A Preoperative Prediction Method For Differentiation Grades Of Hepatocellular Carcinoma Based On Multi-feature Fusion

Hepatocellular carcinoma(HCC)accounts for about 90% of primary liver cancer,is prone to metastasis and recurrence,and is the second leading cause of cancer-related death.Recurrence of HCC is closely related to its differentiation grade,and preoperative determination of differentiation grade is essential for the selection of treatment strategies.Currently,pathological sections are the gold standard for grading and are invasive diagnostic means,which may lead to missed diagnosis due to inaccurate sections.Magnetic resonance examination is a conventional imaging method for screening and diagnosis of HCC patients,which has the characteristics of noninvasive,rapid and economical,but its diagnosis is based on the experience of doctors to give diagnostic opinions,which is prone to missed diagnosis and misdiagnosis.Using magnetic resonance imaging and computer technology to achieve non-invasive and rapid auxiliary diagnosis of HCC differentiation grade can reduce the pain and risk of patients caused by pathological sections and reduce medical costs.To solve the problem of accurate grading of HCC,this paper proposes a classification method based on radiomics.The steps of this method include: data acquisition,image preprocessing,tumor segmentation,feature extraction,feature selection and classification model training.In this paper,we collected 189 T2 weighted MRI images of patients with HCC from the hospital.In order to alleviate the image brightness difference caused by different acquisition equipment,we used local adaptive histogram equalization algorithm to preprocess the image.The imaging doctors divided the tumor,extracted four types of radiomics characteristics for the tumor interest area drawn.In order to obtain the characteristics of stable and redundant removal,Finally,SVM,random forest,XGBoost and light GBM are used to train the classification model.The AUC value of the model is 0.773 and the accuracy is 0.776.It is proved that the image features extracted by this method can express tumor information and the classification model has great grading performance.Because the radiomics features are fixed quantitative features,it is difficult to fully reflect the heterogeneity of tumor.The deep convolution neural network features have the characteristics of automatic extraction and strong expression ability,which can supplement the shortcomings of image features.This paper proposes a feature extraction method based on depth learning,which includes data augmentation,depth model training,depth feature extraction and feature fusion.The depth feature which can express the tumor information is extracted by model migration and parameter migration,and the multi-modal feature fusion training is carried out by using the logistic regression model.The AUC value of the classification model is 0.828,the accuracy is 0.796,the sensitivity is 0.708,and the specificity is 0.880.The AUC value of the classification model using this method is 5.5% higher than that of the traditional radiomics method,which shows that the fusion of image features and depth features can improve the classification performance of the model.Calibration curve evaluation shows that the classification stability of the model is good,which can provide valuable information for clinical decision-making.