A Multimodal Deep Learning Framework Integrating CT Radiomics and Transcriptomics for Predicting Platinum Response in High-Grade Serous Ovarian Cancer

This study aimed to develop and evaluate a multimodal deep learning framework that integrates pre-treatment contrast-enhanced CT radiomics with tumor transcriptomics to predict platinum response in patients with high-grade serous ovarian cancer (HGSOC). A retrospective analysis was conducted using publicly available data from TCGA-OV and TCIA. The cohort included patients with baseline CT scans, bulk tumor RNA-seq, and clinical annotations for deriving platinum-free interval (PFI). A dual-tower neural network with attention-based fusion was designed to process radiomic features and RNA pathway activity scores. Model performance was rigorously assessed using nested cross-validation and compared against unimodal and clinical baselines. The multimodal fusion model achieved a mean ROC-AUC of [AUC_fusion] for predicting 6-month platinum resistance, outperforming radiomics-only (AUC_rad), transcriptomics-only (AUC_rna), and clinical-only (AUC_clin) models. The model demonstrated strong calibration, clinical utility on decision-curve analysis, and significant stratification of PFI in survival analysis. Performance remained robust across subgroup and sensitivity analyses. Integration of CT radiomics and transcriptomics using deep learning improves prediction of platinum response in HGSOC compared to single-modality approaches. This proof-of-concept study supports the potential of multimodal fusion, though external validation is required before clinical translation.