Multimodal AI in high-grade serous ovarian cancer: integrated prediction and clinical decision-making.
Authors
Affiliations (6)
Affiliations (6)
- Precision Imaging as a New Key in Cancer Care (PINKCC) Lab, Montpellier Cancer Research Institute (IRCM), Univ Montpellier, Inserm, ICM, Montpellier, France.
- Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, NY, United States.
- Research Center for Medical Image Analysis and Artificial Intelligence, Department of Medicine, Faculty of Medicine and Dentistry, Danube Private University, Krems an der Donau, Austria.
- Montpellier Cancer Institute, Department of Surgery, Montpellier, France.
- Fudan University Shanghai Cancer Center, Department of Radiology, Shanghai, China.
- Montpellier Cancer Institute, Department of Radiology, Montpellier, France.
Abstract
Ovarian cancer remains the most lethal gynecologic malignancy, with high-grade serous ovarian carcinoma (HGSOC) accounting for the majority of deaths. It is typically diagnosed at an advanced stage, characterized by extensive peritoneal dissemination and high rates of recurrence, which contribute to persistently poor survival despite advances in cytoreductive surgery, platinum-based chemotherapy, and targeted therapies such as PARP inhibitors. Clinical outcomes are highly variable, reflecting substantial inter- and intratumoral heterogeneity at the genomic, cellular, and microenvironmental levels. Accurate survival prediction and treatment response assessment therefore remain major challenges. Current clinical and radiological evaluation tools provide only limited resolution of this complexity and fail to fully capture the spatial and temporal dynamics of the disease. As a result, patient stratification and treatment decision-making are often suboptimal, particularly in the context of evolving therapeutic strategies. Artificial intelligence (AI) has emerged as a powerful approach to extract clinically relevant information from diverse data sources, including CT and MRI radiomics, histopathology whole-slide images, and multi-omics data. While promising results have been reported within individual modalities, each captures only a partial view of the disease. Integrating these complementary data sources through multimodal AI approaches offers the potential to more comprehensively model tumor biology and improve predictive performance. In this review, we analyze 31 HGSOC studies using a unified multimodal pipeline taxonomy structured into five sequential stages: feature extraction, intra-modal aggregation, fusion strategy, inter-modal integration, and prediction head. For each stage, we identify and categorize the technical approaches employed across studies, highlighting both common design patterns and novel methodological contributions. To complement this analysis, we provide visual overviews, a paper-by-paper comparative visualization, and structured summary tables. We further discuss key challenges in the field, including inconsistent validation strategies, limited reproducibility, heterogeneous integration of genomic data, and insufficient attention to model explainability.