Morphological signatures and genomic correlates in glioblastoma

Abstract

Large multimodal datasets such as The Cancer Genome Atlas present an opportunity to perform correlative studies of tissue morphology and genomics to explore the morphological phenotypes associated with gene expression and genetic alterations. In this paper we present an investigation of Cancer Genome Atlas data that correlates morphology with recently discovered molecular subtypes of glioblastoma. Using image analysis to segment and extract features from millions of cells, we calculate high-dimensional morphological signatures to describe trends of nuclear morphology and cytoplasmic staining in whole-slide images. We illustrate the similarities between the analysis of these signatures and predictive studies of gene expression, both in terms of limited sample size and high-dimensionality. Our top-down analysis demonstrates the power of morphological signatures to predict clinically-relevant molecular tumor subtypes, with 85.4% recognition of the proneural subtype. A complementary bottom-up analysis shows that self-aggregating clusters have statistically significant associations with tumor subtype and reveals the existence of remarkable structure in the morphological signature space of glioblastomas.

Full text

Page 1

MORPHOLOGICAL SIGNATURES AND GENOMIC CORRELATES IN GLIOBLASTOMA
Lee A.D. Cooper?, Jun Kong?, Fusheng Wang?, Tahsin Kurc?, Carlos S. Moreno†?, Daniel J. Brat†, Joel H. Saltz?
?Center for Comprehensive Informatics, Emory University, Atlanta, GA 30322
†Department of Pathology and Laboratory Medicine, Emory University, Atlanta, GA 30322
ABSTRACT
Large multimodaldatasets such as The Cancer Genome Atlas
present an opportunity to perform correlative studies of tis-
sue morphology and genomics to explore the morphological
phenotypes associated with gene expression and genetic al-
terations. In this paper we present an investigation of Cancer
Genome Atlas data that correlates morphology with recently
discovered molecular subtypes of glioblastoma. Using im-
age analysis to segment and extract features from millions
of cells, we calculate high-dimensional morphological sig-
natures to describe trends of nuclear morphology and cy-
toplasmic staining in whole-slide images. We illustrate the
similarities between the analysis of these signatures and pre-
dictive studies of gene expression, both in terms of limited
sample size and high-dimensionality. Our top-down analy-
sis demonstrates the power of morphological signatures to
predict clinically-relevant molecular tumor subtypes, with
85.4% recognition of the proneural subtype. A complemen-
tary bottom-up analysis shows that self-aggregating clusters
have statistically significant associations with tumor sub-
type and reveals the existence of remarkable structure in the
morphological signature space of glioblastomas.
Index Terms— bioinformatics, in silico, digital pathol-
ogy, image analysis, microscopy
1. INTRODUCTION
The National Institute of Health’s The Cancer Genome At-
las (TCGA)1project is producinga large multimodal datasets
containing pathology imaging, radiology, genomic, and clin-
ical data for glioblastoma and other tumor types [1]. These
datasets present a unique opportunity to perform correlative
studies of tissue morphology and genomics to explore the
morphological phenotypes associated with patient outcome,
gene expression and genetic alterations. To investigate these
relationshipsin glioblastomabrain tumorswe havedeveloped
an imaging system to analyze millions of cells within hun-
This research is supported in part by by NCI Contract No.
CO-12400 and 94995NBS23 and HHSN261200800001E, by Grant Number
R01LM009239 from the NLM, and by NSF CNS 0615155, 79077CBS10,
and 0403342.
1http://cancergenome.nih.gov/
N01-
dreds of whole-slide pathology images to determine the sur-
vival,geneexpressionandgenomiccorrelatesofcellularmor-
phology.
A study of TCGA molecular data has identified four
clinically-relevant subtypes of human glioblastoma [2]. The
Proneural (PN), Neural (NR), Classical (CL) and Mesenchy-
mal (MS) tumor subtypes are each defined by characteristic
gene expression profiles and genetic alterations including
mutations and chromosomal amplifications/deletions, and
differ in response to treatment and survival expectations.
This analysis also compared the gene expression profiles of
tumor subtypes to those of ordinary neural cell types and
found evidence suggesting that the glioblastoma subtypes are
reminiscent of distinct neural cell types.
In a previous study we examined the links between nu-
clear morphology and these four tumor subtypes in glioblas-
toma [3]. In this paper we extend this analysis to include fea-
tures describing cytoplasmic staining to produce an enhanced
morphologicalsignature describing both nuclear morphology
and the surrounding cytoplasm. A top-down analysis is per-
formed to determine effectiveness of enhanced signatures in
predicting tumor subtype. A bottom-up self-aggregation of
morphological signatures is also performed to examine the
links between the natural clustering of signatures and tumor
subtype.
2. MORPHOLOGICAL ANALYSIS
We have developed a system for the quantitative morpholog-
ical analysis of microanatomy in whole-slide images. The
system consists of a number of stages, as depicted in Figure
1. Nucleiarefirst segmentedandthesurroundingcytoplasmic
spaces are identified. A set of features is extracted to describe
the morphology and texture of each individual nucleus and
the staining characteristics of its cytoplasm. The segmented
objects and corresponding extracted features are stored in a
database for further analysis. Using the database, a morpho-
logical signature is calculated for each slide using first and
second order statistics. This workflow executes on a comput-
ing cluster and currently supports over 600 slides containing
an estimated 254 million nuclei.

