Research On Automatic Segmentation Of Liver Tumor Image Based On Deep Learning

Liver tumor CT segmentation is an important task in the field of medical image analysis,which aims to automatically extract the contours of the liver and intrahepatic tumors from CT images,providing doctors with more accurate diagnosis and treatment recommendations.There are three main challenges in liver tumor CT image segmentation.Firstly,the morphology,size,and location of the liver and intrahepatic tumors vary greatly,requiring segmentation algorithms to have strong robustness and adaptability.Secondly,the grayscale distribution of the liver and intrahepatic tumors in CT images is similar,requiring segmentation algorithms to have high discrimination and accuracy.Thirdly,there are often problems such as artifacts,noise,and non-uniformity in CT images,requiring segmentation algorithms to have strong robustness and denoising ability.This article mainly focuses on these pain points,with the following contents:To address the challenge of the variability of liver tumor morphology,this article proposes a liver tumor segmentation method based on a multi-scale residual atrous convolutional neural network.The multi-scale and atrous convolution can effectively extract multi-scale features,helping the model to discriminate liver tumors with significant differences in shape and size.The design of residual networks can not only prevent gradient explosion during training but also preserve more image details,which is helpful for identifying smaller tumors.Experimental results show that the proposed method based on multi-scale residual atrous convolutional neural network can achieve a segmentation performance(DICE index)of 65.34% on 3D-IRCADb1,which is better than the compared methods.To address the challenge of the similarity between the grayscale distribution of intrahepatic tumors and the background,this article proposes a method based on asymmetric atrous convolutional networks.Asymmetric atrous convolutional networks can solve the grid problem caused by atrous convolution,enabling more image details to be preserved and effectively improving the discriminative power of the model.Experimental results show that the average DICE score of the proposed method based on asymmetric atrous convolution in liver tumor segmentation is 76.16%,demonstrating the effectiveness of the model.To address the problems of artifacts,noise,and non-uniformity in CT images,this article proposes a liver tumor segmentation method based on the orthogonal fusion of local and global deep features.Effective modeling of local and global features can help the model learn lesion semantics more accurately and effectively distinguish lesions from artifacts and noise.Specifically,the compressed and excited network is combined with the latest convolutional neural network architecture Conv Ne Xt as the backbone network for effective feature extraction.Secondly,in the decoder,this method uses a local and global deep feature orthogonal fusion module to extract fine-grained contextual features.Experimental results show that the proposed method based on the orthogonal fusion of local and global deep features achieves a DICE index of 85.80% in liver tumor segmentation,demonstrating the effectiveness of the proposed method.