CT Images-Based 3D Convolutional Neural Network In Prediction Of Early Recurrence Of Hepatocellular Carcinoma After Radical Hepatectomy

PurposeFirst,establish the 3D convolutional neural network model to accurately predict the early recurrence(≤2 years)of solitary hepatocellular carcinoma after radical hepatectomy according to the preoperative dynamic enhanced CT images.Second,explore the hidden prediction logic of 3D convolutional neural network by drawing the gradient-weighted class activation mappings and the class saliency maps.MethodsChapter Ⅰ:according to the inclusion and exclusion criteria,a total of 220 patients pathologically proved hepatocellular carcinoma who received radical hepatectomy as the initial treatment were included in the study cohort.There were 178 patients in the training group(73 patients with early recurrence)and 42 patients in the verification group(21 patients with early recurrence).The basic structure of 3D convolutional neural network model was constructed by Keras package based on Python.After resampling,segmentation and extraction of volumes of Interest,dataset augmentation and gray scale normalization,the sample files were input into the model for training.The training process of the model was supervised by cross-entropy loss function to select the optimal model.Finally,the predictive ability and clinical practical value of the model were verified and evaluated.Chapter Ⅱ:reconstruct the 3D convolutional neural network model and import the weight of the optimal model into the reconstructed model to restore the optimal model.By applying the Keras-vis package based on Python,we use the preprocessed samples to draw the gradient-weighted class activation mappings and the class saliency maps,to analyze and explain the hidden prediction logic of 3D convolutional neural network model.ResultsChapter Ⅰ:In the training group,the values of area under the receiver operating characteristic curve(AUC)of early recurrence predicted by 3D convolution neural network model was 0.829(95%CI 0.770-0.889).In the validation group,the AUC of early recurrence predicted by 3D convolution neural network model was 0.787(95%CI 0.634-0.940).According to the Youden index of ROC in training group,the best cut-off value is 0.402,to divide patients into high-risk group and low-risk group.In the training group,100 patients were at a high risk of early recurrence and 78 patients were at a low risk.In the validation group,17 people with a high risk of early recurrence and 25 people with a low risk.The Kaplan-Meier survival curve of the training group and the validation group showed that there was a significant difference in recurrence-free survival between low-and high-risk groups(P<0.001).Chapter Ⅱ:the gradient-weighted class activation mappings show that the model pays more attention to the edge and surrounding area of the tumor.The class saliency maps show that the model can extract 3D spatial features well,and the pixels with high spatial support related to early recurrence are distributed in the surrounding area of tumor,high enhancement area,non enhancement area and the background of the tumor in the volume of interest.Conclusions3D convolutional neural network model has high predictive performance in predicting early recurrence of patients with solitary hepatocellular carcinoma after radical hepatectomy.Based on the analysis of the gradient weighted class activation map and class saliency map to visually explain the model,we speculate that the 3D convolutional neural network model has the ability to predict the prognosis of patients with hepatocellular carcinoma,which may be related to tumor size,specific texture of high enhancement area and specific texture of non enhancement area.However,these features are advanced and complex which are difficult to understand and quantify.