A mathematical-descriptor of tumor-mesoscopic-structure from computed-tomography images annotates prognostic- and molecular-phenotypes of epithelial ovarian cancer

Abstract

The five-year survival rate of epithelial ovarian cancer (EOC) is approximately 35-40% despite maximal treatment efforts, highlighting a need for stratification biomarkers for personalized treatment. Here we extract 657 quantitative mathematical descriptors from the preoperative CT images of 364 EOC patients at their initial presentation. Using machine learning, we derive a non-invasive summary-statistic of the primary ovarian tumor based on 4 descriptors, which we name "Radiomic Prognostic Vector" (RPV). RPV reliably identifies the 5% of patients with median overall survival less than 2 years, significantly improves established prognostic methods, and is validated in two independent, multi-center cohorts. Furthermore, genetic, transcriptomic and proteomic analysis from two independent datasets elucidate that stromal phenotype and DNA damage response pathways are activated in RPV-stratified tumors. RPV and its associated analysis platform could be exploited to guide personalized therapy of EOC and is potentially transferrable to other cancer types.

Conflict of interest statement

H.G. is an employee of AstraZeneca. The other authors declare no competing interests.

Figures

Fig. 1

Unsupervised clustering analysis of radiomic data in EOC. a Heatmap illustrating clustered matrix of sample-wise similarities (HH cohort) based on whole tumor radiomic profiles of primary ovarian tumors from all EOC histology. Black bar, group 1; yellow bar, group 2; green bar, group 3. b Distribution of high-grade serous ovarian carcinomas over patient groupings defined by similarities of radiomic profile (n = 84, p = 0.02, Fisher’s exact test). c Differences in the numbers of genes affected by copy-number aberration in tumors with spectral radiomic clusters. Black line, group 1; yellow line, group 2; green line, group 3. d Unsupervised hierarchical clustering of radiomic profile from primary HGSOC identified two distinct subgroups (blue and red as shown on the top row above heatmap). The associations between radiomic subgroups with the presence of ascites, lateral and tumor stage are indicated on the right. A summary of radiomic features are given on the y-axis. Blue bar, cluster 1; red bar, cluster 2. e Kaplan−Meier analysis of the radiomic subgroups with progression-free survival (n = 136). Blue line, cluster 1; red line, cluster 2. p value from log-rank test is included. HH Hammersmith Hospital, EOC epithelial ovarian cancer, HGSOC high-grade serous ovarian cancer

Fig. 2

Prognostic model based on radiomic profile in HGSOC. a Summary of univariate Cox regression between each radiomic feature and overall survival in the discovery set. Each black point represents the p value (y-axis) and hazard ratio (HR; x-axis) of a radiomic feature. Red horizontal dashed line indicates the false discovery rate (FDR) of 0.05; red vertical dashed line indicates HR of 1. b Least absolute shrinkage and selection operator (LASSO) regression analysis was performed to select radiomic features for prognostic model-building for HGSOC patients. Feature coefficients were plotted against shrinkage parameter (Lambda). c Partial likelihood deviance from Cox regression (y-axis) was generated under different shrinkage parameters (x-axis). Number of features selected corresponding to each lambda are given above the plot. Kaplan−Meier analyses were performed between radiomic prognostic vector (RPV) and overall survival in d HH discovery cohort (n = 136), e TCGA validation cohort (n = 70) and f HH validation cohort (n = 77). Red line, RPV low; green line, RPV medium; blue line, RPV high. p values are given by log-rank test. HGSOC high-grade serous ovarian cancer, TCGA the Cancer Genome Atlas

Fig. 3

Molecular characteristics associated with RPV in HGSOC. Gene set enrichment analysis identified a RPV-positively correlated biological pathways and b RPV-negatively correlated biological pathways from KEGG pathway database (FDR < 0.05). NES normalized enrichment score. c Volcano plot showing the differential expressed genes between stroma and tumor epithelial component from a public dataset, GSE40595. The genes that positively correlated with RPV are highlighted in red (r > 0.3, Spearman correlation); genes that are negatively correlated with RPV are highlighted in blue (r < −0.3, Spearman correlation). d Heatmap showing correlation of protein expression (Fibronectin, Rad51 and FoxM1) with RPV for 47 cases in the TCGA validation dataset. The significance between these protein features with RPV was indicated for 119 cases in the HH cohort and with eRPV from 353 additional TCGA cases. Top panel, RPV ranked from low to high (left to right) and their corresponding eRPV (light blue). Lower panel, protein expression level of Fibronectin, Rad51 and FoxM1. p values are given by one-sided Spearman’s correlation test as validation of the transcriptomic analyses. e Clinical, histological and genetic characteristics associated with RPV in the TCGA and HH cohorts. Each rectangle block represents one patient in the TCGA validation dataset. The significance of association between these characteristics with RPV in the TCGA validation dataset, HH cohort and eRPV in additional TCGA dataset is indicated on the right side. The significance is indicated on the right from Kruskal−Wallis test (molecular subtype) or two-tailed Wilcoxon rank-sum test (others). The association between RPV and stromal component is shown in (a), (c), (d) and (e); The association between RPV and proliferation or DNA damage response is highlighted in (b) and (d). ***p < 0.001, **p < 0.01, *p < 0.05, ns p > 0.1. RPV radiomic prognostic vector, HGSOC high-grade serous ovarian cancer, TCGA the Cancer Genome Atlas, KEGG Kyoto Encyclopedia of Genes and Genomes

Fig. 4

The reliability and reproducibility of radiomic profile. a Principal component analysis (PCA) plot of radiomic profile by scanner manufacturers. b Radiomic feature-wise correlations of two independent cohorts. x-axis indicates the correlation between any two radiomic features in the HH cohort, y-axis indicates the corresponding feature-wise correlation in the TCGA cohort. The Pearson correlation coefficient and p value is indicated. c An illustration of erosion (red) and dilation (green) on the original segmented CT scan (blue). d Impact of segmentation deformations on RPV. x-axis indicates the range of deformations from erosion by 4 voxels (−4) to dilation by 4 voxels (+4) and 0 is the original segmentation. y-axis indicates the difference between deformed RPV and original RPV. HH Hammersmith Hospital, TCGA the Cancer Genome Atlas, CT computed tomography, RPV radiomic prognostic vector