, and volume al patients.

There is paucity of evidence on computational image analysis Sub-regional variability

of tumor morphology and its prognostic implications. Prior
studies often used 2D or simple 3D features such as tumor Several studies have demonstrated that GBM heterogeneity is
volume, with no control for prognostic variables, such as the not only limited to tumor margins but also involves peri-
Karnofsky Performance Score (KPS) or patient age. To over- tumoral brain parenchyma tissue. Analysis of these regions
come this shortcoming, Czarnek et al. demonstrated that after and its microenvironment suggests cellular and molecular in-
controlling for these variables, algorithmic analysis of GBM teraction that contributes to tumor infiltration, breakdown of
shapes was significantly prognostic of survival [27]. Using blood brain barrier, and microvascular proliferation, ultimate-
automated tumor segmentation from FLAIR sequences, three ly leading to poorer prognosis [79]. Radiomics allows one to
morphological features were found to be independently prog- study these subtle macro- and micro-scale changes within the
nostic of survival (p < 0.05): (a) glioma bounding ellipsoid lesion through quantitative measurement. The clinical rele-
volume ratio, (b) margin fluctuation, and (c) angular standard vance of these micro-architectural changes stands on two prin-
deviation. On FLAIR alone, margin fluctuation and angular cipal hypotheses: (a) radiomic features derived from
standard deviation were not statistically significant for prog- multiparametric MRI sequences can reveal subtle quantitative
nosis; however, when analyzed with post-contrast T1-weight- traits associated with tumor aggressiveness, and (b) these traits
ed MR images, both of these features were significant for are distinct between long- and short-term GBM survivors
survival prediction [27]. Thus, this proves the importance of [79].
Neuroradiology (2020) 62:771–790 779

Table 3 Major developments in radiomics modeling for gliomas

Author (year) Study Study methodology Major findings

population

Georgiadis et al. 67 1. Volume of interest (VOI) segmented from MRI series. 3D volumetric textural analysis improved discrimination
(2009) [72] 2. Volumetric textural features extracted (gray-level accuracy between metastases, gliomas and
co-occurrence and run-length matrices). meningiomas.
3. Bagging (bootstrap aggregation) of 3 LSFT-SVMs for Modified support vector machine (SVM) classifier using
classification scheme. least square features transformation (LSFT) improved
discrimination accuracy.
Zhou et al. 32 1. Linear normalization of tumor region. Long-term survival group had tumor habitats with high
(2014) [50] 2. Manual segmentation of ROI on T1+C. enhancement and high cell density.
3. 2D and 3D histogram analysis. Poor survival group had tumors with increased regions of
low enhancement.
Yang et al. 82 1. Tumor regions manually segmented T1-w and FLAIR Textural features are predictive of molecular subtypes and
(2015) [73] MR images. survival status in GBM.
2. 5 sets of textural features extracted: segmentation-based
fractal texture analysis, histogram of oriented gradients,
run-length matrix, local binary patterns, Haralick fea-
tures.
3. Ensemble classifier (random forest) used to predict GBM
molecular subtype and 12-month survival status.
Zhou et al. 16 1. Image data acquisition. Slow progression (> 500 days)—smaller distances between
(2013) [48] 2. Tumor region identification. ROI compared with fast progression.
3. Data normalization.
4. Image segmentation with OTSU algorithm.
5. Distance between two segmented regions measured.
Zhou et al. 32 1. Pair of tumor MRI slices selected as inputs. Spatial characteristics derived from tumor sub-regions of
(2017) [3] 2. Each sequence segmented by OTSU. edema (T2 and FLAIR) had the highest predictive value
3. Tumor region separated into two sub-regions. of prognosis (81.25% accuracy).
4. Spatial mapping to impose an overlap between
segmented sequences.
5. Features extracted from contrast-enhanced regions given
to machine-learning algorithm to build classifier to pre-
dict survival.
Molina et al. 79 1. Semi-automated image segmentation. 3D textural heterogeneity measures computed on
(2016) [74] 2. Segmented image manually corrected. post-contrast T1 MRI are predictors of survival.
3. 16 heterogeneity measures computed automatically.
4. Run-length matrix features used for regional heteroge-
neity.
5. Co-occurrence matrix features used for local heteroge-
neity.
Chang et al. 126 1. Volumetric tumor segmentation Machine learning techniques to analyze multimodal
(2016) [75] 2. Image registration, normalization and ADC submask imaging features could accurately predict survival in
generation. patients with recurrent glioblastoma treated with
3. Imaging features extraction: histogram, shape, bevacizumab.
multimodal parametric and textural.
4. Machine learning algorithm generated.
5. Kaplan-Meier analysis to evaluate progression-free sur-
vival and overall survival.
Gutman et al. 75 1. Pre-surgical MR images interpreted by 3 neuroradiolo- Overall survival was highly correlated to degree of contrast
(2013) [4] gists for size, location and morphology using standard- enhancement and length of major axis of lesion.
ized feature set. A semiquantitative computed method using standardized
2. Inter-rater analysis performed using Krippendorff α sta- visual feature set improved estimation of contrast
tistic and intra-class correlation coefficient. enhancement.
3. Multivariate Cox regression models for association
between survival and tumor size/morphology.
4. Fisher exact test for relationship between imaging
features and genomics.
Czarnek et al. 68 1. Five shape features automatically extracted from Algorithmic 3D analysis of tumor shape is a strong
(2017) [27] manually segmented tumor regions. prognostic marker of survival independent of patient age,
2. 3D features: nearest neighbor interpolation between MRI Karnofsky Performance Score and tumor volume.
slices to reconstruct 3D tumor shape. Fitted smallest
780 Neuroradiology (2020) 62:771–790

Table 3 (continued)

Author (year) Study Study methodology Major findings

population

bounding ellipsoid to the 3D tumor shape based on

Khachiyan algorithm.
3. 2D features: extracted from axial FLAIR and T1+C
images with largest tumor cross-section. Measured
margin fluctuation and angular standard deviation.
Mazurowski 22 1. Manual segmentation of pre-operative axial FLAIR im- The proportion of the tumor volume to the volume of the
et al. (2016) ages. smallest bounding ellipsoid is strongly predictive of
[28] 2. Extracted set of 5 features using computer algorithms. patient survival.
3. 2D features: margin fluctuation and angular standard
deviation were calculated.
4. 3D features: minimum bounding ellipsoid ratio,
semi-axis diameter ratios.
Sanghani et al. 163 1. Volumetric, shape and texture features extracted from Overall survival can be predicted with high accuracy using
(2018) [76] regions of edema, contrast-enhancement and necrosis. machine learning to analyze tumor volumetric, shape and
2. Feature selection using recursive feature elimination texture features.
(RFE).
3. Linear support vector machine (SVM) used for survival
group prediction.
Lao et al. (2017) 112 1. Segmentation of tumor sub-regions: necrosis, Deep learning–based radiomics model can accurately pre-
[77] enhancement and edema. dict overall survival by stratifying patients into high and
2. Handcrafted features extracted: geometry, intensity and low-risk groups. Radiomic signatures identified through
texture. deep learning outperformed manual extraction.
3. Deep features extracted from pre-trained CNN model via
transfer learning.
4. Four-step feature selection. Six most predictive deep
features selected.
5. Radiomics signature and radiomics nomogram
constructed.
Li et al. (2017) 92 1. Image pre-processed then automatically segmented into Introduced a fully automatic multiparametric radiomics
[78] 5 classes: non-tumor region and 4 tumor sub-regions model for pre-operative prediction of overall survival in
(necrosis, edema, enhancing and non-enhancing area). GBM patients.
2. High-throughput radiomics features extracted from tu- Multiparametric radiomics signature offered better
mor sub-regions. prognostic performance than fixed-parameter signatures.
3. Feature reproducibility and prognostic performance
assessed.
Prasanna et al. 65 1. Image pre-processing and registration. T2-w and FLAIR Peri-tumoral radiomic features outperformed features from
(2017) [79] were co-registered with T1+C. other regions (enhancing, necrosis) in predicting
2. Segmentation of tumor into 3 regions: parenchymal zone, survival.
necrosis and enhancement. Peri-tumoral radiomic features combined with clinical
3. 134 radiomic features obtained, resulting in 9 feature features (age, KPS) were more predictive of survival than
sets. radiomic features alone.
4. Identified 10 most predictive features.
5. Random forest classifier used to determine ability of each
feature set in predicting survival groups.
6. Randomized 3-fold cross-validation performed.
7. Kaplan-Meier survival analysis used to compare survival
times between short and long-term survivors.
Bahrami et al. 33 1. Volumes of interest (VOI) within FLAIR hyperintense Texture analysis using edge-contrast of FLAIR hyperin-
(2018) [80] region were segmented. tense regions may be predictive of survival in high-grade
2. Edge-contrast for each VOI was calculated using gradi- gliomas treated with bevacizumab.
ents of 3D FLAIR images. Low edge-contrast (vague borders) has poorer
3. Cox proportional hazard models were used to determine progression-free survival and overall survival compared
relationship between edge contrast and with patients with high edge-contrast (sharp border).
progression-free/overall survival. Age and extent of
surgical resection were used as covariates.
Bisdas et al. 37 1. DKI acquired using spin-echo echo planar imaging DWI Diffusional kurtosis imaging (DKI) accurately predicts IDH
(2018) [81] sequence. mutational status.
2. Tumor VOIs manually segmented around FLAIR Texture analysis and SVM analysis of DKI maps produced
abnormality. biomarkers to distinguish IDH-mutant from
3. Texture features extracted from both DKI and FLAIR IDH-wildtype as well as grade II from grade III gliomas.
VOIs.
Neuroradiology (2020) 62:771–790 781