Page 2

Fig. 1. Morphological analysis. Characterizations of nuclear shape
and cytoplasmic staining describing each cell are indexed to support
calculation of whole-slide morphological signatures.
2.1. Segmentation and Feature Extraction
The first stage of our analysis segments individual nuclei
using a combination of simple image processing operations.
Color images are first thresholded to identify and remove
blood and nontissue regions. The remaining areas are con-
verted to grayscale and a morphological reconstruction is
applied to remove debris. Overlapped nuclei are then sepa-
ratedusingawatershedoperation. Eachregioncorresponding
to a segmented nucleus is then dilated by a specific margin
to identify the surrounding cytoplasmic space.
segmentation, a collection of features is calculated for each
segmented nucleus to represent characteristics of nuclear
morphology and nuclear and cytoplasmic staining. A com-
plete list of these features is available in [3]. A color decon-
volution algorithm is first applied to the cytoplasmic space
to isolate hematoxylin and eosin stain signals into separate
channels prior to analyzing cytoplasmic intensity, texture,
and gradients [4]. Features representing morphology are not
calculatedfor the cytoplasmic space since the shape is strictly
derived from the nucleus boundary.
Following
2.2. Morphological Signature Calculation
The final stage of our analysis calculates a high-dimensional
morphological signature for each whole-slide image.
each feature fi, we calculate the first moment μi = E{fi},
and the second moment Ci,j = E{(fi− μi)(fj− μj)} for
each pair of features (i,j) ∈ {1,2,...,N} × {1,2,...,N}
where N is the number of features calculated for each seg-
mented entity. Second order statistics are necessary to rep-
resent the relationships between features describing nuclear
morphology, nuclear morphology and nuclear staining, or
nuclear morphology/stainingand cytoplasmic staining within
the whole-slide image.This produces an N(N + 3)/2-
dimensional feature vector to represent the morphology of
each slide in high dimensional space. Our current implemen-
tation uses N = 74 features resulting in a 2849-dimensional
signature.
For
2.3. Pathology Analytical Imaging Standards Database
The scale of our derived morphological datasets requires a
coordinated approach to data management to support virtual
experiments like those presented in this paper. Currently we
maintaina glioblastomadataset containingmorethan 600im-
ages with an average of 400 thousand nuclei per slide. Mor-
phologicalcharacterizationsproduce1.5GB/slideofmetadata
describing algorithm parameters, object boundaries, and fea-
tures. To address this issue the Pathology Analytical and
Imaging Standards (PAIS) model was developed to provide
flexible representation of analysis and characterization data
[5].The PAIS model is implemented using a relational
database to provide permanent management of analysis re-
sults and knowledge discovery through query support.
3. RESULTS AND DISCUSSIONS
We performed two experiments to explore the relationship
between the molecularly-defined tumor subtypes and the
morphological signatures derived from image analysis. In
the first experiment, a top-down classification was applied
to predict tumor subtype from morphological signature. In
the second experiment, a bottom-up unsupervised clustering
was performed to examine structure in the high-dimensional
morphological feature space. The bottom-up clusters were
further analyzed to determine if the tumor subtypes were
overrepresentedor underrepresented within each cluster.
To analyze morphology, 200 slides corresponding to 77
distinct patients were acquired from the National Cancer In-
stitute’s TCGA public data portal. These images are 20×
magnification whole-slide scans of formalin-fixed paraffin
embedded sections stained with hematoxylin and eosin. The
slides were analyzed using the pipeline from Section 2 to
calculate signatures of nuclear morphometryand cytoplasmic
staining. The tumor subtypes for a subset of these patients
were acquired from [2]. The subtype labels for the remaining
samples not included in [2] were predicted using Prediction
Analysis of Microarrays [6] with Affymetrix HTU133 gene
expression data also acquired from the TCGA portal. In all,
our dataset consisted of 49 PN samples, 38 NR samples, 61
CL samples and 52 MS samples.
3.1. Prediction of Molecular Tumor Subtype using Mor-
phological Signatures
We performed a classification experiment to examine the
power of morphometric signatures to predict molecular tu-
mor subtype. This experiment is similar to a gene-expression
study that aims to predict a given biological state or other
external truth using gene expression profiles obtained from
microarray experiments. The challenges encountered in pre-
dictive gene expression studies remain the same in our case.
Sample size is limited and so the high-dimensional space is
sparsely populated, making meaningful learning of structure

