DINOSAM+: Structure-Guided Prompt Learning for Semi-Supervised 3D Medical Image Segmentation.

Abstract

Semi-supervised 3D medical image segmentation under low-label settings often suffers from noisy pseudo labels and anatomically inconsistent predictions, which limit segmentation reliability. Existing methods primarily emphasize training strategies or loss engineering, but they still fail to sufficiently exploit structural priors to guide pseudo-label generation and unlabeled supervision. Hence, we propose DINOSAM+, a structure-guided semi-supervised framework for reliable 3D medical image segmentation. Specifically, a frozen self-supervised representation model, DINOv2, is employed to extract coarse structural attention from volumetric slices, which is further refined by a lightweight structure-aware network into foreground, boundary, and background probability maps, thereby providing more discriminative structural cues for pseudo-label generation. Based on these refined structural maps, structure-aware prompts are automatically generated for SAM-Med3D to produce pseudo labels with improved spatial consistency and structural completeness. The resulting pseudo labels are then incorporated into a Mean Teacher framework with confidence filtering, improving the reliability of unlabeled supervision and reducing error accumulation during training. Experiments on the LA and Synapse datasets show improved Dice, HD95, and worst-case performance under limited annotations.