Table 3 (continued)

Author (year) Study Study methodology Major findings

population

4. SVM analysis for binary classification: glioma grading

and IDH mutation status.
5. Biomarker selection using recursive feature elimination.
Bae et al. (2018) 217 1. Radiomic features extracted from multiparametric MRI. Radiomic phenotyping based on multiparametric MRI data
[82] 2. Random survival forest model trained with radiomic improves survival prediction when integrated with
features along with clinical and genetic profiles. clinical and genetic status in patients with GBM.
3. Incremental values of radiomic features assessed using
integrated area under the receiver operating characteristic
curve.
Bahrami et al. 61 1. Patients with grade II/III gliomas with molecular data and Quantitative FLAIR textural features (signal heterogeneity
(2018) [83] MRI prior to radiation included. and border sharpness) may serve as a useful imaging
2. Quantitative MRI features extracted—tissue heteroge- biomarker for determining tumor molecular status in
neity (homogeneity and pixel correlation) and FLAIR grade II/III gliomas.
border distinctiveness (edge contrast).
3. T tests performed to determine whether patients with
different genotypes differed across the features.
4. Logistic regression with LASSO regularization used to
determine optimal combination of imaging and clinical
features for predicting molecular subtypes.
Chaddad et al. 40 1. Acquisition of pre-treatment MR images. Using Laplacian-of Gaussian (LoG) filter to generate
(2018) [84] 2. Registration of T1-w image with corresponding FLAIR multiscale texture features has the potential to predict
images and labelling of GBM subtypes (phenotype). GBM survival.
3. Multiscale texture feature extraction.
4. Survival analysis.
Ditmer et al. 94 1. ROI manually segmented on T1+C images. Quantitative measurement of heterogeneity using MRI
(2018) [29] 2. Textural analysis performed using filtration-histogram textural analysis can accurately discriminate high versus
method. low grade gliomas.
3. Parameters were correlated with WHO glioma grade
using Spearman correlation.
4. AUC calculated using ROC curve analysis to distinguish
tumor grades.
Darbar et al. 48 1. ADC values calculated in areas of greatest restriction in ADC of tumor regions on pre-operative MRI can discrim-
(2018) [85] solid tumor components. inate high- and low-grade gliomas.
2. Pattern of contrast enhancement recorded. Low grade gliomas have significantly higher mean lowest
3. ROC analysis used to evaluate predictive potential of ADCs than high grade gliomas.
ADC values for low grade gliomas.
Osman A 163 1. Radiomic image features extracted locally from 3 tumor A derived gray-level co-occurrence matrix feature was
(2019) [86] sub-regions on multi-parametric MR images. found to be highly associated with survival—suggesting
2. LASSO regression applied for feature selection. intra-tumoral heterogeneity has an essential role in sur-
3. Radiomic signature model of 9 features constructed. vival stratification.
4. Model tested for patient stratification into short Ensemble learning showed superior performance over the
(< 10 months), medium (10–15 months) and long tested ML classifiers.
survivors (> 15 months).
5. ML classification models trained then cross-validated.
Petrujkić et al. 55 1. Each tumor outlined on T1+C images. Textural features are more significant than fractal-based
(2019) [87] 2. ROI over imposed to corresponding T2-w and SWI im- features in differentiating glioblastoma from solitary
ages. metastasis.
3. Tumor representation in these 3 sequences were
segmented.
4. Binary image obtained.
5. Quantitative parameters of fractal and texture analysis
were estimated—using box-counting method and
GLCM methods.
Yang et al. 1. Textural features from 30 parametric maps were Gray-level size-zone matrix (GLSZM) combined with
(2019) [88] extracted using 4 models: global, GLCM, GLRLM, gray-level 64 may be the optimal texture retrieving
GLSZM. model for glioma grading.
2. These features were then input into RBF-SVM combined
with attribute selection using SVM-RFE.
3. SVM model was trained and established using 10-fold
cross validation.
782 Neuroradiology (2020) 62:771–790

In a recent experimental study, Zhou et al. identified quan- combined across multiparametric sequences, it revealed
titative spatial imaging biomarkers with prognostic value in higher predictive ability than radiomic features derived from
predicting survival outcomes [3]. Through two datasets of necrotic or enhancing areas. This result is reflected in other
patients with unresected GBM, tumor habitats were quantified studies that have demonstrated peri-tumoral edema on MRI as
on multiple MRI slices (including contrast-enhanced T1, a negative prognostic marker; however, the precise role of
FLAIR, and T2 sequences). Quantitative features from signal peri-tumoral brain parenchyma zones in GBM prognosis re-
enhancing tumor sub-regions revealed discriminative ability mains controversial. For instance, in a multi-institutional
in predicting survival groups. More specifically, spatial char- study by Schoenegger et al., surrounding areas of edema were
acteristics derived from sub-regions of edema (co-occurring identified as an independent prognostic marker, where pa-
MRI signals in FLAIR and T2-weighted images) displayed tients displaying extensive edema had significantly poorer
the highest predictive ability in separating long-term (> overall survival compared with those with minimal edema
400 days) and short-term (< 400 days) survival groups. [89]. However, this finding was contradicted by Lacroix
Mapping sub-regions of edema yielded the highest accuracy et al. in a large series of 416 GBM patients where the extent
of 81.25% in predicting survival groups (p < 0.05)—indicat- of edema was not a prognostic marker for overall survival
ing strong prognostic value of MRI-defined sub-regions in [90]. Thus, the prognostic implications of peri-tumoral edema
GBMs. in GBM have been inconclusive in the literature. The reason
With the same fundamental concepts, Lao et al. developed for these inconsistencies may be due to prior studies only
a deep learning–based radiomics model to predict survival examining gross volumetric measurement of these areas.
outcomes in GBM patients [77]. Three tumor sub-regions Radiomics holds the potential to overcome these limitations
were segmented from multimodality MR images (T1, post- through capturing subtle local variations in image intensities
contrast T1, T2, and FLAIR), including areas of necrosis, that are otherwise visually not appreciable.
enhancement, and edema. Handcrafted and higher-order deep
features were then extracted for selection, with the final aim of Classification of glioma subtypes
selecting features with prognostic value. A six-feature
radiomics signature was constructed, and these signatures The relationship between intra-tumoral heterogeneity and tu-
were shown to accurately stratify 75 patients into high- and mor infiltrative capacity, response to treatment and overall
low-risk groups, successfully predicting overall survival. All survival has been well established in the literature. Although
six features were deep features derived from multiple tumor the current gold standard for grading gliomas involves histo-
sub-regions in post-contrast T1, T2, and FLAIR images. It is pathological analysis, stereotactic biopsies, or resection, these
not surprising that deep features extracted via transfer learning techniques are inherently limited by its invasiveness, sampling
outperformed traditional manual extraction in predicting over- error, and interpreter variability [17]. Moreover, there is sig-
all survival—as higher-order imaging patterns can capture nificant overlap of conventional and multiparametric imaging
more intra-tumoral heterogeneity [77]. Such radiographic het- features for differentiating high- and low-grade tumors. Thus,
erogeneity of GBMs may reflect underlying genetic heteroge- radiomic analysis to quantify intra-tumoral heterogeneity may
neity, which could explain treatment resistance and poorer improve diagnostic and prognostic accuracy allowing tailored
prognosis. However, this remains a complex hypothesis and treatment planning and monitoring of therapeutic response.
defining the correlation between deep features and genetic As mentioned previously, textural analysis is a key tool of
characteristics requires further research. Despite the study be- radiomics in unraveling complex imaging patterns.
ing retrospective with a relatively small sample size, the pro- Algorithms can examine spatial distribution of gray levels in
posed radiomics model has the potential to shape pre- an image, by incorporating a filtration-histogram approach
operative management of patients with GBM. where textural features of varying intensities are quantified
Prasanna et al. similarly extracted radiomic features from using histogram-based statistical metrics [29]. Several studies
sub-regions of GBM habitat including enhancing tumors, ne- suggest that these extracted features may be of potential use as
crotic areas, and the peri-tumoral brain parenchyma zones spatial imaging biomarkers for GBM heterogeneity. This is
[79]. These features were derived from 65 pre-treatment exemplified by a recent study by Skogen et al., where
multiparametric MRI sequences (T1, T2, and FLAIR) in order histogram-based textural analysis of GBM on MR images
to distinguish long- and short-term survivors. Peri-tumoral was able to accurately discriminate high- and low-grade tu-
features were found to be more predictive across T2 and mors [91]. Ditmer et al. further extended this notion through a
FLAIR (p = 0.0006 and p = 0.003, respectively) compared retrospective study of 94 patients to determine the accuracy of
with enhancing areas or necrotic features. However, on post- radiomic-based filtration-histogram textural analysis in grad-
contrast T1, radiomic features from necrotic sub-regions were ing gliomas [29]. Their analysis found that fine texture fea-
more predictive of survival prognosis than peri-tumoral zones tures on post-contrast T1 images have the strongest ability in
(p = 0.006). Interestingly, when peri-tumoral features were discriminating high- and low-grade gliomas. This reiterates
Neuroradiology (2020) 62:771–790 783

