Boundary-Selective Pseudo-Label Repair and Model Absorption for Semi-Supervised Medical Image Segmentation

Abstract

Pseudo-labels can exploit unlabeled medical images, but high-overlap masks may still contain local contour errors and selected corrections may disappear during fine-tuning. We recast pseudo-label refinement as a repair-to-model problem and present Boundary-Selective Pseudo-Label Repair and Model Absorption (CPPA). A validation-selected frozen source supplies the no-edit reference. Wavelet-frequency evidence, spatial responses, and disagreement between source predictors constrain a morphological Snake to bounded add/remove candidates. A local policy scores each contour arc against the paired no-edit source and abstains when the estimated benefit is unsafe. The selected mask supervises an edit-aware absorption bank, after which a validation-only router promotes one checkpoint or retains the source. Across 27 internal settings spanning CVC-ClinicDB, BUSI-benign, and BUSI-malignant, three label ratios, and three seeds, the final single-model outputs improve the selected source by 0.0241 Dice, 0.0536 BF1, and 0.0402 Boundary IoU on average. Adapted checkpoints are promoted in 26 of 27 settings, and all nine dataset-ratio means improve Dice, BF1, and Boundary IoU. Frozen transfer to CVC-ColonDB improves Dice and BF1 at every label ratio. The evidence supports a bounded, abstaining repair policy whose local corrections can be absorbed into one deployable network.