Silu Zhang1, Zoltan Patay1, Bogdan Mitrea2, Angela Edwards1, Lydia McColl Makepeace1, and Matthew A. Scoggins1
1Diagnostic Imaging, St. Jude Children's Research Hospital, Memphis, TN, United States, 2Activ Surgical, Boston, MA, United States
Synopsis
Diffuse
intrinsic pontine glioma (DIPG) is a pediatric brain tumor with very poor
prognosis. In this study, we identify two subtypes of DIPG based on radiomic
features. The two subtypes show a significant difference in survival rates.
Subtype 1 has a mean progression-free survival (PFS) and overall survival (OS)
of 8.9 and 12.7 months, respectively. Subtype 2 has a mean PFS and OS of 5.7
and 9.1 months, respectively. Our results suggest that shape features and
intensity features extracted from FLAIR and T1-post contrast predict the prognosis
of DIPG.
INTRODUCTION
Diffuse intrinsic
pontine glioma (DIPG) is a pediatric brain tumor with very poor prognosis. The
median overall survival (OS) of patients is less than 1 year, and the 2-year
survival rate is less than 10%.1 Studies have identified subgroups of
DIPG based on genetic profiles.2-4 In this study, we identify DIPG
subtypes based on radiomic features extracted from magnetic resonance imaging (MRI).METHODS
This
study analyzed baseline imaging from 126 patients diagnosed with DIPG treated
at St. Jude Children's Research Hospital. The workflow of image preprocessing
and analysis is shown in Figure 1.
For each patient, 4 MRI sequences were used for
analysis: T1-weighted (T1-pre), T2-weighted (T2), T1-post contrast (T1-post),
and T2 fluid-attenuated inversion recovery (FLAIR). All MRIs were co-registered,
smoothed, bias corrected and intensity normalized using WhiteStripe5 normalization. Tumors
were first auto segmented using Cascaded Anisotropic Convolutional Neural
Networks6,7 and then manually adjusted by our image processing
specialists if necessary. Radiomic features were extracted from segmented tumor
regions using the PyRadiomics package.8 Original images and filtered
images were used for calculating first order statistics, shape, and texture features.
The bin width used for image discretization was customized for each image filter, such that the number of bins for each volume of interest was between 16 and 128
bins, as per PyRadiomics guidelines. This setup resulted in a total of 4746
features. Feature selection was then performed
by measuring the dependence between
the progression free survival (PFS) and each radiomic feature using distance correlation.9,10 Unlike Pearson correlation
that measures only linear dependence (resulting in false negatives) and is sensitive
to outliers (resulting in false positives),
distance correlation measures
the distance between
2 distributions and therefore
ensures independence screening.11 Patients were randomly divided
into training (80%) and validation (20%) sets, and feature selection was
performed in only the training set. Features with P<0.05 were
selected. Selected features were used for clustering analysis using
unsupervised K-means clustering.12,13 The algorithm was first run on
training data, and the centroid of each cluster
was used to predict the cluster membership for validation data. To determine the best value
for K, the K-means
were run 100 times with random starting
seeds on the training data for each value of K (2–4) and the K
value resulting in the most stable cluster
assignments was selected. The
reproducibility of cluster
assignment was measured
by the average adjusted
Rand index.14 After K
was determined, the most frequent clustering assignment from the 100 runs was used for predicting validation data and survival
analysis. The subtype
determined by K-means clustering was then evaluated
by survival prediction. Both PFS and OS were used as outcomes. Kaplan–Meier
survival curves were plotted, and the Cox proportional hazard model was used to
calculate the hazard ratio. The significance of difference in survival rates
between different subtypes was tested by the log rank test.
RESULTS
A total of 549 radiomic features
with P<0.05 were selected. The K-means clustering analysis was
first performed on the training set using the selected features, with K=2
providing the most reproducible results. For K=2, 3, and 4, the average
adjusted Rand indexes were 0.97, 0.87, and 0.79, respectively. The partition
results of K=2 on training and test data are shown in Figure 2. Frequencies
of the 2 subtypes in our data are shown in Table
1, with the frequency of subtype 1 being 76% and subtype 2 being 24%.
