Residual UNet with spatial and channel attention for automatic magnetic resonance image segmentation of rectal cancer

The precise segmentation of rectal tumors is a key step in the diagnosis and treatment of rectal cancer. This paper aims to study the automatic segmentation task of rectal tumors based on deep learning methods, and proposes a residual UNet network model that combines spatial attention and channel attention. The model uses residual convolution for feature extraction, and uses squeeze-and-excitation module and attention gating module to focus on more useful features. In this study, we established a rectal tumor dataset for model evaluation, and used a combination of two-class cross-entropy and DICE loss function in the training process. Comparative experiments show that the DICE similarity coefficient is 0.8476, the Hausdorff distance reaches 9.5622, the prediction accuracy of the model is 0.9938, and the evaluation indicators are better than the segmentation results of UNet and AttUNet, which can effectively segment the rectal tumor area, and the combined loss function can also improve the segmentation accuracy by about 15% to a certain extent.

Automated summary

This paper proposes a novel automatic segmentation method for rectal tumors based on deep learning methods. The method utilizes a residual UNet network model that combines spatial attention and channel attention. Experimental results show that this method can effectively segment the rectal tumor area and achieve better evaluation indicators compared to traditional methods.