Foundation Model-empowered Multimodal Prediction of Platinum Resistance in High-grade Serous Ovarian Cancer: A Multicenter Retrospective Study.
Authors
Affiliations (9)
Affiliations (9)
- Department of MRI, the First People's Hospital of Yunnan Province, the Affiliated Hospital of Kunming University of Science and Technology, Kunming, Yunnan 650032, China (Q.B., J.Y., L.W.).
- Fudan University, Shanghai, China (L.Q.).
- Department of Radiology, the First Affiliated Hospital of Chongqing Medical University, Chongqing, China (Y.L.).
- Department of Pathology, the First People's Hospital of Yunnan Province, the Affiliated Hospital of Kunming University of Science and Technology, Kunming, Yunnan, China (Q.W.).
- Department of Pathology, the Third Affiliated Hospital of Kunming Medical University, Yunnan Cancer Hospital, Yunnan Cancer Center, Kunming, Yunnan, China (Q.M.).
- Department of Radiology, Zhongda Hospital, School of Medicine, Southeast University, Nanjing, Jiangsu, China (Y.W.); Jiangsu Key Laboratory of Intelligent Medical Image Computing, School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing, China (Y.W.).
- Department of Radiology, the Third Affiliated Hospital of Kunming Medical University, Yunnan Cancer Hospital, Yunnan Cancer Center, Kunming, Yunnan, China (C.A.).
- Department of Medical Oncology, the First People's Hospital of Yunnan Province, the Affiliated Hospital of Kunming University of Science and Technology, Kunming, Yunnan, China (K.M.).
- Department of MRI, the First People's Hospital of Yunnan Province, the Affiliated Hospital of Kunming University of Science and Technology, Kunming, Yunnan 650032, China (Q.B., J.Y., L.W.). Electronic address: [email protected].
Abstract
To develop and validate a foundation model (FM)-empowered multimodal framework to predict platinum resistance in patients with high-grade serous ovarian cancer (HGSOC). This multicenter retrospective cohort study included 744 HGSOC patients who underwent primary debulking surgery across four centers. Perioperative multimodal data were collected, including clinical variables, preoperative pelvic magnetic resonance imaging (MRI), and postoperative hematoxylin and eosin (H&E)-stained whole slide images (WSIs). Pretrained radiology and pathology FMs were used as frozen feature extractors, and attention-based aggregation was applied to summarize variable-length MRI slices and WSI tiles. A cross-modal fusion module integrated modalities and supported missing-modality inference. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and compared with late-fusion (LF) baselines. In the internal validation cohort, unimodal and multimodal FM variants achieved AUCs ranging from 0.708 to 0.780, and in external test cohorts A, B, and C achieved AUCs ranging from 0.653 to 0.790. The trimodal FM (clinical+MRI+WSI) achieved the highest AUC in all cohorts (internal validation: 0.780; external cohorts A, B, and C: 0.779, 0.781, and 0.790, respectively), outperforming the trimodal LF model (internal validation: 0.753; external cohorts A, B, and C: 0.729, 0.715, and 0.730, respectively). A FM-empowered multimodal model integrating clinical, MRI, and pathological data was constructed to predict platinum resistance in HGSOC, demonstrating superior and stable performance across centers.