Puzzles and Segmentation

Abstract:

The clinical reality of missing data severely hampers multimodal deep learning, as strong modalities often suppress the learning from weaker ones. This talk presents PRISM, a module that stabilizes training for tasks like MRI-based segmentation, even in the face of this data imbalance. PRISM introduces an internal teacher-student dynamic with two key components. The first is a dual-distillation scheme that transfers knowledge at both the pixel and feature-prototype levels for a more comprehensive understanding. The second is a relative preference mechanism that acts as an intelligent regulator, actively balancing the learning process. It identifies and accelerates underperforming modalities while modulating gradients across epochs. The result is a robust segmentation model, resilient to the unequal modality absences found in real-world clinical datasets.

About David:

David Rutkevich is a computer science student at TU Berlin whose software development work focuses on data efficient models and has earned him national recognition. He was the national winner of the 2025 “Jugend forscht” competition for developing a method to improve medical image segmentation, particularly when dealing with incomplete data sources. He also secured first place in the German Artificial Intelligence Competition for creating an AI model that fully automates the analysis of blood cells.