the findings by Skogen et al. where fine texture scales were brain cortex, thus suitable to distinguish pathophysiological
also found to be the best discriminative feature, with a sensi- states in MR imaging [38].
tivity and specificity of 93% and 81% (p < 0.05) [29]. Several experimental studies have used fractal analysis for
Despite the accumulating work in improving differentia- MR brain imaging classifications, given its unique ability to
tion of gliomas through radiomic models, there remains a evaluate the self-affinity at multiple scales and the long-range
question of adding value to molecular sub-classifications. To correlations of an image [38]. Authors have hypothesized that
address this, Macyszyn et al. applied machine learning and normal brain MRI has higher self-affinity and long-range cor-
pattern recognition methods to predict GBM molecular sub- relations compared with those with neuropathologies, and that
types by extracting imaging phenotypes [92]. Tumors were these morphological changes can be quantified using compu-
categorized into four subtypes through an isoform-level assay tational fractal-based analysis [38, 93–95]. For further details
classifier: (a) proneural, (b) neural, (c) mesenchymal, and (d) regarding the theoretical principles of fractal-based analysis
classical. Molecular subtypes were predicted with an overall and its clinical applications into the basic and clinical neuro-
accuracy of 75.76%. Imaging phenotypes that were most pre- sciences, see Di Ieva et al., references [37, 38, 96].
dictive for each subtype were (a) histogram of T2-FLAIR
intensity and mean T1 in enhancing tumors for proneural sub- Fractal geometry of brain tumors
type, (b) T2 intensity histogram in areas of edema and tumor
location for neural subtype, (c) histogram of T2-FLAIR inten- The geometrical structure of tumors tends to be complex
sity and mean T1 signal in edema for proneural subtype, and and irregular due to the uneven spatial distribution of their
(d) size of enhancing regions, T2-FLAIR intensity histogram, cells and microvessels. In particular, brain tumors exhibit
and peak height on perfusion signals in areas of edema for irregular geometry during their growth process and are
classical subtype. Several studies have shown similar findings apparent even in their microvascular networks and spatial
where specific molecular subtypes of GBM displayed unique diffusion through time. Fractal-based analysis in neuroim-
imaging phenotypes that could be extracted and used as non- aging has been fundamental in the geometrical evaluation
invasive biomarkers. For instance, mesenchymal tumors have and quantification of tumor irregularities. Its precise abil-
been found to have lower non-enhancing tumor volume and ity to characterize geometric features of irregular and
surrounding edema intensity, proneural subtypes have signif- complex natural objects has led to novel techniques for
icantly lower blood-brain barrier breakdown, and classical tumor segmentation [58, 97, 98], tumor grading [97, 99],
subtypes are strongly associated with features of necrosis and therapeutic monitoring [70, 97]. Parameters computed
and edema [92]. Unlike prior studies that involved tissue spec- by means of fractal analysis can be used to not only val-
imen analysis, Macyszyn et al. demonstrated that molecular idate tumor growth models but also to gain further clinical
subtypes of GBM can be accurately predicted using imaging and prognostic information in oncological patients [100].
alone. Fractal-based parameters, such as the fractal dimension
(FD), have been increasingly used for tumor segmentation in
neuroimaging, oncologic grading, and evaluation of therapy
[101]. More specifically, MRI with contrast enhancement
Computational fractal-based analysis [97], susceptibility-weighted MRI (known as SWI) [70, 99],
and histological specimens have been assessed by means of
The application of fractalomics in neuroscience is relatively a fractal analysis [102–106]. For instance, Di Ieva et al. evalu-
new paradigm [37, 38]. Fractal analysis is a tool used to math- ated the fractal dimension on 7 Tesla SWI-MRI for grading of
ematically assess morphological features (e.g., roughness and gliomas [99]. Their findings revealed an increasing trend of
geometrical complexity) of natural objects [37, 38]. Within the the intra-tumoral SWI patterns’ fractal dimension with tumor
last decade, fractal analysis has become an attractive method grade—1.682 ± 0.278 for grade II, 2.018 ± 0.517 for grade III,
to quantify complex morphological features in computed to- and 2.247 ± 0.358 for grade IV gliomas. Statistically signifi-
mography (CT) and magnetic resonance imaging [38]. More cant difference was found between grade II and grade IV
recently, it has been applied in neuroimaging for automated gliomas (p < 0.05), which proved that fractal geometric anal-
classifications to improve diagnostic and prognostic accuracy ysis can accurately distinguish high- and low-grade tumor. In
[93–95]. Fractal dimension, a parameter used in fractal anal- an earlier study, fractal capacity dimension was used to eval-
ysis, is a non-integer number between 0 and 2, in a two- uate the effects of antiangiogenic treatments [70]. This was
dimensional space, or 0 and 3, in a volume of interest, that also performed on 7 Tesla SWI-MR images to monitor in vivo
quantifies the geometrical complexity of natural objects and the therapeutic response. These promising results testify to the
their ability to fill the surrounding space in which it is embed- value of using fractal analysis on 7 T SWI-MRI to quantita-
ded [38]. Fractal dimension computation has been shown to tively examine malignant brain tumors and their dynamics
be useful in characterizing the complex morphology of the during antiangiogenic therapy.
784 Neuroradiology (2020) 62:771–790

Fractal geometry analysis on post-contrast MR images has technical challenges in managing big data. With multiple data
also been applied by Iftekharuddin et al. [58, 97] and Zook et al. sources (e.g., institutions) and various data types (e.g.,
[98] for brain tumor detection and FD estimation. They pro- multiparametric imaging data, gene expression profiles, and
posed three modified box-counting algorithms where pixel in- clinical records), standardizing data collection and sharing be-
tensities can be viewed in the third dimension, rendering them come incredibly complex [46]. Differences in image acquisi-
more suitable for fractal textural analysis. Using a feature ex- tion and reconstruction are covariates that must be addressed
traction methodology with a self-organizing map, multiple frac- in the mining of quantitative features. Standards will have to
tal parameters were derived from post-contrast T1, T2, and be established across different image protocols and parameters
FLAIR MRI modalities. Following this extraction, the authors to validate results from radiomic models. Thus, there is a need
could train a supervised neural network to automatically classi- for mutual agreements between national, international, and
fy image regions as tumorous or non-tumorous. Several authors multi-institutional consortia to share data through centralized
have also used FD to analyze the 3D tumor interface in GBMs. or federated networks [1]. Initiatives such as The Cancer
Interestingly, Smitha et al. analyzed FLAIR sequences to assess Genome Atlas [109], The Cancer Imaging Archive [110],
variations in fractal dimensions of the tumor contours in low- and the Quantitative Imaging Network [111] have allowed
and high-grade gliomas [39]. Low-grade gliomas yielded a FD efficient sharing of clinical data and help validate imaging
of 1.243 ± 0.127, while high-grade gliomas revealed 1.338 ± biomarkers against an independent dataset. However, estab-
0.248, with a statistically significant difference (p < 0.05). lishing high-quality benchmarks with complete clinical labels,
A dataset of fractal dimensions correlates to a particular standard radiomic features, and molecular profiles remains to
feature of the brain tumor lesion, such as enhancing regions, be challenging at a larger scale. Growing efforts to improve
geometric texture, vascularity, and tumor interface. These fea- data sharing, experimental evaluation, and reproducibility will
tures are invaluable in characterizing the dynamic evolution of push radiomics closer toward precision medicine.
brain tissue from normal to dysplastic and to neoplastic. Such
descriptors may greatly contribute to improving diagnosis and The role of radiologists
therapeutic monitoring, and serve as the platform for develop-
ing innovative tumor growth models for optimizing therapy In current practice, radiologic investigations are qualitatively
and drug delivery. examined. The finalized reports often do not use a standard
lexicon, despite recent efforts such as the RadLex® [112].
Although guidelines exist for reporting, none are available
Future perspectives for reporting quantitative imaging features, let alone for
reporting highly complex radiomics features. Due to the lack
Substantial progress has been made in the field of of standards, huge existing image repositories are essentially
radiomics to improve our understanding of the biology inaccessible for curation. Moreover, archived medical images
and evolution of brain tumors. Quantitative characteristics are rarely re-accessed. The most practical solution would be to
derived from neuroimaging modalities enables imaging capture data prospectively at the point of care. This may lead
surrogate biomarkers to be validated through machine to a transition from classic radiology to a future where radiol-
learning and fractal-based analyses. This allows subtle ogists actively participate in the curation of quantitative image
variations in the intra-tumoral microenvironment to be databases [1]. Generating high-quality image data would re-
monitored throughout the course of treatment. Despite quire considerable expertise in identifying, segmenting (with
the growing body of literature, there remains a need to computer assistance), and annotating (using a standardized
develop more specific and precise methods to apply quan- and mineable lexicon) the regions of interest [1]. For the
titative imaging features in a clinical setting. Here, we curation of high-dimensional data to become a reality, radiol-
explore the challenges and opportunities that pertain to ogists should be first convinced of its value, and the process
brain cancer imaging and the application of radiomics in must be refined to work within the limitations of clinical prac-
neuro-oncology. tice. Indeed, we envision radiomics to become a valuable asset
to improving diagnostic accuracy and clinical decision mak-
The role of big data ing. With further involvement of radiologists in the curation
and analysis of big data, radiomics will continue to push the
Modern healthcare has seen an exponential growth in biomed- boundaries of precision medicine.
ical data generation and extraction from individual patients
[1]. Massive datasets, so-called “big data,” are required by Developments in machine learning
radiomic studies to validate deep learning–based approaches
and expand its clinical applications [107, 108]. Despite the As described previously, identifying multiparametric prognos-
growing potential for radiomics, there remain logistical and tic imaging biomarkers remains a challenge when extracting
Neuroradiology (2020) 62:771–790 785

