Identifying The Subtypes Of Primary Liver Tumors Based On Deep Learning

Liver cancer is one of the most common malignant tumors in clinic,with high incidence rate and high mortality.Early diagnosis and treatment of liver cancer can effectively improve the survival rate of patients,which has very important clinical significance.At present,imaging examination based on enhanced CT is one of the routine detection methods in clinical diagnosis of subtypes of liver tumors.However,imaging examinations usually require experienced radiologists to diagnose the subtypes of liver tumors,and the large amount of liver tumor cases has been contradicted by the few professional radiologists.Moreover,due to the different practical experience of different radiologists,there are observer differences in the diagnosis results of the same liver tumor image.In recent years,with the rapid development of computer technology,researchers tend to solve these medical problems by establishing computer-aided detection system.In this thesis,a computer aided detection system for diagnosing subtypes of liver tumors was established based on deep learning technology.This system can not only rapidly process a large number of medical image data in a short time,improve the work efficiency of radiologists,but also has a high prediction accuracy without observer differences,which has important clinical significance for preoperative diagnosis of liver tumors.The specific research content of this thesis is as follows:（1）An improved ConvNeXt model,Sim-ConvNeXt,is proposed based on a simple,parameter-free attention module for convolutional neural networks（SimAM）.In order to solve the problem that feature information of different channels in the same spatial position cannot be extracted effectively in ConvNeXt,we introduced SimAM attention module into ConvNeXt.SimAM is a parameterless 3D attention module,which can explore the importance of each neuron and strengthen the connection between space and channels.In addition,in order to accelerate the convergence of the model,this thesis uses a cosine warmup learning rate update strategy to optimize the learning rate.This experiment uses accuracy,precision,recall,₁value,and area under receiver operating characteristic curve（AUC）to evaluate the performance of the prediction model.From the experimental results,we can draw two conclusions:（i）Sim-ConvNeXt is superior to the other four typical deep learning models in the evaluation index of preoperative CT slice image data set of liver tumor patients,（ii）the use of Cosine Warmup can make the model converge faste.（2）Based on Sim-ConvNeXt,a neural network model with depthwise atrous spatial pyramid pooling（Dw ASPP）,Sim-Dw AConvNeXt,is proposed,which enables the prediction model to have the ability to fuse multi-scale features.First of all,we introduce the structure of the atrous spatial pyramid pooling（ASPP）in Sim-ConvNeXt,which strengthens the ability of the prediction model to extract multi-scale features.Finally,in order to balance the relationship between the accuracy of the prediction model and FLOPs,we set the expansion rates of the four dilated convolution in the ASPP module to 1,2,3,and 4 respectively.At the same time,depthwise convolution operation is used in the traditional dilated convolution layer in the ASPP module to make it become deeply detachable dilated convolution.The improved ASPP module is called the Dw ASPP module.In this experiment,accuracy,₁value,AUC value,parameter number of the model and floating point operation per second（FLOPs）are used to evaluate the performance of the prediction model.The results of ablation experiments on the preoperative CT section image data set of liver tumor patients show that the improvement of Sim-Dw AConvNeXt algorithm is effective,and Sim-Dw AConvNeXt can further improve the accuracy of predicting liver tumor subtypes.