Abstract: To achieve accurate classification of liver tumor lesions, a liver tumor lesion classification method was proposed based on joint feature learning and multiple transfer learning. The data augmentation was completed on the image of input network through the preprocessing method of expansion channel, and the network could extract more feature information from the original input image. A joint feature learning dual-stream convolutional neural network was designed to extract features to avoid the loss of some feature information due to the increasing in network depth. An ensemble classifier was adopted to achieve the final classification, and the entire ensemble classifier was optimized through the multiple loss constraints. The parameter transfer and domain adaptation were combined in the training process of the model to reduce loss and improve the fitting performance of the model. In the experiment, 155 plain CT images of the abdomen were used, and several evaluation indices of specificity, sensitivity, precision, F1-score, accuracy and error rate were designed. The results show that the proposed method can realize the classification of hepatocellular carcinoma（HCC）, metastatic liver cancer（MET）, hemangioma（HEM） and normal liver tissue, and the average classification accuracy rate reaches 96%.