Transfer Learning and Attention Mechanism for Breast Cancer Classification

Abstract

Since the existing deep learning frameworks still cannot meet the high requirements of accuracy and efficiency in practical clinical diagnosis, a breast cancer classification model based on deep transfer learning and visual attention mechanism is proposed. Firstly, in order to overcome the overfitting effect brought by a small number of samples of breast histopathological images, transfer learning is carried out on the basis of two network frameworks, namely VGG16 and ResNet50, and the ImageNet dataset is used as the transfer source domain and the shallow convolutions are frozen. By doing this, training efficiency can be effectively improved since there is no need to train shallow general features repeatedly. Secondly, the attention modules are combined to focus on the feature description of breast lesions. The classification accuracy and efficiency of the model can be effectively improved since they vastly reduce the redundant and interference information. Finally, the soft voting is applied to fuse the class probability outputs from the two individual classifiers to obtain the final classification result. The experimental results of the BreaKHis dataset show that the two-classification accuracy of the proposed network reaches 99.45%, and the eight-classification accuracy reaches 94.11%. Compared with some recent breast image classification algorithms, such as BiCNN, CSDCNN, BHCNet, the proposed method is better in terms of many indicators, such as Accuracy, Precision, Recall, and F1_ score.