TransDeeplabv3: Multi-Prior Segmentor for Medical Image Segmentation

At present, medical image segmentation is a challenging and developing task in medical image field, in which semantic segmentation algorithms play an important role in many clinical applications. The atrous spatial pyramid pooling (ASPP) module of Deeplab-V3 can well balance the diversity, robustness, and connectivity of global feature map extraction under different scenarios. However, there is still a large improving prospect for medical image segmentation. Therefore, we introduce mechanisms of multi-group parallel processing, mass guided loss focusing on irregular shape, and k-nearest attention architecture, respectively, to achieve higher robustness and stronger global connectivity with more sufficient receptive field compared to original methods. The proposed model effectively alleviates the deficiencies of local semantic receptive fields, improves the ability of generalization of our model, and overcomes the agnosticism of feature spatial distribution. Under the premise of maintaining feature diversity and enhancing heterogeneous information correlation, we obtain better segmentation performance in our experiments by enriching global feature representations. We conduct a series of segmentation experiments using the proposed model and on medical image datasets: STARE, CHASE, DRIVE and HRF. Compared with Deeplab-V3 baseline, the performance is significantly improved in terms of several metrics, which verifies the effectiveness of our method.