Description:
Background
Ovarian cancer causes approximately 14,000 deaths annually in the U.S., with many patients receiving standard chemotherapy regimens that prove ineffective. Deep learning offers promise in predicting treatment response by uncovering patient-specific disease factors. However, challenges such as limited model transparency and reliance on large datasets underscore the need for more scalable and interpretable approaches to guide treatment decisions.
Innovation
USC researchers have developed a multimodal deep learning framework utilizing H&E whole slide images (WSIs) and clinical data to predict ovarian cancer treatment effectiveness. Unannotated WSIs and a set of 22 clinical features were input into a multi-instance learning classifier. Interpretability models were also employed to provide insights into relationships between WSI areas and clinical variables with the classifier predictions. After demonstrating performance in ovarian cancer, feature-level transfer learning was applied to predict recurrence in a renal cell carcinoma dataset. This framework offers an interpretable and adaptable approach for precision oncology.
Advantages
- Integrates multimodal data to increase prediction accuracy and AUC vs unimodal analysis
- Enhances interpretability by detailing feature interactions and identifying informative tumor regions within WSIs
- Transferable framework is data-efficient and adaptable to different solid tumor settings
Publications
https://pmc.ncbi.nlm.nih.gov/articles/PMC10785847/