Semi-supervised Polyp Segmentation via a Dual-Teacher Student Framework

Automatic polyp segmentation is very critical for the early detection and treatment of colon cancer. Most existing studies in polyp segmentation are devoted to using a fully supervised training, which requires a substantial amount of pixel-wise annotated data. In this paper, we propose a dual-teacher student framework for semi-supervised polyp segmentation. Our framework, named DTS-PolypNet, introduces a dual-teacher mechanism. To optimize the model training process, DTS-PolypNet proposes an innovative data augmentation technique called Attention-Enhanced Mixer (AEMixer). AEMixer effectively combines outputs from two teacher models to create high-quality mixed images and pseudo-labels. To assess the performance of our proposed method, extensive experiments on the Kvasir-SEG and CVC-ClinicDB datasets show that our DTS-PolypNet achieves SOTA performance.