Multimodal, Multispecies and Pathology Invariant Skullstripping

Résumé fourni par la source

Motivation: Recent skullstripping models aim towards better generalizability, but still suffer from innate assumptions from the training data. Goal(s): Our goal is to minimize assumptions and priors, so that a single model would be generalizable between modalities, pathology, and species. Approach: We train a skullstripping model on synthetic data that uses two assumptions: 1. The brain would be inside the skull/head 2. The brain is the center main object in the image. Post processing step of removing false positives is done after inference. Results: Our model succeeds in removing the skull various multimodal, multispecies and pathology tasks. Impact: This is the first demonstration of a multimodal, multispecies and pathology invariant skullstripping model, only trained on synthetic data with minimal assumptions. Results suggest that with correct assumptions, a single model could be all we need for any skullstripping task.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.