A boundary-enhanced and target-driven deformable convolutional network for abdominal multi-organ segmentation

It is crucial to accurately segment organs from abdominal CT images for clinical diagnosis, treatment planning, and surgical guidance, which remains an extremely challenging task due to low contrast between organs and surrounding tissues and the difference of organ size and shape. Previous works mainly focused on complex network architectures or task-specific modules but frequently failed to learn irregular boundaries and did not consider that different slices from the same case might contain targets of different numbers of categories. To tackle these issues, this paper proposes UAMSNet for abdominal multi-organ segmentation. In UAMSNet, a hybrid receptive field extraction (HRFE) module is introduced to adaptively learn the features of irregular targets, which has an adaptive dilation factor containing distance information to facilitate spatial and channel attention. The HRFE module can simultaneously learn multiple scales and deformations of different organs. Furthermore, a multi-organ boundary-enhanced attention (MBA) module in the encoder and decoder is designed to provide effective boundary information for feature extraction based on the large peak of the organ edge. Finally, the difference in the number of organ categories between different slices is first considered using a loss function, which can adjust the loss computation based on organ categories in the image. The loss function mitigates the effect of false positives during training to ensure the model can adapt to small organ segmentation. Experimental results on WORD and Synapse datasets demonstrate that our UAMSNet outperforms the existing state-of-the-art methods. Ablation experiments confirm the effectiveness of our designed modules and loss function. Our code is publicly available on https://github.com/HeyJGJu/UAMSNet.