An automated imaging pipeline has been developed to support preoperative risk stratification in high-grade serous ovarian carcinoma (HGSOC) using routine contrast-enhanced CT. The approach integrates deep learning–based segmentation of primary ovarian lesions with radiomics-driven survival modelling. Designed to reduce reliance on labour-intensive manual contouring, the framework was trained and internally validated in a UK cohort and externally assessed in independent datasets from the United States and Germany. In addition to imaging features, selected clinical variables were incorporated into multivariable analyses. Performance was benchmarked against established prognostic indicators, including Cancer Antigen 125, residual disease status and FIGO stage, as well as previously described radiomics models. The objective was to evaluate whether a fully automated workflow could provide reproducible, clinically relevant risk stratification for overall survival across heterogeneous, multicentre imaging data.
Multicentre Cohorts and Imaging Standardisation
The analysis included more than 600 contrast-enhanced CT examinations from adults with histologically confirmed primary HGSOC across 3 institutions. The largest cohort originated from a German centre, with additional cases from a UK specialist hospital and a US imaging repository. The UK dataset was divided into training and validation subsets, while the US and German cohorts were reserved for external testing. Eligibility required a visible primary adnexal mass on portal venous phase CT and no prior oncological surgery or systemic therapy. Scans lacking appropriate contrast phases or containing artefacts that precluded reliable segmentation were excluded.
Must Read: Automated CT Pipeline Advances Sacral Tumour Classification
Manual delineation of primary adnexal masses, including both solid and cystic components, was performed by experienced radiologists using dedicated software. Ascites and indeterminate regions were excluded. When bilateral disease was present, lesions were segmented either jointly or separately depending on confluence. Imaging data were acquired across multiple scanners and protocols and were resampled to isotropic voxel spacing to support radiomics standardisation. Clinical information, including staging and postoperative residual disease, was obtained from medical records and multidisciplinary discussions. Overall survival was defined from surgery to death or last follow-up.
Automated Segmentation and Radiomics Modelling
Automated lesion segmentation was implemented using nnU-Net with a combination of two-dimensional and three-dimensional configurations. Models were trained with cross-validation and ensembled to improve robustness. Segmentation performance was evaluated using dice similarity coefficients across cohorts. Median dice scores were high in the training dataset and remained acceptable in validation and external test sets, although lower values were observed in the German cohort. Inter-reader variability assessed in a subset of cases demonstrated substantial agreement between radiologists.
From the segmented regions, several hundred texture-based radiomics features were extracted using fixed Hounsfield Unit discretisation. Multiple feature-selection techniques and survival modelling strategies were compared, including tree-based, regularised and ensemble approaches. Hyperparameters were optimised using cross-validation with concordance index as the primary metric. A bootstrap-ensemble model combining permutation-variable importance random forest feature selection with a random survival forest predictor achieved the strongest validation performance among the evaluated approaches. Concordance indices were highest in the training cohort and remained moderate in validation and external datasets. When age, stage and residual disease were added to the radiomics model, performance improved in external cohorts.
Risk Stratification and Biological Associations
Predicted survival probabilities were stratified into risk groups using clustering methods. A binary threshold distinguished high-risk from low-risk patients with significant separation in overall survival across cohorts. A three-tiered stratification was also explored to identify an intermediate-risk group. Survival differences between strata were statistically significant in training, validation and external test datasets.
Model performance was compared with conventional clinical variables, including age, CA-125, FIGO stage and residual disease, as well as selected genomic alterations. Deep learning survival models trained directly on contrast-enhanced CT without predefined segmentations were also developed using modified convolutional neural networks. While these approaches demonstrated prognostic capability, the radiomics-based ensemble model showed the highest validation concordance. Sensitivity analyses indicated that modest alterations to the segmentation boundaries did not materially change radiomics performance.
Associations between predicted risk and tumour biology were explored using histological, transcriptomic and proteomic data in subsets of patients. High-risk predictions were linked to pathways related to vesicular transport and MAPK signalling. Increased tumour cellularity was observed in the high-risk group within the UK cohort, and predicted probabilities correlated positively with cellularity but not with fibronectin expression. Proteomic analyses in the US cohort identified correlations between predicted risk and several proteins, including positive association with STAT5ALPHA and negative associations with proteins such as JNK2, YB1, RAD51, MTORPS448 and HER3PY1298.
An automated end-to-end CT radiomics pipeline demonstrated feasibility for risk stratification in HGSOC across multinational cohorts. The framework combined deep learning–based segmentation with ensemble survival modelling to generate clinically interpretable risk groups associated with overall survival. External validation supported moderate prognostic performance, which improved when integrated with selected clinical variables. Biological analyses suggested links between imaging-derived risk and tumour cellularity, signalling pathways and proteomic profiles. The findings indicate potential for automated radiomics to contribute to preoperative prognostic assessment, while highlighting the need for further validation in broader populations and continued evaluation of alternative deep learning survival strategies.
Source: The Lancet Digital Health
Image Credit: iStock
