MBDA-Net: Multi-source boundary-aware prototype alignment domain adaptation for polyp segmentation

Accurate segmentation of polyps in colonoscopy images is important for the prevention and treatment of colorectal cancer. However, samples collected from different centers often possess diverse distributions, leading to poor generalization when a segmentation model trained in one center is directly employed in another. This paper proposes a multi-source boundary-aware prototype alignment domain adaptation network (MBDA-Net) to improve the performance of cross-center colonoscopy image segmentation. Specifically, we first design an image translation (IT) module based on the discrete cosine transform (DCT) to reduce the distribution gap between source and target domains by translating source domain styles into target domain styles. Then we propose a mutual perception prototype alignment (MPPA) module containing prototype inference, prototype interaction and adaptive mutual feature fusion. By learning relationships between prototypes and features among multiple domains, one can obtain mutual perception features that fuse prototype information from multiple domains. In order to fully exploit the supervised information of the source domains and optimize the prediction boundaries, we develop a boundary-aware learning (BAL) module to align the boundaries of the source domain predictions and ground truths. Moreover, to mitigate the foreground–background imbalance present in small-sized polyp images and reduce the biased predictions of the model, this study proposes a double normalization strategy (DNS) during the inference stage to improve the detection rate of small polyps. Experimental results on three challenging public datasets show that the proposed MBDA-Net outperforms existing methods on cross-center colonoscopy image segmentation, achieving state-of-the-art performance.