Survival rates of the 2 subtypes are shown in Figure 3. There was a significant
difference in Kaplan–Meier curves of the 2 subtypes in both training data and
validation data. For PFS, results showed P=0 on training and P=0.012
on validation. The hazard ratio (HR) between subtype 2 and subtype 1 was 2.48
for training and 3.53 for validation. Because of the high correlation between
PFS and OS, subtypes assigned using PFS-related features could also accurately predict OS: P=0.006, HR=1.94 on training,
and P=0.026, HR=3.05 on validation. The mean PFS and OS of the 2 subtypes
are listed in Table 1. The mean PFS and OS of subtype 1 was 8.9 and 12.7
months and for subtype 2 was 5.7 and 9.1 months,
respectively.DISCUSSION
Of 15 shape
features extracted, 7 were selected. Of 542 intensity-based features selected, 268 were
extracted from FLAIR, 168 from T1-post, 64 from T2, and 42 from T1. These
results suggest that tumor shape, FLAIR, and T1-post intensity
features (particularly nonuniformity, variance, energy, and contrast) inform the
prognosis of DIPG.CONCLUSION
MRI imaging is the primary tool for diagnosis of DIPG. In one of the largest single institution
cohorts of DIPG patients, radiomic features at baseline identified 2 subtypes
with differing prognoses.
Subtype 1 had a mean
PFS and OS of 8.9 and 12.7 months, respectively. Subtype 2 had a mean PFS and OS
of 5.7 and 9.1 months, respectively.
A critical next step is to determine the
relationship between these imaging-derived subtypes and genetic profiles. Biopsies
have been historically performed infrequently due to the location of the tumor.
The potential of subtyping by imaging information would be valuable. Acknowledgements
No acknowledgement found.References
- Hargrave, D., Bartels, U., & Bouffet, E. (2006). Diffuse brainstem glioma in children: critical review of clinical trials. The lancet oncology, 7(3), 241-248.
- Khuong-Quang, D. A., Buczkowicz, P., Rakopoulos, P., Liu, X. Y., Fontebasso, A. M., Bouffet, E., ... & Bourgey, M. (2012). K27M mutation in histone H3. 3 defines clinically and biologically distinct subgroups of pediatric diffuse intrinsic pontine gliomas. Acta neuropathologica, 124(3), 439-447.
- Buczkowicz, P., Hoeman, C., Rakopoulos, P., Pajovic, S., Letourneau, L., Dzamba, M., ... & Zuccaro, J. (2014). Genomic analysis of diffuse intrinsic pontine gliomas identifies three molecular subgroups and recurrent activating ACVR1 mutations. Nature genetics, 46(5), 451-456.
- Castel, D., Philippe, C., Calmon, R., Le Dret, L., Truffaux, N., Boddaert, N., ... & Mackay, A. (2015). Histone H3F3A and HIST1H3B K27M mutations define two subgroups of diffuse intrinsic pontine gliomas with different prognosis and phenotypes. Acta neuropathologica, 130(6), 815-827.
- Shinohara, R.T., et al., Statistical
normalization techniques for magnetic resonance imaging. NeuroImage:
Clinical, 2014. 6: p. 9-19.
-
Wang,
G., et al. Automatic brain tumor segmentation
using cascaded anisotropic convolutional neural networks. in International MICCAI Brainlesion Workshop.
2017. Springer.
-
Gibson, E., et al., NiftyNet: a
deep-learning platform for medical imaging. Computer methods and programs
in biomedicine, 2018. 158: p.
113-122.
-
Van
Griethuysen, J.J., et al., Computational
radiomics system to decode the radiographic phenotype. Cancer research,
2017. 77(21): p. e104-e107.
-
Székely, G.J. and M.L. Rizzo, Brownian
distance covariance. The annals of applied statistics, 2009. 3(4): p. 1236-1265.
-
Székely,
G.J., M.L. Rizzo, and N.K. Bakirov, Measuring
and testing dependence by correlation of distances. The annals of
statistics, 2007. 35(6): p.
2769-2794.
-
Li, R., W. Zhong, and L. Zhu, Feature
screening via distance correlation learning. Journal of the American
Statistical Association, 2012. 107(499):
p. 1129-1139.
-
Lloyd,
S., Least squares quantization in PCM.
IEEE transactions on information theory, 1982. 28(2): p. 129-137.
- MacQueen, J. Some methods for classification and analysis of multivariate
observations. in Proceedings of the
fifth Berkeley symposium on mathematical statistics and probability. 1967.
Oakland, CA, USA.
- Hubert, L. and P. Arabie, Comparing
partitions. Journal of classification, 1985. 2(1): p. 193-218.