Reproduce “PROFUSEme: PROstate Cancer Biochemical Recurrence Prediction via FUSEd Multi-modal Embeddings” on OpenResearch

About this paper

Suhang You, Carla Pitarch-Abaigar, Sanket Kachole, Sumedh Sonawane, Juhyung Ha, Anish Sudarshan Gada, David Crandall, Rakesh Shiradkar and 1 more

The PROFUSEme framework, developed by Indiana University, predicts prostate cancer biochemical recurrence by integrating clinical, radiology, and pathology data through a Transformer-based intermediate fusion strategy. This approach yielded a C-index of 0.861 in internal cross-validation, outperforming late fusion methods, and achieved a 0.7107 C-index on an external challenge leaderboard.

Key insights

• Learning explicit cross-modal interactions through an intermediate fusion approach significantly improves prognostic accuracy compared to late fusion methods.
• Integrating diverse patient data—clinical, radiology, and pathology—provides a more comprehensive and robust assessment for prostate cancer biochemical recurrence.
• Pre-trained foundation models for medical imaging are effective for extracting robust features, allowing focus on multi-modal fusion mechanisms rather than raw pixel processing.

Summary by alphaXiv.