Analysis Of Microvascular Infiltration Of Hepatocellular Carcinoma Based On ResNet Depth Feature

Hepatocellular carcinoma is the most common form of liver cancer,accounting for 90% of primary liver cancer.Hepatocellular carcinoma is the top three most common malignancy in the world.Microvascular infiltration MVI as an indicator of tumor invasion has a good effect on liver cancer and its postoperative diagnosis.The diagnosis of hepatocellular carcinoma by MRI is extremely laborious and requires great expertise from doctors.Extracting the texture features of hepatocellular carcinoma images and training the classification model through texture features can provide reliable recommendations for doctors' diagnosis.The difficulty in microvascular infiltration analysis of hepatocellular carcinoma is that it is difficult to obtain effective features by artificial design.For this reason,a deep learning feature such as ResNet is needed to obtain a large number of depth features,and a large number of depth features and a small number of training samples easily lead to over-fitting of the classification model.In order to meet the needs of Shanghai Lian Ying's actual projects,it is necessary to develop an effective model and system for extracting image features,selecting features,and accurately classifying them.To this end,the feature selection experiment was carried out by the correlation coefficient method combined with SVM model,PCA combined with SVM model and LDA combined with SVM model.Experiments show that the Filter-based feature selection method can not get the expected classification accuracy.Finally,using the Recursive Feature Elimination(RFE-SVM)algorithm based on Support Vector Machine,the Wrapper-based feature selection method is used to obtain the desired classification accuracy.On this basis,a software system capable of diagnosing the presence or absence of "microvascular infiltration" in hepatocellular carcinoma using a small sample of hepatocellular carcinoma images was developed.The main algorithms include:(1)extracting liver texture features of hepatocellular carcinoma images,including traditional image texture features and ResNet depth features;(2)designing recursive feature elimination algorithm RFE filtering uncorrelated features;(3)establishing support vector machine model pairs There is no "microvascular infiltration" for diagnosis.Through the 5-fold cross-validation experiment,it was found that RFE combined with SVM classifier obtained satisfactory prediction performance,and the accuracy rate of "microvascular infiltration" was 81%.The study showed that the proposed method has good applicability in the extracted liver texture features in this project,and the classification of “microvascular infiltration” is superior to the previous system of the joint medical company.