Weakly Supervised In-context Learning-powered High-efficiency Universal Models for Medical Image Segmentation
Year: 2025
Type: preprint
Abstract: In medical image segmentation, universal models-including interactive and in-context learning (ICL) variants-deliver robust generalization capabilities but suffer from heavy reliance on extensive annotations. Interactive models depend on repeated user prompts for each individual image, while ICL models require dense, pixel-level labels to construct context sets. To address these inefficiencies, we propose Weakly Supervised In-Context Learning (WS-ICL), a novel ICL paradigm that uses weak prompts (e.g., bounding boxes or points) instead of dense masks to build context. This approach drastically cuts annotation effort by removing the need for fine-grained segmentation masks and avoiding repeated prompting for every image. We evaluated the proposed WS-ICL model on three held-out benchmark datasets. Experimental findings show that WS-ICL matches the performance of conventional ICL models while slashing annotation costs considerably. Additionally, WS-ICL maintains strong competitiveness even when operating under the interactive paradigm. These results position WS-ICL as a valuable stride toward developing more efficient and unified universal models for medical image segmentation.
Author: Naranhvwar Tvlg
Institution: Mongolian University of Science and Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4415122871
Cites:
Cited by:
Related to: