Universal medical image segmentation via in-context cross-attention.

Semantic segmentation is critical in medical image processing, with traditional specialist models facing adaptation challenges to new tasks or distribution shifts. While both generalist pre-trained models and universal segmentation approaches have emerged as solutions, universal methods offer advantages in versatility, sample efficiency, and integration ease into annotation pipelines. We introduce a novel universal segmentation method based on the premise that pre-selecting relevant regions from support sets improves segmentation accuracy. Our approach implements cross-attention between query images and support set images, coupled with an innovative attention up-scaling mechanism that efficiently computes cross-attention on small-scale features with upscaling to higher resolutions. The design inherently supports explainability by allowing inspection of relevant support set locations for each input region. Extensive evaluation across 29 medical datasets spanning 9 imaging modalities and 135 segmentation tasks demonstrates consistent performance improvements, even with lightweight models. Our experiments show proportional gains in segmentation performance as support set size increases, with the cross-attention mechanism effectively selecting the most relevant support images from larger annotation pools. Additionally, our explainability module demonstrates competitive or improved interpretability when compared to established methods like LayerCAM.