SegMoTE: Token-Level Mixture of Experts for Medical Image Segmentation
Researchers have developed SegMoTE, a new framework designed to adapt general image segmentation models like SAM for medical imaging tasks. This approach introduces a small number of learnable parameters to dynamically adjust for different modalities and anatomies, overcoming limitations of previous fine-tuning methods. SegMoTE also features a progressive prompt tokenization mechanism for fully automatic segmentation with significantly reduced annotation needs, achieving state-of-the-art results on diverse medical datasets with minimal training data. AI
IMPACT Enables more efficient and cost-effective deployment of advanced segmentation models in clinical settings.