Integrating ultrasound-CT-MR for preoperative multi-task prediction in ovarian cancer: achieving diagnostic parity with multidisciplinary team consensus

BackgroundOvarian cancer, with its high mortality rate, requires precise preoperative integration of ultrasound, computed tomography (CT), and magnetic resonance imaging (MRI) for guide therapy, yet current clinical workflows rely on specialized multidisciplinary team (MDT) evaluations, a barrier in resource-limited settings. Current artificial intelligence (AI)-driven studies, primarily focused on single-modality analysis and distinguishing malignant from benign ovarian lesions, fail to address the complexity of multimodal decision-making.MethodsWe retrospectively collected preoperative ultrasound, CECT, and MRI data from 1844 patients with ovarian lesions. Using radiomics and AI, we developed the OVUCM system, which integrates multi-modal features through intermediate fusion and machine learning. This system predicts five clinical tasks, which are benign vs non-benign, borderline vs malignant, epithelial vs non-epithelial, FIGO I-II vs FIGO III-IV, and HGSOC vs non-HGSOC, using nested cross-validation for optimal model selection. Our system's performance was compared to the MDT at the National Regional Medical Centre for Obstetrics and Gynaecology.FindingsThe OVUCM system simultaneously predicts five clinical endpoints: benign vs non-benign (AUC 0·929), borderline vs malignant (0·889), non-epithelial vs epithelial (0·897), International Federation of Gynaecology and Obstetrics (FIGO) staging I-II vs III-IV (0·853), and non-high-grade serous ovarian cancer (HGSOC) vs HGSOC (0·847). This system achieved diagnostic parity with MDT consensus (ΔAUC 0-0·25; Delong test, P>0·05 in five tasks), while outperforming independent gynaecologists in some tasks. External validation in 102 patients from a general hospital confirmed its generalizability (AUCs 0·899-0·974 in five tasks).InterpretationThe OVUCM, a multi-task and multi-modality system standardizes preoperative workflows by bridging the gap between single-modality tools and the comprehensive, multidisciplinary decision-making required for personalized therapy. Its ability to replicate MDT expertise in resource-limited settings positions it as a transformative tool for global health equity.FundingMedical Health Science and Technology Plan of Zhejiang Provincial Health Commission (WKJ-ZJ-2408) and Zhejiang Key Research & Development Program (2023C03G1753905, 2024SSYS0010).