Semi-Supervised Segmentation Method for Craniopharyngioma Based on SAM Annotations

Craniopharyngioma is a rare epithelial tumor primarily located in the sellar/suprasellar region, adjacent to critical structures such as the hypothalamus and pituitary stalk. Its complex growth patterns pose significant threats to patients’ neurological functions and endocrine systems. Accurate image segmentation is crucial for preoperative planning and postoperative monitoring. However, the low incidence of craniopharyngioma results in limited annotated data across medical institutions, and the tumor’s high morphological heterogeneity and ambiguous boundaries make reliable interpretation of craniopharyngioma images time-consuming and labor-intensive, requiring experienced radiologists. To address these challenges, this paper proposes a semi-supervised learning framework based on the Segment Anything Model (SAM). By integrating confidence-weighted training and anatomically constrained post-processing, the framework achieves high-precision segmentation with limited annotations. Leveraging SAM’s generalization capability to generate pseudo-labels, the framework employs a confidence-weighted strategy to suppress misleading supervision in noisy regions. Additionally, an anatomy-constrained mechanism based on spatial priors of the sellar region ensures that segmentation results align with clinical knowledge. This approach significantly reduces annotation costs while maintaining anatomical plausibility and reliability, providing an efficient tool for the precise diagnosis and treatment planning of craniopharyngioma.Our code is available at https://github.com/kongdeqing2025/CP-SemiSeg.