Abstract: A method for the benign and malignant classification of lung tumors was put forward due to Challenges with the training of deep convolutional neural networks, network degradation and a weak ability to express the features based on non-negative representation classification and a multi-modal residual neural network. The pre-trained residual neural network model was initialized using transfer learning. three data sets(CT, PET and PET/CT)were used to train the network and extract the feature vectors of the fully connected layer, then a non-negative representation classifier was used for the non-negative representation of the feature vector, and used to solve the non-negative coefficient matrix. The residual similarity was used to classify benign and malignant lung tumors. Comparative experiments were conducted with the AlexNet, GoogleNet and ResNet-18/50/101 models. The experimental results showed that the classification accuracy of the ResNet-NRC was better than the other models, and the specificity and sensitivity indices were also higher. The proposed method has improved robustness and generalization ability.

Key words: residual neural network, multimodal medical image, lung tumor, transfer learning, non-negative representation classification algorithm