MAPSeg: self-supervised colorectal polyp segmentation via memory-augmented framework and synthetic polyp simulation

IntroductionColorectal cancer originates in most cases from polyps that progressively become malignant over time and colonoscopy offers a highly studied early diagnostic strategy to ensure a timely treatment plan. Automatic image segmentation of polyps, using intelligent supervised approaches, achieved good performance, but the need of their large annotated datasets limits the clinical applicability.MethodsThis study presents MAPSeg (Memory-Augmented Polyp Segmentation), a fully self-supervised and annotation-free framework for colorectal polyp segmentation, trained exclusively on images of healthy mucosa within an anomaly detection paradigm. The key novelty of MAPSeg mainly lies in SIMPO (Simulation of Polyps), a synthetic augmentation strategy that generates realistic polyp shapes and textures in a colon-specific context, combined with a memory-augmented encoder that models structural priors of normal tissue.ResultsExtensive experiments demonstrate that MAPSeg outperforms the strongest unsupervised methods by approximately 23% in Intersection over Union and 12% in DICE score on the Hyper-Kvasir dataset and consistently maintains this performance margin across multiple out-of-distribution benchmarks, indicating strong generalization capability.DiscussionThe results highlight MAPSeg supported by SIMPO as a viable solution for unsupervised colorectal polyp segmentation, significantly reducing the dependency on manual annotations while maintaining high segmentation accuracy.