U-Net rectal cancer image segmentation based on depth separable convolution

Abstract

Aiming at the problem that manual diagnosis of rectal cancer has a large workload and is likely to cause subjective misdiagnosis, this paper proposes an improved U-Net rectal tumor segmentation method based on the U-Net network and combined with deep separable convolution: DS-UNet (Depthwise Separable Convolution U-Net). First, replace the convolution operation in the extraction part of the network with a deep separable convolution to deepen the depth of the network so that it can fully learn the image details, thereby improving the feature extraction ability of the model; finally, use the DiceLoss loss function to replace the cross entropy loss function, Optimize network parameters to solve the problem of sample imbalance, and then obtain higher sensitivity and generalization ability. The experimental results show that the accuracy of rectal cancer tumor segmentation of the proposed method reaches 91.20%, which is 3.94%, 24.31%, and 5.45% higher than that of U-Net, SegNet and DeepLab segmentation models, respectively, which verifies the effectiveness of the proposed method.