large-scale radiomic features from numerous imaging modal- understanding of GBM biology. In the coming years,
ities. Development of machine learning algorithms to analyze radiomic analysis, using fractal geometry or higher-order
rich databases could prove to be beneficial in identifying clin- statistical methods such as machine learning or deep
ically relevant feature descriptors [46]. Sparse-learning learning, will be eventually able to redefine tumor sub-
models, also referred to as lasso regularization [113], have types paving way for discovery of new biomarkers, with
been used in other cancer pathologies to identify prognostic the final aim to improve decision making and patients’
imaging biomarkers—such as non-small cell lung cancer treatment.
[114]. Although its applicability in neuroimaging is still lack-
ing, recent developments in SVM learning and deep learning Developments in fractal analysis
models prove to be encouraging in classifying GBM subtypes.
However, the lack of widely available labeled medical data In regard to fractal-based analysis, its efficacy in classifying
poses a challenge for developing novel deep learning models brain MR images has been well established through many
[46]. For instance, collecting a large enough database of can- breakthrough studies in the field of biomedical engineering.
cer images with accurate histopathological labels is costly at Prior studies [93–95] could be extended through optimizing
scale; thus, developing methods to integrate data with varying feature selection and applying the computational of the fractal
clinical labels may provide the opportunity to input large dimension and related parameters (e.g., lacunarity) to other
datasets for deep learning models [115]. The clinical relevance neuroimaging modalities such as functional MR images, CT,
of deep learning outputs from multi-scale medical data (e.g., and positron-emission tomography (PET). Zook and
multimodal MR imaging and genomics) remains uncertain; Iftekharuddin proposed integrating tumor subtypes, geometri-
however, the ability to extract concise imaging patterns via cal size, and the effect of noise when analyzing fractal dimen-
artificial neural networks may guide future studies in devel- sions [98]. Wardlaw et al. suggested the removal of cardiac
oping large-scale radiomic models [116]. and respiratory factors from the blood oxygen level–
dependent (BOLD) signals to better identify tumor sub-
Developments in targeted therapy regions of active metabolism [122]. Iftekharrudin et al. recom-
mended future studies to improve discrimination of various
Optimizing treatment selection for GBM requires further brain tissues, i.e., white matter, gray matter, cerebrospinal flu-
research into developing specific radiomic signatures. id, and skull, in order to better distinguish solid tumors and
With neuroimaging modalities such as diffusion- areas of edema [58]. In the computational era, fractal-based
weighted sequences, there is growing evidence that ADC analysis may be incorporated into diffusion tensor imaging
maps may be beneficial to differentiate clinical outcomes studies and even nuclear medicine tools. For instance, fractal
in GBM treated with radiation therapy concurrently with analysis of single photon emission computed tomography
temozolomide [117]. In fact, early variations in ADC (SPECT) and PET imaging may deepen our insight of
maps were identified as a potential marker for predicting radiomics and may offer novel biomarkers that are clinically
GBM recurrence [118]. As mentioned earlier, machine applicable [123].
learning and fractal-based analyses were also used to pre-
dict treatment responses to bevacizumab—where a de-
crease in the volume of FLAIR signal and contrast en- Conclusion
hancement was found as a potential biomarker in estimat-
ing therapeutic success [119]. To further improve our un- With novel imaging biomarkers being uncovered at an accel-
derstanding of intra-tumoral dynamics expressed as imag- erating rate through radiomics, comes a new frontier for inte-
ing phenotypes, larger collection of radiomic features ex- grating multiparametric data to improve the treatment of brain
tracted at various diagnostic periods could provide an op- tumors. Computational models are expanding with the use of
portunity to describe tumor evolution before and after machine learning and fractal-based analysis, which are in-
treatment [120]. Distinguishing tumor growth between creasingly becoming paramount for diagnostic and prognostic
pseudo-progression and pseudo-response continues to be accuracy. Over time, these models will have to be aligned with
a challenge through imaging alone [121]; thus, develop- tumor biology to maximize the clinical implications of
ment of radiomic models to better characterize treatment radiomics. Current obstacles in understanding tumor hetero-
outcomes will push the field of neuro-oncology a step geneity may be overcome through increased research in com-
closer to precision medicine. Although validation of putational models and extending those findings to the clinical
radiomic features as true predictors of treatment response realm. Our review of the newly emerging radiomic techniques
is yet to be defined, the growing depth of radiomic find- via machine learning and fractal-based analyses demonstrate
ings combined with growing genomic and clinical data the potential for improving diagnostic and prognostic accura-
may provide the opportunity to redefine our cy of gliomas. The field of radiomics is a rapidly developing
786 Neuroradiology (2020) 62:771–790

field with many avenues yet to be explored for further discov- Ellison DW (2016) The 2016 World Health Organization classifi-
cation of tumors of the central nervous system: a summary. Acta
ery and innovation.
Neuropathol 131(6):803–820. https://doi.org/10.1007/s00401-
016-1545-1
Acknowledgments Professor Antonio Di Ieva received the 2019 John 7. Bredel M, Scholtens DM, Harsh GR, Bredel C, Chandler JP,
Mitchell Crouch Fellowship from the Royal Australasian College of Renfrow JJ, Yadav AK, Vogel H, Scheck AC, Tibshirani R,
Surgeons (RACS) which, along with Macquarie University co-funding, Sikic BI (2009) A network model of a cooperative genetic land-
supported the opening of the Computational NeuroSurgery (CNS) Lab at scape in brain tumors. JAMA 302(3):261–275. https://doi.org/10.
Macquarie University, Sydney, Australia. Moreover, he is supported by 1001/jama.2009.997
an Australian Research Council (ARC) Future Fellowship (2019-2023, 8. Sturm D, Witt H, Hovestadt V, Khuong-Quang D-A, Jones David
FT190100623). KJ would like to personally thank Professor Antonio Di TW, Konermann C, Pfaff E, Tönjes M, Sill M, Bender S, Kool M,
Ieva, Dr. Carlo Russo, and Dr. Abhishta Bhandari for their untiring sup- Zapatka M, Becker N, Zucknick M, Hielscher T, Liu X-Y,
port during the course of this work. We also thank Dr. Bhandari for Fontebasso Adam M, Ryzhova M, Albrecht S, Jacob K, Wolter
contributing to the figure used in this manuscript. M, Ebinger M, Schuhmann Martin U, van Meter T, Frühwald
Michael C, Hauch H, Pekrun A, Radlwimmer B, Niehues T, von
Funding information No funding was received for this study. Komorowski G, Dürken M, Kulozik Andreas E, Madden J,
Donson A, Foreman Nicholas K, Drissi R, Fouladi M, Scheurlen
W, von Deimling A, Monoranu C, Roggendorf W, Herold-Mende
Compliance with ethical standards C, Unterberg A, Kramm Christof M, Felsberg J, Hartmann C,
Wiestler B, Wick W, Milde T, Witt O, Lindroth Anders M,
Conflict of interest The authors declare that they have no conflict of Schwartzentruber J, Faury D, Fleming A, Zakrzewska M,
interest. Liberski Pawel P, Zakrzewski K, Hauser P, Garami M, Klekner
A, Bognar L, Morrissy S, Cavalli F, Taylor Michael D, van Sluis P,
Ethical approval All procedures performed in the studies involving hu- Koster J, Versteeg R, Volckmann R, Mikkelsen T, Aldape K,
man participants were in accordance with the ethical standards of the Reifenberger G, Collins VP, Majewski J, Korshunov A, Lichter
institutional and/or national research committee and with the 1964 P, Plass C, Jabado N, Pfister Stefan M (2012) Hotspot mutations in
Helsinki Declaration and its later amendments or comparable ethical H3F3A and IDH1 define distinct epigenetic and biological sub-
standards. groups of glioblastoma. Cancer Cell 22(4):425–437. https://doi.
org/10.1016/j.ccr.2012.08.024
Informed consent Informed consent was obtained from all individual 9. Yan H, Parsons DW, Jin G, McLendon R, Rasheed BA, Yuan W,
participants included in the study. Kos I, Batinic-Haberle I, Jones S, Riggins GJ, Friedman H,
Friedman A, Reardon D, Herndon J, Kinzler KW, Velculescu
VE, Vogelstein B, Bigner DD (2009) IDH1 and IDH2 mutations
in gliomas. N Engl J Med 360(8):765–773. https://doi.org/10.
References 1056/NEJMoa0808710
10. Sanson M, Marie Y, Paris S, Idbaih A, Laffaire J, Ducray F, El
Hallani S, Boisselier B, Mokhtari K, Hoang-Xuan K, Delattre J-Y
1. Gillies RJ, Kinahan PE, Hricak H (2016) Radiomics: images are
(2009) Isocitrate dehydrogenase 1 codon 132 mutation is an im-
more than pictures, they are data. Radiology 278(2):563–577.
portant prognostic biomarker in gliomas. J Clin Oncol 27(25):
https://doi.org/10.1148/radiol.2015151169
4150–4154. https://doi.org/10.1200/JCO.2009.21.9832
2. Kumar V, Gu Y, Basu S, Berglund A, Eschrich SA, Schabath MB, 11. Eckel-Passow JE, Lachance DH, Molinaro AM, Walsh KM,
Forster K, Aerts HJ, Dekker A, Fenstermacher D, Goldgof DB, Decker PA, Sicotte H, Pekmezci M, Rice T, Kosel ML, Smirnov
Hall LO, Lambin P, Balagurunathan Y, Gatenby RA, Gillies RJ IV, Sarkar G, Caron AA, Kollmeyer TM, Praska CE, Chada AR,
(2012) Radiomics: the process and the challenges. Magn Reson Halder C, Hansen HM, McCoy LS, Bracci PM, Marshall R,
Imaging 30(9):1234–1248. https://doi.org/10.1016/j.mri.2012.06. Zheng S, Reis GF, Pico AR, O’Neill BP, Buckner JC, Giannini
010 C, Huse JT, Perry A, Tihan T, Berger MS, Chang SM, Prados MD,
3. Zhou M, Chaudhury B, Hall LO, Goldgof DB, Gillies RJ, Wiemels J, Wiencke JK, Wrensch MR, Jenkins RB (2015) Glioma
Gatenby RA (2017) Identifying spatial imaging biomarkers of groups based on 1p/19q, IDH, and TERT promoter mutations in
glioblastoma multiforme for survival group prediction. J Magn tumors. N Engl J Med 372(26):2499–2508. https://doi.org/10.
Reson Imaging 46(1):115–123. https://doi.org/10.1002/jmri. 1056/NEJMoa1407279
25497 12. Jenkins RB, Blair H, Ballman KV, Giannini C, Arusell RM, Law
4. Gutman DA, Cooper LA, Hwang SN, Holder CA, Gao J, Aurora M, Flynn H, Passe S, Felten S, Brown PD, Shaw EG, Buckner JC
TD, Dunn WD Jr, Scarpace L, Mikkelsen T, Jain R, Wintermark (2006) A t(1;19)(q10;p10) mediates the combined deletions of 1p
M, Jilwan M, Raghavan P, Huang E, Clifford RJ, Mongkolwat P, and 19q and predicts a better prognosis of patients with
Kleper V, Freymann J, Kirby J, Zinn PO, Moreno CS, Jaffe C, oligodendroglioma. Cancer Res 66(20):9852–9861. https://doi.
Colen R, Rubin DL, Saltz J, Flanders A, Brat DJ (2013) MR org/10.1158/0008-5472.CAN-06-1796
imaging predictors of molecular profile and survival: multi- 13. Hegi ME, Liu L, Herman JG, Stupp R, Wick W, Weller M, Mehta
institutional study of the TCGA glioblastoma data set. MP, Gilbert MR (2008) Correlation of O6-methylguanine methyl-
Radiology 267(2):560–569. https://doi.org/10.1148/radiol. transferase (MGMT) promoter methylation with clinical outcomes
13120118 in glioblastoma and clinical strategies to modulate MGMT activ-
5. Kotrotsou A, Zinn PO, Colen RR (2016) Radiomics in brain tu- ity. J Clin Oncol 26(25):4189–4199. https://doi.org/10.1200/JCO.
mors: an emerging technique for characterization of tumor envi- 2007.11.5964
ronment. Magn Reson Imaging Clin N Am 24(4):719–729. 14. Delfanti RL, Piccioni DE, Handwerker J, Bahrami N, Krishnan A,
https://doi.org/10.1016/j.mric.2016.06.006 Karunamuni R, Hattangadi-Gluth JA, Seibert TM, Srikant A,
6. Louis DN, Perry A, Reifenberger G, von Deimling A, Figarella- Jones KA, Snyder VS, Dale AM, White NS, McDonald CR,
Branger D, Cavenee WK, Ohgaki H, Wiestler OD, Kleihues P, Farid N (2017) Imaging correlates for the 2016 update on WHO
Neuroradiology (2020) 62:771–790 787

