Content and shape attention network for bladder wall and cancer segmentation in MRIs.

Accurate segmentation of the bladder wall and cancer is the key to preoperatively predicting patients' muscle-invasive status. However, the segmentation of bladder wall and cancer have many challenges, including complex background distribution, a variety of bladder shapes, and weak boundary. For these issues, we propose a deep network that consists of a content attention module and a shape attention module. In the content attention module, we employ the attention U-Net to emphasize salient image features that are useful for the segmentation task. The shape attention module uses a spatial transform network to introduce a shape prior, which ensures a closed bladder wall in segmentation results. Experimental results show that the proposed model has a competitive performance compared to the existing methods. The mean DSCs of the 5-fold cross-validation was 0.80 and 0.84 for bladder wall and cancer respectively. From the visualization, our approach can mitigate the issue of complex background and weak boundary in bladder wall and cancer segmentation effectively.