Page 3

Table 1. One-versus-one binary SVM classifier accuracy and
featurecount. Meanandstandarddeviationofpositivepredic-
tion are listed for 100 independent 10-fold cross validations.
Positive
Prediction
PN vs. NR94.4±1.2%
PN vs. CL91.4±1.0%
PN vs. MS 92.7±0.6%
NR vs. CL90.3±1.2%
NR vs. MS92.0±1.3%
CL vs. MS86.4±1.2%
Selected
Features
20
39
32
20
27
33
in feature-space difficult. We addressed these challenges us-
ingtechniquescommonlyfoundinpredictivegeneexpression
studies. Feature selection was used to reduce dimensionality
and increase signal-to-noiseratio by eliminating correlated or
irrelevant features. Classification results were validated us-
ing K-fold cross validation to provide evidence of classifier
generality.
Tumorsubtypeswereusedas groundtruthtotest andtrain
a multiclass support vector machine (SVM). The multiclass
SVM consisted of a collection of six binary SVMs trained to
distinguish between each pair of tumor subtypes. Given a test
sample, the collection of binary SVMs was evaluated and the
test sample was assigned to the subtype according to major-
ity vote. Exponential kernels were used for training SVMs
to deal with nonlinearities with σ ranging from 1 to 3. For
each binary SVM, the relevant features separating the tumor
subtype pair were identified using a margin-maximizingopti-
mizationwithL1-normpenaltytoencouragesparsityinselec-
tion [7]. Classifier training and testing was performedusing a
10-foldcrossvalidationwith stratificationtopreservethesub-
type proportions. The validation was repeated 100 times with
randomizedfold assignments to report the mean and standard
deviation of positive prediction for each class, defined as the
ratio of true positives to the sum of true positives and false
negatives.
Table 1 shows the positive prediction values for each of
the binary subtype SVMs and the number of features selected
for each. The multiclass SVM results are presented in Table
2. The mean and standard deviation of positive prediction are
presented along with the aggregated confusion matrix. Most
binaryclassifiers achieve 90%+ accuracywith only ≈ 1−2%
Table 2. Multiclass SVM accuracy and aggregate confusion.
Positive
Prediction
85.4±1.0%
72.3±2.5%
84.7±1.8%
79.2±2.2%
PN
4184
357
245
208
NR
98
2748
189
242
CL
319
536
5167
631
MS
299
132
499
4119
PN
NR
CL
MS
Sample Index
Sample Index
20 4060 80100 120140160 180200
20
40
60
80
100
120
140
160
180
200
Fig. 2. Consensus clustering. Bottom-up clustering of morphologi-
cal signatures reveals four distinct clusters.
of features selected in each case. The distinction between
classical and mesenchymal tumor subtypes is the least accu-
rate at 86.4%. The four-way multiclass classifier performs
reasonably for proneural, classical, and mesenchymal sub-
types, but has some problems with morphological classifica-
tion of the neural subtype. From the confusion matrix we
note that neural is most often mistaken for either proneural or
classical subtypes, but not the mesenchymal subtype.
We note that in similar predictive studies of gene expres-
sion, selected features (genes) offerbiological insight into the
differences between sample classes. These selected features
can be mined for biological associations using any number of
tools based on ontologies of function,disease, or localization.
Currently the biological interpretation of most morphological
features is unclear. We recognize that providing biological
insight is the central aim of correlative morphological stud-
ies and we are currently working on methods to annotate and
describe signature features in the glioblastoma context.
3.2. Consensus Clustering and Subtype Enrichment
A consensus clustering procedurewas applied to the morpho-
logical signatures to determine if natural morphological clus-
ters exist. This method uses repeated application of K-means
clustering to measure, for each pair of morphological signa-
tures, the frequency with which they are clustered together
[8]. Model selection on K was used to identify that K = 4
producedthe highest cophenetic correlation (0.98) among the
possibilities K = 2,3,4,5,6. The structure that corresponds
to this exceptional correlation is apparent in Figure 2. Each
row/column of Figure 2 contains the co-clustering frequen-
cies between all samples. The strong block-diagonal struc-
ture indicates that the four clusters reliably form regardless

