Enhancing Pseudo-Label Reliability for Semi-Supervised Gastric Polyp Segmentation

Accurate gastric polyp segmentation is essential for effective diagnosis and prevention of gastric cancer, given the malignant potential of gastric polyps such as adenomas and hyperplastic lesions. However, gastric polyps present unique segmentation challenges due to multifocal and flat lesions as well as the complex gastric mucosal folds. To this end, we propose a novel semi-supervised gastric polyp segmentation framework. Conventional confidence-based pseudo-label refinement often suffers from overconfidence and unreliability, further exacerbated by the severe class imbalance between polyp regions and background. These challenges aggravate the impact of noisy pseudo-labels, motivating our reliability-aware pseudo-label refinement module. The module combines a similarity-confidence alignment score with the confidence score to yield a unified and recalibrated reliability measure. Based on this measure, pseudo-labels are adaptively filtered using a dynamic threshold, and small false-positive regions are subsequently removed. We further introduce a progressive augmentation-based consistency learning strategy that enforces prediction consistency across gradually intensified augmentations, enabling the model to achieve knowledge alignment and learn more robust representations under challenging endoscopic conditions. In addition, we curate a new benchmark dataset, named GPSD5K, consisting of one in-domain subset and three out-of-domain test subsets with diverse endoscopic systems. Extensive experiments show that our method outperforms eleven state-of-the-art approaches on the in-domain subset and two colorectal polyp datasets, and maintains robust performance on out-of-domain validation.