Automated segmentation of pretreatment computed tomography for evaluating intersite radiomic heterogeneity between platinum-resistant and platinum-sensitive ovarian cancer cohorts: a proof of concept study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Obstetrics and Gynecology, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India. Electronic address: [email protected].
- Department of Biostatistics, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India. Electronic address: [email protected].
- Department of Radiodiagnosis, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India. Electronic address: [email protected].
- Department of Medical Oncology, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India. Electronic address: [email protected].
- Department of Obstetrics and Gynecology, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India. Electronic address: [email protected].
Abstract
To evaluate the feasibility of utilizing automated segmentation of pretreatment contrast-enhanced computed tomography (CECT) images to characterize and evaluate intersite radiomic heterogeneity between platinum-resistant and platinum-sensitive cohorts in high-grade serous ovarian cancer. Pretreatment CECTs of 114 High Grade Serous Ovarian Cancer patients diagnosed from January 2016 till January 2023 were formatted with the dcm2niix tool. Masks for pelvic disease, omental disease and supradiaphragmatic and inguinal lymph node disease were created using the ovseg library. The surface area and volumes of these segments were extracted using numpy, pydicom and SimpleITK libraries. The patient data was divided into the platinum resistant group (clinical, radiological or biochemical relapse or progression within 6 months of completion of chemotherapy) and the platinum sensitive group (no relapse or progression within 6 months) and compared using the Chi-square and Mann-Whitney U tests. Machine learning classifiers were developed using LightGBM, CatBoost, and XGBoost algorithms within a modified four-fold stratified nested cross-validation framework to evaluate the capability of these multi-site features to distinguish between the two treatment-response cohorts. Models were selected using the cut offs- mean validation Area Under Curve (AUC) > 0.6, train-validation AUC gap < 0.1 for all folds and validation AUC standard deviation < 0.05. Surface area-volume ratios of omental disease, pelvic disease and supradiaphragmatic lymph nodes were found to differ significantly between the groups. Nine models satisfied the selection criteria. The best model used Catboost and surface area-volume ratio of omental disease and pelvic disease. Metrics- Mean Train AUC-0.69, Mean Validation AUC-0.665(95%CI- 0.591-0.739 Standard Deviation-0.038), Mean Train-Validation AUC Gap-0.049, Mean F1-0.608, Mean F2-0.705, Mean Precision-0.505, Mean Recall-0.797. Automated segmentation of pretreatment CECT images in high-grade serous ovarian cancer is feasible with ovseg and enables the automated extraction of multi-site imaging features that reflect distinct structural variations between treatment-response cohorts, demonstrating modest capabilities for retrospective cohort classification.