classification of grade II/III gliomas: implications for IDH, 1p/19q institutional study. J Neuro-Oncol 132(1):55–62. https://doi.org/
and ATRX status. J Neuro-Oncol 135(3):601–609. https://doi.org/ 10.1007/s11060-016-2359-7
10.1007/s11060-017-2613-7 28. Mazurowski MA, Czarnek NM, Collins LM, Peters KB, Clark K
15. Haas-Kogan DA, Prados MD, Tihan T, Eberhard DA, Jelluma N, (2016) Predicting outcomes in glioblastoma patients using com-
Arvold ND, Baumber R, Lamborn KR, Kapadia A, Malec M, puterized analysis of tumor shape - preliminary data. In: Progress
Berger MS, Stokoe D (2005) Epidermal growth factor receptor, in Biomedical Optics and Imaging - Proceedings of SPIE. https://
protein kinase B/Akt, and glioma response to erlotinib. JNCI: J doi.org/10.1117/12.2217098
Natl Cancer Instit 97(12):880–887. https://doi.org/10.1093/jnci/ 29. Ditmer A, Zhang B, Shujaat T, Pavlina A, Luibrand N, Gaskill-
dji161 Shipley M, Vagal A (2018) Diagnostic accuracy of MRI texture
16. Noushmehr H, Weisenberger DJ, Diefes K, Phillips HS, Pujara K, analysis for grading gliomas. J Neuro-Oncol 140(3):583–589.
Berman BP, Pan F, Pelloski CE, Sulman EP, Bhat KP, Verhaak https://doi.org/10.1007/s11060-018-2984-4
RGW, Hoadley KA, Hayes DN, Perou CM, Schmidt HK, Ding L, 30. Galloway MM (1975) Texture analysis using gray level run
Wilson RK, Van Den Berg D, Shen H, Bengtsson H, Neuvial P, lengths. Comput Graphics Image Process 4(2):172–179. https://
Cope LM, Buckley J, Herman JG, Baylin SB, Laird PW, Aldape K doi.org/10.1016/S0146-664X(75)800086
(2010) Identification of a CpG island methylator phenotype that 31. Amadasun M, King R (1989) Textural features corresponding to
defines a distinct subgroup of glioma. Cancer Cell 17(5):510–522. textural properties. IEEE Trans Syst Man Cybern 19(5):1264–
https://doi.org/10.1016/j.ccr.2010.03.017 1274. https://doi.org/10.1109/21.44046
17. Raab SS, Grzybicki DM, Janosky JE, Zarbo RJ, Meier FA, Jensen 32. Ojala T, Pietikäinen M, Mäenpää T (2002) Multiresolution gray-
C, Geyer SJ (2005) Clinical impact and frequency of anatomic scale and rotation invariant texture classification with local binary
pathology errors in cancer diagnoses. Cancer 104(10):2205– patterns, vol 24. vol 7 %. J IEEE Trans Pattern Anal Mach. Intell.
2213. https://doi.org/10.1002/cncr.21431 IEEE Computer Society. https://doi.org/10.1109/tpami.2002.
18. Gerlinger M, Rowan AJ, Horswell S, Math M, Larkin J, 1017623
Endesfelder D, Gronroos E, Martinez P, Matthews N, Stewart 33. Ambrosini RD, Wang P, O'Dell WG (2010) Computer-aided de-
A, Tarpey P, Varela I, Phillimore B, Begum S, McDonald NQ, tection of metastatic brain tumors using automated three-
Butler A, Jones D, Raine K, Latimer C, Santos CR, Nohadani M, dimensional template matching. J Magn Reson Imaging 31(1):
Eklund AC, Spencer-Dene B, Clark G, Pickering L, Stamp G, 85–93. https://doi.org/10.1002/jmri.22009
Gore M, Szallasi Z, Downward J, Futreal PA, Swanton C 34. Lowe DG (1999) Object recognition from local scale-invariant
(2012) Intratumor heterogeneity and branched evolution revealed features. In: Proceedings of the Seventh IEEE International
by multiregion sequencing. N Engl J Med 366(10):883–892. Conference on Computer Vision, 20–27 Sept. 1999. pp 1150–
https://doi.org/10.1056/NEJMoa1113205 1157 vol.1152. https://doi.org/10.1109/ICCV.1999.790410
19. Friedmann-Morvinski D (2014) Glioblastoma heterogeneity and 35. Dalal N, Triggs B (2005) Histograms of oriented gradients for
cancer cell plasticity. Crit Rev Oncog 19(5):327–336. https://doi. human detection. In: 2005 IEEE Computer Society Conference
org/10.1615/critrevoncog.2014011777 on Computer Vision and Pattern Recognition (CVPR'05), 20–25
20. Inda MM, Bonavia R, Seoane J (2014) Glioblastoma multiforme: June 2005. vol. 881. pp 886–893. https://doi.org/10.1109/CVPR.
a look inside its heterogeneous nature. Cancers (Basel) 6(1):226– 2005.177
239. https://doi.org/10.3390/cancers6010226 36. Mandelbrot BB, Freeman WH, Company (1983) The fractal ge-
21. Gevaert O, Mitchell LA, Achrol AS, Xu J, Echegaray S, Steinberg ometry of nature. Henry Holt and Company
GK, Cheshier SH, Napel S, Zaharchuk G, Plevritis SK (2014) 37. Di Ieva A, Grizzi F, Jelinek H, Pellionisz AJ, Losa GA (2013)
Glioblastoma multiforme: exploratory radiogenomic analysis by Fractals in the neurosciences, part I: general principles and basic
using quantitative image features. Radiology 273(1):168–174. neurosciences. Neuroscientist 20(4):403–417. https://doi.org/10.
https://doi.org/10.1148/radiol.14131731 1177/1073858413513927
22. Zinn PO, Mahajan B, Sathyan P, Singh SK, Majumder S, Jolesz 38. Di Ieva A, Esteban FJ, Grizzi F, Klonowski W, Martín-Landrove
FA, Colen RR (2011) Radiogenomic mapping of edema/cellular M (2013) Fractals in the neurosciences, part II: clinical applica-
invasion MRI-phenotypes in glioblastoma multiforme. PLoS One tions and future perspectives. Neuroscientist 21(1):30–43. https://
6(10):e25451. https://doi.org/10.1371/journal.pone.0025451 doi.org/10.1177/1073858413513928
23. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van 39. Smitha KA, Gupta AK, Jayasree RS (2015) Fractal analysis: frac-
Stiphout RG, Granton P, Zegers CM, Gillies R, Boellard R, tal dimension and lacunarity from MR images for differentiating
Dekker A, Aerts HJ (2012) Radiomics: extracting more informa- the grades of glioma. Phys Med Biol 60(17):6937–6947. https://
tion from medical images using advanced feature analysis. Eur J doi.org/10.1088/0031-9155/60/17/6937
Cancer 48(4):441–446. https://doi.org/10.1016/j.ejca.2011.11. 40. Lim JS (1990) Two-dimensional signal and image processing.
036 Prentice-Hall, Inc.
24. Djuricic GJ, Radulovic M, Sopta JP, Nikitovic M, Milosevic NT 41. Gonzalez RC, Woods RE, Eddins SL (2004) Digital image pro-
(2017) Fractal and gray level cooccurrence matrix computational cessing using MATLAB. Pearson Education
analysis of primary osteosarcoma magnetic resonance images pre- 42. Aerts HJ, Velazquez ER, Leijenaar RT, Parmar C, Grossmann P,
dicts the chemotherapy response. Front Oncol 7:246. https://doi. Carvalho S, Bussink J, Monshouwer R, Haibe-Kains B, Rietveld
org/10.3389/fonc.2017.00246 D, Hoebers F, Rietbergen MM, Leemans CR, Dekker A,
25. Parekh V, Jacobs MA (2016) Radiomics: a new application from Quackenbush J, Gillies RJ, Lambin P (2014) Decoding tumour
established techniques. Expert Rev Precis Med Drug Dev 1(2): phenotype by noninvasive imaging using a quantitative radiomics
207–226. https://doi.org/10.1080/23808993.2016.1164013 approach. Nat Commun 5:4006. https://doi.org/10.1038/
26. Haralick RM, Shanmugam K, Dinstein I (1973) Textural features ncomms5006
of image classification. IEEE Trans Syst Man Cybern 3(6):610– 43. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N,
621. https://doi.org/10.1109/TSMC.1973.4309314 Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts H
27. Czarnek N, Clark K, Peters KB, Mazurowski MA (2017) (2017) Computational Radiomics system to decode the radio-
Algorithmic three-dimensional analysis of tumor shape in MRI graphic phenotype. Cancer Res 77(21):e104–e107. https://doi.
improves prognosis of survival in glioblastoma: a multi- org/10.1158/0008-5472.Can-17-0339
788 Neuroradiology (2020) 62:771–790