Page 4

Table 3. Enrichment of tumor subtypes in bottom-up consensus clustering. The minimum p-value of either over/under repre-
sentation is listed with an O or U. Statistically significant results (p < 0.05) are highlighted below.
ConsensusProneuralNeural
Cluster% Samples
p-O/p-U
% Samples
p-O/p-U
1 46.2%
8e-4 O 2.6%
1.3e-3 U
2
46.0%
1.3e-3 O24.3%0.24 O
3
15.6%0.15 U 15.6%0.40 U
4
9.8%
4.6e-6 U 25.0%
3.5e-2 O
ClassicalMesenchymal
% Samples
18.0%
27.0%
25.0%
29.4%
% Samples
33.3%
2.7%
43.8%
35.9%
p-O/p-U
0.40 O
5.8e-6 U
6.1e-2 O
8.6e-2 O
p-O/p-U
0.14 U
0.51 O
0.54 U
0.20 O
of K-means initialization, and provides evidence that signifi-
cant structure exists within the high-dimensional morpholog-
ical signatures.
The content of the consensus clusters was further ana-
lyzed to determine relationships to the tumor subtypes. For
each consensus cluster we asked the following question: Are
any of the tumor subtypes significantly more or less frequent
in this cluster than can be expected? The distribution of the
entire dataset provides the expectation in terms of subtype
proportions (24.5% PN, 19% NR, 30.5% CL, and 26% MS).
Given these proportions, we used the hypergeometric distri-
bution to calculate the probability of a given subtype being
over-represented or under-represented in each consensus
cluster. The hypergeometric probability mass function p(k)
with parameters (Ω,S,C) models the probability of finding
k instances of a subtype in a consensus cluster containing
C samples if S total instances out of Ω total samples are
expected
?S
The significance of over or under representation was calcu-
lated by summing over p(k) to determine the probabilities
p − O,p − U of finding an over or under representation that
is at least as extreme as what is observed
p(k) =
k
??Ω−S
C
C−k
?Ω
?
?
.
p − O =
C
?
k=M
p(k),p − U =
M
?
k=0
p(k),
where M is the observed number of instances of the tumor
subtype in question within the consensus cluster.
The breakdown of the consensus clusters in terms of tu-
mor subtype is presented in Table 3. Cluster one (39 samples)
is enriched in proneural samples and neural samples are con-
spicuously absent. Cluster two (39 samples) is similarly en-
riched in proneural samples with classical samples underrep-
resented. Cluster three (32 samples) is enriched with classi-
cal subtypesamples and the proneuralsubtypeis significantly
underrepresented in cluster four (92 samples). Mesenchymal
samples are neither over or under represented in any cluster.
4. CONCLUSION
