The persistent clinical problem of positive tumor margins in cancer surgery presents a computational grand challenge: to process up to 2 TB of microscopic imaging data within 10 minutes in the operating room, providing the surgeon with actionable information in real time.
We address this challenge with FASMP, a human-centered high-performance computing cyberinfrastructure designed to train ML models on datasets of hundreds of terrabytes of medical imaging data. We address the extreme data requirements with a multiscale, cache-oblivious data model that mitigates I/O bottlenecks, a transformer architecture that drastically reduces trainable parameters, and the integration of Topological Data Analysis to ensure robustness.
User-optimized, task-specific labeling tools enable the creation of effective annotations quickly, improving training and closing the validation loop with medical professionals.
Ultimately, FASTMAP serves as the computational engine for an imaging platform, enabling real-time, automated cancer detection in the operating room. This work fundamentally transforms surgical outcomes for cancer patients by providing immediate, data-driven feedback during the procedure.