Research On Liver And Liver Tumour Segmentation Algorithm Based On MobileViT-Unet

With the development of computer imaging technology and deep learning technology today,the data volume of CT images is constantly increasing,the workload of imaging experts is increasing,and their analysis has many drawbacks.Therefore,designing an automatic CT segmentation scheme for liver and liver tumours will improve the diagnostic efficiency of liver cancer and be beneficial for patient treatment.The current segmentation schemes mainly focus on the difficult problems of small target missed detection,strong individual differences,and blurred target boundary segmentation in liver and liver tumour segmentation.This article also proposes a liver and liver tumour segmentation scheme based on Mobile Vi T-UNet based on the above issues.1.This article proposes a liver and liver tumour segmentation network(Mobile Vi T UNet)scheme based on Transformer’s self-attention lightweight module(Mobile Vi T block).This method uses the encoder and decoder structure of the "U" network architecture and introduces the Mobile Vi T block lightweight self-attention module to make up for the lack of a global Receptive field in convolution,improve the network’s ability to calculate the semantic correlation of the global pixels of the feature map,and thus improve the performance of model segmentation.2.Considering the complexity and high computational complexity of the Mobile Vi T-UNet model,it is difficult to train the model.This article improves the structure of the Mobile Vi T-UNet model by introducing the Mobile Vi T-block module only in the encoder part of the algorithm model to reduce the complexity and computational complexity of the model;(2)Introducing a multi-scale information fusion module at the skip connections of the model to increase the information redundancy and enhance the robustness of the model,alleviating the impact of pixel semantic gaps caused by skip connections;(3)This article proposes a novel attention gate mechanism for SA-CBAM,which integrates multi-scale information with decoder information to improve the extraction of target liver and liver tumours in the model graph.3.To explore the three-dimensional liver and liver tumour automatic segmentation scheme of Mobile Vi T-UNet,the improved model was designed in a three-dimensional manner.To solve the problem of difficult segmentation of 3D segmented objects,this paper designs the loss function as Focal Loss plus Tversky Loss to improve the segmentation performance of the model for individuals with difficult segmenting small objects.This article conducted ablation and comparative experiments,and the experimental results showed that the 2D segmentation algorithm based on Mobile Vi T-UNet had a global average Dice similarity coefficient of 92.23% for liver tumours and 87.75% for liver tumours.In addition,the two scoring indicators of3D-Mobile Vi T-UNet also reached 88.42% and 74.54%.