bioRxiv2019
AbstractO_ST_ABSBackgroundC_ST_ABSHigh grade serous ovarian carcinoma shows marked intra-tumoral heterogeneity which is associated with decreased survival and resistance to platinum-based chemotherapy. Pre-treatment quantification of spatial tumor heterogeneity by multiple tissue sampling is not clinically feasible. Using standard-of-care CT imaging to non-invasively quantify heterogeneity could have high clinical utility and would be highly cost-effective. Texture analysis measures local variations in computed tomography (CT) image intensity. Haralick texture methods are typically used to capture the heterogeneity of entire lesions; however, this neglects the possible presence of texture habitats within the lesion, and the differences between metastatic sites. The primary aim of this study was to develop texture analysis of intra-site and inter-site spatial heterogeneity from standard-of-care CT images and to correlate these measures with clinical and genomic features in patients with HGSOC. Methods and findingsWe analyzed the data from a retrospective cohort of 84 patients with HGSOC consisting of 46 patients from Memorial Sloan Kettering Cancer Center (MSKCC) and 38 non-MSKCC cases selected from The Cancer Imaging Archive (TCIA). Inclusion criteria consisted of FIGO stage II-IV HGSOC, attempted primary cytoreductive surgery, intravenous contrast-enhanced CT of abdomen and pelvis performed prior to surgery and availability of molecular tumor data analysed as per the Cancer Genome Atlas (TCGA) Research Network ovarian cancer project. Manual segmentation and image analysis was performed on 463 metastatic tumor sites from 84 patients. In the MSKCC cohort the median number of tumor sites was 7 (interquartile range 5-9) and 4 (interquartile range 3-4) in the TCIA patients. Sub-regions were produced within each tumor site by grouping voxels with similar Haralick texture using the Kernel K-means method. We derived statistical measures of intra- and inter-site tumor heterogeneity (IISTH) including cluster sites entropy (cSE), cluster sites standard deviation (cluDev) and cluster sites dissimilarity (cluDiss) from sub-regions identified within and between individual tumor sites. Unsupervised clustering was used to group patient IISTH measures into low, medium, high, and ultra-high heterogeneity clusters from each cohort. The IISTH measure cluDiss was an independent predictor of progression-free survival (PFS) in multivariable analysis in both datasets (MSKCC hazard ratio [HR] 1.04, 95% CI 1.01-1.06, P = 0.002; TCIA HR 1.05, 95% CI 1.00-1.10, P = 0.049). Low and medium IISTH clusters were associated with longer PFS in multivariable analysis (MSKCC HR 2.94, 90% CI 1.29-6.70, P = 0.009, TCIA HR 5.94, 95% CI 1.05-33.6, P = 0.044). IISTH measures were robust to differences in the CT imaging systems. Average Haralick textures contrast (TCIA HR 1.08, 95% CI 1.01-1.10, P = 0.019) and homogeneity (TCIA HR 1.09, 95% CI 1.02-1.16, P = 0.008) were associated with PFS in mutivariate analysis only in the TCIA dataset. All other average Haralick textures and total tumor volume were not associated with PFS in either dataset. ConclusionsTexture measures of intra- and inter-site tumor heterogeneity from standard of care CT images are correlated with shorter PFS in HGSOC patients. These quantitative methods are independent of the CT imaging system and can thus be applied in clinical practice. The methodology proposed here enables the non-invasive quantification of intra-tumoral heterogeneity and disease stratification for future experimental medicine studies and clinical trials, particularly in cases where total tumour volume and averaged textures have low predictive power. Author summaryO_ST_ABSWhy was this study done?C_ST_ABSO_LITumor heterogeneity is a feature of many solid malignancies including ovarian cancer. C_LIO_LIRecent genomic research suggests that intra-site tumor heterogeneity (heterogeneity within a single tumor site) and inter-site tumor heterogeneity (heterogeneity between different metastatic sites in the same patient) correlate with clinical outcome in HGSOC. C_LI What did the researchers do and find?O_LIWe developed quantitative and non-invasive image-analysis based measures for predicting outcome in HGSOC patients by combining image-based information from within and between multiple tumor sites. C_LIO_LIUsing datasets from two sources, we demonstrate that these image-based tumor heterogeneity measures predict progression free survival in patients with HGSOC. C_LI What do these findings mean?O_LINon-invasive measures of CT image heterogeneity may predict outcomes in HGSOC patients. C_LIO_LIWider application of these CT image heterogeneity measures could prove useful for stratifying patients to different therapies given that total tumour volume and averaged textures have low predictive power. C_LI