Semi-Supervised Medical Image Segmentation
Overview
Semi-supervised medical image segmentation studies how to train models to outline organs, lesions, and tissues using a small set of annotated scans alongside many unannotated ones. Common approaches include teacher–student networks that generate pseudo-labels, consistency losses across augmented views, and feature or prototype alignment; newer methods also use foundation models, prompts, and diffusion-based refinement. Researchers evaluate these methods across settings such as cardiac MRI, abdominal CT, ultrasound, OCT, and histopathology, with benchmarks including ACDC, Synapse, AMOS, and GlaS. A central challenge is making pseudo-labels reliable for difficult or rare structures and across institutions, while ensuring gains under scarce labels generalize beyond benchmark-specific evaluation.
Latest papers
49 papers1 paper in the last four weeks, down 83% on the four weeks before. 0.0% of all new papers.