44. Rizzo S, Botta F, Raimondi S, Origgi D, Fanciullo C, Morganti image analysis on magnetic resonance imaging: a review. Artif
AG, Bellomi M (2018) Radiomics: the facts and the challenges of Intell Med 95:64–81. https://doi.org/10.1016/j.artmed.2018.08.
image analysis. Eur Radiol Exp 2(1):36–36. https://doi.org/10. 008
1186/s41747-018-0068-z 60. Zikic D, Ioannou Y, Brown M, Criminisi A (2014) Segmentation
45. Earnest F, Kelly PJ, Scheithauer BW, Kall BA, Cascino TL, of brain tumor tissues with convolutional neural networks. In:
Ehman RL, Forbes GS, Axley PL (1988) Cerebral astrocytomas: Proceedings MICCAI-BRATS. pp 36–39
histopathologic correlation of MR and CT contrast enhancement 61. Brébisson Ad, Montana G (2015) Deep neural networks for ana-
with stereotactic biopsy. Radiology 166(3):823–827. https://doi. tomical brain segmentation. In: 2015 IEEE Conference on
org/10.1148/radiology.166.3.2829270 Computer Vision and Pattern Recognition Workshops
46. Zhou M, Scott J, Chaudhury B, Hall L, Goldgof D, Yeom KW, Iv (CVPRW), 7–12 June 2015. pp 20–28. https://doi.org/10.1109/
M, Ou Y, Kalpathy-Cramer J, Napel S, Gillies R, Gevaert O, CVPRW.2015.7301312
Gatenby R (2018) Radiomics in brain tumor: image assessment, 62. Havaei M, Davy A, Warde-Farley D, Biard A, Courville A,
quantitative feature descriptors, and machine-learning approaches. Bengio Y, Pal C, Jodoin P-M, Larochelle H (2017) Brain tumor
Am J Neuroradiol 39(2):208–216. https://doi.org/10.3174/ajnr. segmentation with deep neural networks. Med Image Anal 35:18–
A5391 31. https://doi.org/10.1016/j.media.2016.05.004
47. Gatenby RA, Grove O, Gillies RJ (2013) Quantitative imaging in 63. Zhang W, Li R, Deng H, Wang L, Lin W, Ji S, Shen D (2015)
cancer evolution and ecology. Radiology 269(1):8–15. https://doi. Deep convolutional neural networks for multi-modality isointense
org/10.1148/radiol.13122697 infant brain image segmentation. NeuroImage 108:214–224.
48. Zhou M, Hall L, Goldgof D, Gillies R, Gatenby R (2013) Survival https://doi.org/10.1016/j.neuroimage.2014.12.061
time prediction of patients with glioblastoma multiforme tumors 64. Lyksborg M, Puonti O, Agn M, Larsen R (2015) An ensemble of
using spatial distance measurement, vol 8670. SPIE Medical 2D convolutional neural networks for tumor segmentation. In:
Imaging. SPIE Paulsen RR, Pedersen KS (eds) Image Analysis, Cham, 2015.
49. Cao Y, Tsien CI, Nagesh V, Junck L, Ten Haken R, Ross BD, Springer International Publishing, p 201–211
Chenevert TL, Lawrence TS (2006) Survival prediction in high-
65. Li W, Wang G, Fidon L, Ourselin S, Cardoso MJ, Vercauteren T
grade gliomas by MRI perfusion before and during early stage of
(2017) On the compactness, efficiency, and representation of 3D
RT [corrected]. Int J Radiat Oncol Biol Phys 64(3):876–885.
convolutional networks: brain parcellation as a pretext task. In:
https://doi.org/10.1016/j.ijrobp.2005.09.001
Niethammer M, Styner M, Aylward S, et al (eds) Information
50. Zhou M, Hall L, Goldgof D, Russo R, Balagurunathan Y, Gillies
Processing in Medical Imaging, Cham, 2017. Springer
R, Gatenby R (2014) Radiologically defined ecological dynamics
International Publishing, p 348–360
and clinical outcomes in glioblastoma multiforme: preliminary
66. Kamnitsas K, Ledig C, Newcombe VFJ, Simpson JP, Kane AD,
results. Transl Oncol 7(1):5–13. https://doi.org/10.1593/tlo.13730
Menon DK, Rueckert D, Glocker B (2017) Efficient multi-scale
51. Chang PD, Malone HR, Bowden SG, Chow DS, Gill BJA, Ung
3D CNN with fully connected CRF for accurate brain lesion seg-
TH, Samanamud J, Englander ZK, Sonabend AM, Sheth SA,
mentation. Med Image Anal 36:61–78. https://doi.org/10.1016/j.
McKhann GM 2nd, Sisti MB, Schwartz LH, Lignelli A,
media.2016.10.004
Grinband J, Bruce JN, Canoll P (2017) A multiparametric model
for mapping cellularity in glioblastoma using radiographically lo- 67. McInerney T, Terzopoulos D (1996) Deformable models in med-
ical image analysis: a survey. Med Image Anal 1(2):91–108.
calized biopsies. AJNR Am J Neuroradiol 38(5):890–898. https://
https://doi.org/10.1016/S1361-8415(96)80007-7
doi.org/10.3174/ajnr.A5112
52. Gordillo N, Montseny E, Sobrevilla P (2013) State of the art sur- 68. Wu Y, Zhao Z, Wu W, Lin Y, Wang M (2019) Automatic glioma
vey on MRI brain tumor segmentation. Magn Reson Imaging segmentation based on adaptive superpixel. BMC Med Imaging
31(8):1426–1438. https://doi.org/10.1016/j.mri.2013.05.002 19(1):73. https://doi.org/10.1186/s12880-019-0369-6
53. Gibbs P, Buckley DL, Blackband SJ, Horsman A (1996) Tumour 69. Madabhushi A, Udupa JK (2006) New methods of MR image
volume determination from MR images by morphological seg- intensity standardization via generalized scale. Med Phys 33(9):
mentation. Phys Med Biol 41(11):2437–2446. https://doi.org/10. 3426–3434. https://doi.org/10.1118/1.2335487
1088/0031-9155/41/11/014 70. Di Ieva A, Matula C, Grizzi F, Grabner G, Trattnig S,
54. Hojjatoleslami SA, Kittler J (1998) Region growing: a new ap- Tschabitscher M (2012) Fractal analysis of the susceptibility
proach. IEEE Trans Image Process 7(7):1079–1084. https://doi. weighted imaging patterns in malignant brain tumors during
org/10.1109/83.701170 antiangiogenic treatment: technical report on four cases serially
55. Mie S, Lakare S, Ming W, Kaufman A, Nakajima M (2000) A imaged by 7 T magnetic resonance during a period of four weeks.
gradient magnitude based region growing algorithm for accurate World Neurosurg 77(5-6):785.e711–785.e721. https://doi.org/10.
segmentation. In: Proceedings 2000 International Conference on 1016/j.wneu.2011.09.006
Image Processing (Cat. No.00CH37101), 10–13 Sept. 2000. 71. Sottoriva A, Spiteri I, Piccirillo SGM, Touloumis A, Collins VP,
vol.443, pp 448–451. https://doi.org/10.1109/ICIP.2000.899432 Marioni JC, Curtis C, Watts C, Tavaré S (2013) Intratumor het-
56. Dam E, Loog M, Letteboer M (2004) Integrating automatic and erogeneity in human glioblastoma reflects cancer evolutionary
interactive brain tumor segmentation. In: Proceedings of the 17th dynamics. Proc Natl Acad Sci U S A 110(10):4009–4014.
International Conference on Pattern Recognition, 2004. ICPR https://doi.org/10.1073/pnas.1219747110
2004., 26–26 Aug. 2004. vol 793, pp 790–793. https://doi.org/ 72. Georgiadis P, Cavouras D, Kalatzis I, Glotsos D, Athanasiadis E,
10.1109/ICPR.2004.1334647 Kostopoulos S, Sifaki K, Malamas M, Nikiforidis G, Solomou E
57. Kannan SR (2008) A new segmentation system for brain MR (2009) Enhancing the discrimination accuracy between metasta-
images based on fuzzy techniques. Appl Soft Comput 8(4): ses, gliomas and meningiomas on brain MRI by volumetric tex-
1599–1606. https://doi.org/10.1016/j.asoc.2007.10.025 tural features and ensemble pattern recognition methods. Magn
58. Iftekharuddin KM, Zheng J, Islam MA, Ogg RJ (2009) Fractal- Reson Imaging 27(1):120–130. https://doi.org/10.1016/j.mri.
based brain tumor detection in multimodal MRI. Appl Math 2008.05.017
Comput 207(1):23–41. https://doi.org/10.1016/j.amc.2007.10.063 73. Yang D, Rao G, Martinez J, Veeraraghavan A, Rao A (2015)
59. Bernal J, Kushibar K, Asfaw DS, Valverde S, Oliver A, Martí R, Evaluation of tumor-derived MRI-texture features for discrimina-
Lladó X (2019) Deep convolutional neural networks for brain tion of molecular subtypes and prediction of 12-month survival
Neuroradiology (2020) 62:771–790 789

status in glioblastoma. Med Phys 42(11):6725–6735. https://doi. 87. Petrujkic K, Milosevic N, Rajkovic N, Stanisavljevic D,
org/10.1118/1.4934373 Gavrilovic S, Dzelebdzic D, Ilic R, Di Ieva A, Maksimovic R
74. Molina D, Perez-Beteta J, Luque B, Arregui E, Calvo M, Borras (2019) Computational quantitative MR image features - a poten-
JM, Lopez C, Martino J, Velasquez C, Asenjo B, Benavides M, tial useful tool in differentiating glioblastoma from solitary brain
Herruzo I, Martinez-Gonzalez A, Perez-Romasanta L, Arana E, metastasis. Eur J Radiol 119:108634. https://doi.org/10.1016/j.
Perez-Garcia VM (2016) Tumour heterogeneity in glioblastoma ejrad.2019.08.003
assessed by MRI texture analysis: a potential marker of survival. 88. Yang Y, Yan LF, Zhang X, Nan HY, Hu YC, Han Y, Zhang J, Liu
Br J Radiol 89(1064):20160242. https://doi.org/10.1259/bjr. ZC, Sun YZ, Tian Q, Yu Y, Sun Q, Wang SY, Zhang X, Wang W,
20160242 Cui GB (2019) Optimizing texture retrieving model for multimod-
75. Chang K, Zhang B, Guo X, Zong M, Rahman R, Sanchez D, al MR image-based support vector machine for classifying glioma.
Winder N, Reardon DA, Zhao B, Wen PY, Huang RY (2016) J Magn Reson Imaging 49(5):1263–1274. https://doi.org/10.1002/
Multimodal imaging patterns predict survival in recurrent glio- jmri.26524
blastoma patients treated with bevacizumab. Neuro-Oncology 89. Schoenegger K, Oberndorfer S, Wuschitz B, Struhal W,
18(12):1680–1687. https://doi.org/10.1093/neuonc/now086 Hainfellner J, Prayer D, Heinzl H, Lahrmann H, Marosi C,
76. Sanghani P, Ang BT, King NKK, Ren H (2018) Overall survival Grisold W (2009) Peritumoral edema on MRI at initial diagnosis:
prediction in glioblastoma multiforme patients from volumetric, an independent prognostic factor for glioblastoma? Eur J Neurol
shape and texture features using machine learning. Surg Oncol 16(7):874–878. https://doi.org/10.1111/j.1468-1331.2009.02613.x
27(4):709–714. https://doi.org/10.1016/j.suronc.2018.09.002 90. Lacroix M, Abi-Said D, Fourney DR, Gokaslan ZL, Shi W,
77. Lao J, Chen Y, Li Z-C, Li Q, Zhang J, Liu J, Zhai G (2017) A deep DeMonte F, Lang FF, McCutcheon I, Hassenbusch SJ, Holland
learning-based radiomics model for prediction of survival in glio- E, Hess K, Michael C, Miller D, Sawaya R (2001) A multivariate
blastoma multiforme. Sci Rep 7(1):10353. https://doi.org/10. analysis of 416 patients with glioblastoma multiforme: prognosis,
1038/s41598-017-10649-8 extent of resection, and survival. J Neurosurg 95(2):190–198.
78. Li Q, Bai H, Chen Y, Sun Q, Liu L, Zhou S, Wang G, Liang C, Li https://doi.org/10.3171/jns.2001.95.2.0190
Z-C (2017) A fully-automatic multiparametric radiomics model: 91. Skogen K, Schulz A, Dormagen JB, Ganeshan B, Helseth E,
towards reproducible and prognostic imaging signature for predic- Server A (2016) Diagnostic performance of texture analysis on
tion of overall survival in glioblastoma multiforme. Sci Rep 7(1): MRI in grading cerebral gliomas. Eur J Radiol 85(4):824–829.
14331. https://doi.org/10.1038/s41598-017-14753-7 https://doi.org/10.1016/j.ejrad.2016.01.013
79. Prasanna P, Patel J, Partovi S, Madabhushi A, Tiwari P (2017) 92. Macyszyn L, Akbari H, Pisapia JM, Da X, Attiah M, Pigrish V, Bi
Radiomic features from the peritumoral brain parenchyma on Y, Pal S, Davuluri RV, Roccograndi L, Dahmane N, Martinez-
treatment-naive multi-parametric MR imaging predict long versus Lage M, Biros G, Wolf RL, Bilello M, O'Rourke DM,
short-term survival in glioblastoma multiforme: preliminary find- Davatzikos C (2016) Imaging patterns predict patient survival
ings. Eur Radiol 27(10):4188–4197. https://doi.org/10.1007/ and molecular subtype in glioblastoma via machine learning tech-
s00330-016-4637-3 niques. Neuro-Oncology 18(3):417–425. https://doi.org/10.1093/
80. Bahrami N, Piccioni D, Karunamuni R, Chang YH, White N, neuonc/nov127
Delfanti R, Seibert TM, Hattangadi-Gluth JA, Dale A, Farid N, 93. Lahmiri S, Boukadoum M (2012) Automatic brain MR images
McDonald CR (2018) Edge contrast of the FLAIR hyperintense diagnosis based on edge fractal dimension and spectral energy
region predicts survival in patients with high-grade gliomas fol- signature. Conf Proc IEEE Eng Med Biol Soc 2012:6243–6246.
lowing treatment with bevacizumab. AJNR Am J Neuroradiol https://doi.org/10.1109/embc.2012.6347421
39(6):1017–1024. https://doi.org/10.3174/ajnr.A5620 94. Lahmiri S, Boukadoum M, Di Ieva A (2014) Detrended fluctua-
81. Bisdas S, Shen HC, Thust S, Katsaros V, Stranjalis G, Boskos C, tion analysis of brain hemisphere magnetic resonnance images to
Brandner S, Zhang JG (2018) Texture analysis- and support vector detect cerebral arteriovenousmalformations. In: 2014 IEEE
machine-assisted diffusional kurtosis imaging may allow in vivo International Symposium on Circuits and Systems (ISCAS), 1–5
gliomas grading and IDH-mutation status prediction: a prelimi- June 2014. pp 2409–2412 https://doi.org/10.1109/ISCAS.2014.
nary study Scientific Reports 8. https://doi.org/10.1038/s41598- 6865658
018-24438-4 95. Lahmiri S, Boukadoum M (2013) Automatic detection of
82. Bae S, Choi YS, Ahn SS, Chang JH, Kang SG, Kim EH, Kim SH, Alzheimer disease in brain magnetic resonance images using frac-
Lee SK (2018) Radiomic MRI phenotyping of glioblastoma: im- tal features. In: 2013 6th International IEEE/EMBS Conference
proving survival prediction. Radiology 289(3):797–806. https:// on Neural Engineering (NER), 6–8 Nov. 2013. p 1505–1508.
doi.org/10.1148/radiol.2018180200 https://doi.org/10.1109/NER.2013.6696231
83. Bahrami N, Hartman SJ, Chang YH, Delfanti R, White NS, 96. Di Ieva A (Ed.) (2016) The fractal geometry of the brain. Springer,
Karunamuni R, Seibert TM, Dale AM, Hattangadi-Gluth JA, New York
Piccioni D, Farid N, McDonald CR (2018) Molecular classifica- 97. Iftekharuddin KM, Jia W, Marsh R (2003) Fractal analysis of
tion of patients with grade II/III glioma using quantitative MRI tumor in brain MR images. Mach Vis Appl 13(5):352–362.
characteristics. J Neuro-Oncol 139(3):633–642. https://doi.org/10. https://doi.org/10.1007/s00138-002-0087-9
1007/s11060-018-2908-3 98. Zook JM, Iftekharuddin KM (2005) Statistical analysis of fractal-
84. Chaddad A, Sabri S, Niazi T, Abdulkarim B (2018) Prediction of based brain tumor detection algorithms. Magn Reson Imaging
survival with multi-scale radiomic analysis in glioblastoma pa- 23(5):671–678. https://doi.org/10.1016/j.mri.2005.04.002
tients. Med Biol Eng Comput 56(12):2287–2300. https://doi.org/ 99. Di Ieva A, Göd S, Grabner G, Grizzi F, Sherif C, Matula C,
10.1007/s11517-018-1858-4 Tschabitscher M, Trattnig S (2013) Three-dimensional suscepti-
85. Darbar A, Waqas M, Enam SF, Mahmood SD (2018) Use of bility-weighted imaging at 7 T using fractal-based quantitative
preoperative apparent diffusion coefficients to predict brain tumor analysis to grade gliomas. Neuroradiology 55(1):35–40. https://
grade. Cureus 10(3):e2284. https://doi.org/10.7759/cureus.2284 doi.org/10.1007/s00234-012-1081-1
86. Osman AFI (2019) A multi-parametric MRI-based radiomics sig- 100. Kristin RS, Alvord EC, Murray JD (2003) Dynamics of a model
nature and a practical ML model for stratifying glioblastoma pa- for brain tumors reveals a small window for therapeutic interven-
tients based on survival toward precision oncology. Front Comput tion. Discrete Contin Dynam Syst B 4(1):289–295. https://doi.org/
Neurosci 13:58. https://doi.org/10.3389/fncom.2019.00058 10.3934/dcdsb.2004.4.289
790 Neuroradiology (2020) 62:771–790

