GP-CPS: Gradient Penalty Cross Pseudo Supervision for Semi-Supervised Medical Image Segmentation

Medical imaging has always been constrained due to the challenges in the acquisition process, poor signal-to-noise ratios, high costs, and the complexity of medical image features. This work introduces a semi supervised framework called Gradient Penalty Cross Pseudo Supervision (GP-CPS), which is based on the Cross Pseudo Supervision (CPS) model and innovatively introduces the concept of gradient penalty, sig-nificantly enhancing the model's performance. To our knowledge, this is the first time the concept of gradient penalty has been applied in the field of image segmentation. Furthermore, this work introduces a concept of fused cross pseudo super-vision to enhance the diversity of training and strengthen the robustness of the model. Using the publicly accessible Kvasir-SEG dataset, the proposed model is compared with baselines and advanced models. Across all four groups with varying amounts of unlabeled data, the suggested model consistently shows better performance. The source code for this work is publicly available at github.com/JustinPeKi/GP-CPS.