CT image liver segmentation based on improved Deeplabv3+

In modern society, the incidence of hepatitis is very high, and accurate segmentation of the liver region can assist physicians in the diagnosis of liver disease. However, mainstream CT image segmentation algorithms U-net, U-net++ did not make full use of the high-level and the low-level features. Still, it simply fused them, which leads to the problem of under-segmentation or over-segmentation of some small regions of the liver by these algorithms. As Deeplabv3+ splicing the features extracted from the first and last layers of the backbone network are processed through different modules. The first layer extracted features are processed by convolution as well as the last layer extracted features are processed by expanding the sensory field without increasing the parameters using dilated convolution in the ASPP module as well as global average pooling for feature enhancement, respectively. In this way, high-level and low-level features are fully utilized, and thus rich contextual semantic and multiscale information is obtained. Aiming at the above problem, a segmentation of CT image based on improved Deeplabv3+ was proposed. First of all, the Deeplabv3+ backbone network was replaced by ResNet-50 to address the phenomenon of gradient explosion and gradient disappearance. Secondly, each residual module in ResNet-50 was added to the SE channel attention module, thus enabling the network to pay attention to more significant information channels. Finally, transfer learning was implemented on the improved ResNet-50 network to make the network converge faster and reduce the risk of overfitting under small sample datasets. In this paper, it performed experiments in the 3Dircadb dataset. The outcome of the experiment indicated that the proposed method outperforms U-net and U-net++ in subjective perception as well as objective evaluation metrics, and the mIoU and Dice coefficients reached 95.40% and 94.80%, respectively.