The paper presents a correlative analysis between high-
dimensional morphological signatures and four molecularly
defined subtypes of glioblastoma tumors. Top down classifi-
cation results suggest that signatures of nuclear morphology
and cytoplasmic staining have reasonable power to predict
tumor subtype. A separate bottom-up clustering analysis,
where morphological signatures are free to self-aggregate
irrespective of tumor subtype, shows clear structure in the
high-dimensional space of morphological signatures. In fu-
ture work we plan to investigate this structure further to
identify significant molecular correlates of the consensus
clustering. We are currently focused on developing methods
to provide biological insight into correlative morphological
studies, including ways to better represent the morpholo-
gies of the heterogeneous cell populations encountered in
whole-slide images.
5. REFERENCES
[1] Cancer Genome Atlas Research Network, “Comprehensive ge-
nomic characterization defines human glioblastoma genes and
core pathways,” Nature, vol. 455, pp. 1061–8, 2008.
[2] R.G.W. Verhaak, K.A. Hoadley, et al., “Integrated genomic
analysis identifies clinically relevant subtypes of glioblastoma
characterized by abnormalities in pdgfra, idh1, egfr, and nf1,”
Cancer Cell, vol. 17, pp. 98–110, 2009.
[3] L.A.D. Cooper, J. Kong, et al., “An integrative approach for in
silico glioma research,” Biomedical Engineering, IEEE Trans-
actions on, vol. 57, no. 10, pp. 2617–2621, oct. 2010.
[4] A.C. Ruifrok and D.A. Johnston, “Quantification of histochem-
ical staining by color deconvolution,” Analytical and Quanti-
tative Cytology and Histology, vol. 23, no. 4, pp. 291–9, aug.
2001.
[5] F. Wang, T. Kurc, P. Widener, T. Pan, J. Kong, L. Cooper, et al.,
“High-performance systems for in silico microscopy imaging
studies,” Data Integration in the Life Sciences. Lecture Notes in
Computer Science, vol. 6254/2010, pp. 3–18, 2010.
[6] R. Tibshirani, T. Hastie, B. Narasimhan, and G. Chu, “Diag-
nosis of multiple cancer types by shrunken centroids of gene
expression,” PNAS, vol. 99, pp. 6567–72, 2002.
[7] Y. Sun, S. Todorovic, and S. Goodison, “Local-learning-based
feature selection for high-dimensional data analysis,”
Transactions on Pattern Analysis and Machine Intelligence, vol.
32, pp. 1610–1626, 2010.
[8] S. Monti, P. Tamayo, J. Mesirov, and T. Golub, “Consensus
clustering: A resampling-based method for class discovery and
visualization of gene expression microarray data,”
Learning, vol. 52, no. 1-2, pp. 91–118, 2003.
IEEE
Machine