101. Sedivy R (1996) Fractal tumours: their real and virtual images. 114. Gevaert O, Xu J, Hoang CD, Leung AN, Xu Y, Quon A, Rubin
Wien Klin Wochenschr 108(17):547–551 DL, Napel S, Plevritis SK (2012) Non-small cell lung cancer:
102. Di Ieva A (2010) Angioarchitectural morphometrics of brain tu- identifying prognostic imaging biomarkers by leveraging public
mors: are there any potential histopathological biomarkers? gene expression microarray data–methods and preliminary results.
Microvasc Res 80(3):522–533. https://doi.org/10.1016/j.mvr. Radiology 264(2):387–396. https://doi.org/10.1148/radiol.
2010.09.005 12111607
103. Di Ieva A, Grizzi F, Ceva-Grimaldi G, Russo C, Gaetani P, Aimar 115. Shen W, Zhou M, Yang F, Dong D, Yang C, Zang Y, Tian J (2016)
E, Levi D, Pisano P, Tancioni F, Nicola G, Tschabitscher M, Learning from experts: developing transferable deep features for
Dioguardi N, Baena RRY (2007) Fractal dimension as a patient-level lung cancer prediction. In: Ourselin S, Joskowicz L,
quantitator of the microvasculature of normal and adenomatous Sabuncu MR, Unal G, Wells W (eds) Medical Image Computing
pituitary tissue. J Anat 211(5):673–680. https://doi.org/10.1111/j. and Computer-Assisted Intervention – MICCAI 2016, Cham,
1469-7580.2007.00804.x 2016. Springer International Publishing, p 124–131
104. Di Ieva A, Grizzi F, Gaetani P, Goglia U, Tschabitscher M, Mortini 116. Mamoshina P, Vieira A, Putin E, Zhavoronkov A (2016)
P, Rodriguez y Baena R (2008) Euclidean and fractal geometry of Applications of deep learning in biomedicine. Mol Pharm 13(5):
microvascular networks in normal and neoplastic pituitary tissue. 1445–1454. https://doi.org/10.1021/acs.molpharmaceut.5b00982
Neurosurg Rev 31(3):271–281. https://doi.org/10.1007/s10143- 117. Chu HH, Choi SH, Ryoo I, Kim SC, Yeom JA, Shin H, Jung SC,
008-01277 Lee AL, Yoon TJ, Kim TM, Lee S-H, Park C-K, Kim J-H, Sohn
105. Di Ieva A, Grizzi F, Tschabitscher M, Colombo P, Casali M, C-H, Park S-H, Kim IH (2013) Differentiation of true progression
Simonelli M, Widhalm G, Muzzio PC, Matula C, Chiti A, from pseudoprogression in glioblastoma treated with radiation
Rodriguez y Baena R (2010) Correlation of microvascular fractal therapy and concomitant temozolomide: comparison study of
dimension with positron emission tomography [(11)C]-methio- standard and high-b-value diffusion-weighted imaging.
nine uptake in glioblastoma multiforme: preliminary findings. Radiology 269(3):831–840. https://doi.org/10.1148/radiol.
Microvasc Res 80(2):267–273. https://doi.org/10.1016/j.mvr. 13122024
2010.04.003 118. Chang PD, Chow DS, Yang PH, Filippi CG, Lignelli A (2017)
106. Di Ieva A, Bruner E, Widhalm G, Minchev G, Tschabitscher M, Predicting glioblastoma recurrence by early changes in the appar-
Grizzi F (2012) Computer-assisted and fractal-based morphomet- ent diffusion coefficient value and signal intensity on FLAIR im-
ric assessment of microvascularity in histological specimens of ages. AJR Am J Roentgenol 208(1):57–65. https://doi.org/10.
gliomas. Sci Rep 2(1):429. https://doi.org/10.1038/srep00429 2214/ajr.16.16234
107. Bates DW, Saria S, Ohno-Machado L, Shah A, Escobar G (2014)
119. Ellingson BM, Cloughesy TF, Lai A, Nghiemphu PL, Mischel PS,
Big data in health care: using analytics to identify and manage
Pope WB (2011) Quantitative volumetric analysis of conventional
high-risk and high-cost patients. Health Aff 33(7):1123–1131.
MRI response in recurrent glioblastoma treated with bevacizumab.
https://doi.org/10.1377/hlthaff.2014.0041
Neuro-Oncology 13(4):401–409. https://doi.org/10.1093/neuonc/
108. Abrol S, Kotrotsou A, Salem A, Zinn PO, Colen RR (2017)
noq206
Radiomic phenotyping in brain cancer to unravel hidden informa-
tion in medical images. Top Magn Reson Imaging 26(1):43–53. 120. Zhou M, Hall LO, Goldgof DB, Russo R, Gillies RJ, Gatenby RA
https://doi.org/10.1097/rmr.0000000000000117 (2015) Decoding brain cancer dynamics: a quantitative histogram-
109. Cancer Genome Atlas Research Network (2008) Comprehensive based approach using temporal MRI. In: Progress in Biomedical
genomic characterization defines human glioblastoma genes and Optics and Imaging - Proceedings of SPIE. https://doi.org/10.
core pathways. Nature 455(7216):1061–1068. https://doi.org/10. 1117/12.2075545
1038/nature07385 121. Hygino da Cruz LC, Rodriguez I, Domingues RC, Gasparetto EL,
110. Prior FW, Clark K, Commean P, Freymann J, Jaffe C, Kirby J, Sorensen AG (2011) Pseudoprogression and pseudoresponse: im-
Moore S, Smith K, Tarbox L, Vendt B, Marquez G (2013) TCIA: aging challenges in the assessment of posttreatment glioma. Am J
an information resource to enable open science. Conf Proc IEEE Neuroradiol 32(11):1978–1985. https://doi.org/10.3174/ajnr.
Eng Med Biol Soc 2013:1282–1285. https://doi.org/10.1109/ A2397
embc.2013.6609742 122. Wardlaw G, Wong R, Noseworthy MD (2008) Identification of
111. Kalpathy-Cramer J, Freymann JB, Kirby JS, Kinahan PE, Prior intratumour low frequency microvascular components via BOLD
FW (2014) Quantitative imaging network: data sharing and com- signal fractal dimension mapping. Phys Med 24(2):87–91. https://
petitive algorithmvalidation leveraging the cancer imaging ar- doi.org/10.1016/j.ejmp.2008.01.006
chive. Transl Oncol 7(1):147–152. https://doi.org/10.1593/tlo. 123. Michallek F, Dewey M (2014) Fractal analysis in radiological and
13862 nuclear medicine perfusion imaging: a systematic review. Eur
112. Heilbrun ME (2013) Evaluating RadLex and real world radiology Radiol 24(1):60–69. https://doi.org/10.1007/s00330-013-2977-9
reporting: are we there yet? Acad Radiol 20(11):1327–1328.
https://doi.org/10.1016/j.acra.2013.09.011 Publisher’s note Springer Nature remains neutral with regard to jurisdic-
113. Tibshirani R (1996) Regression shrinkage and selection via the tional claims in published maps and institutional affiliations.
lasso. J R Stat Soc Ser B Methodol 58(1):267–288