End-to-end integrative segmentation and radiomics prognostic models for risk stratification of high-grade serous ovarian cancer: a retrospective multicohort study
Menée à partir de 605 clichés de tomographies numériques avec rehaussement de contraste réalisées sur des patientes atteintes d'un carcinome ovarien séreux de haut grade, cette étude évalue la performance de modèles d'apprentissage automatique de type "end-to-end" pour segmenter les clichés tomographiques et établir un pronostic
Background : Valid stratification factors for patients with epithelial ovarian cancer are still lacking and individualisation of care remains an unmet need. Radiomics derived from routine contrast enhanced CT (CE-CT) is an emerging, highly promising approach for more accurate prognostic models to improve the preoperative stratification of patients with high-grade serous ovarian carcinoma (HGSOC). However, fine manual segmentation requirements limit its use. To enable broader implementation of CE-CT, we aimed to develop an end-to-end model that automates segmentation processes and prognostic evaluation algorithms in HGSOC.
Methods : In this multicohort study, we retrospectively collected and segmented 605 CE-CT scans of patients with epithelial ovarian cancer and individuals who underwent upfront debulking surgery from the UK, Germany, and the USA from June, 2004, to July, 2018. The development cohort comprised patients from Hammersmith Hospital, London, UK (n=211), which was split with a ratio of 7:3 for training and validation. We used CT images and clinical data from The Cancer Imaging Archive (TCIA) digital repository, USA (n=71) and Kliniken Essen-Mitte (KEM), Essen, Germany (n=323) as test sets. We developed an automated segmentation model for primary ovarian cancer lesions in CE-CT scans with U-Net based architectures. We computed radiomics data from the CE-CT scans. For overall survival prediction, combinations of 13 feature reduction methods and 12 machine learning algorithms were developed on the radiomics data and compared with convolutional neural network models trained on CE-CT scans. We compared our model with a published radiomics model for HGSOC prognosis, the radiomic prognostic vector. In the Hammersmith Hospital and TCIA cohorts, additional histological diagnosis, transcriptomics, proteomics, and copy number alterations were collected; and correlations with the best-performing overall survival model were identified. Predicted probabilities of the best-performing overall survival model were dichotomised using k-means clustering to define high-risk and low-risk groups.
Findings : Using the combination of segmentation and radiomics as an end-to-end framework, the prognostic model improved risk stratification of HGSOC over Cancer Antigen 125, residual disease, International Federation of Gynaecology and Obstetrics staging, and the previously reported radiomics prognostic vector. Calculated from predicted and manual segmentations, our automated segmentation model achieved dice scores of 0·90 for the Hammersmith Hospital validation cohort, 0·88 for the TCIA test set and 0·80 for the KEM test set. The top-performing radiomics model of overall survival achieved a concordance index (C-index) of 0·66 (SE 0·06) for the Hammersmith Hospital validation cohort, 0·72 (0·05) for the TCIA test set, and 0·60 (0·01) for the KEM test set. In a multivariable model of this radiomics model with age, residual disease, and stage, the C-index values were 0·73 (SE 0·07) for the Hammersmith Hospital validation cohort, 0·73 (0·06) for the TCIA test set, and 0·73 (0·03) for the KEM test set. After dichotomising radiomics-based model predictions into risk groups, high-risk groups had poor prognosis (overall survival) with hazard ratios of 4·81 (95% CI 1·61–14·35) for the Hammersmith Hospital validation cohort, 6·34 (2·08–19·34) for the TCIA test set, and 1·71 (1·10–2·65) for the KEM test set, adjusted for stage, age, performance, and residual disease. These groups were linked to an invasive phenotype involving SNARE interactions in vesicular transport and activated MAPK pathways.
Interpretation : An automated end-to-end artificial intelligence pipeline using CE-CT radiomics improves risk stratification in HGSOC and outperforms existing clinical and radiomic benchmarks. This approach could support non-invasive, scalable prognostic assessment to inform treatment planning and trial enrolment.
Funding : The UK Medical Research Council, AstraZeneca, Cancer Research UK, National Institute for Health and Care Research, UK Research and Innovation, the National Natural Science Foundation of China, University of Hong Kong, Material Innovation Institute for Life Sciences and Energy, and the North West London Pathology Research & Education Board.
The Lancet Digital Health , article en libre accès, 2026