Abstract:

Breast tumors adversely affect the holistic well-being of women. Histopathological images are a critical substantiation for doctors to diagnose breast tumor types. The structure of various types of tumor cells exhibits significant correlations, thereby posing challenges to the diagnosis using conventional methods. In this work, the enhanced EfficientNet was employed for the diagnosis of breast tumors, which enabled the network model to learn the features of the disease automatically and improve the accuracy of the diagnosis of breast tumor types. Firstly, the convolutional block attention module was used to extract effective features. Secondly, the group convolution and channel shuffle operations were introduced to improve the feature representation ability of the model. Thirdly, the Hard-Swish activation function was applied to improve the convergence speed of the model. Finally, Experiments showed that the improved EfficientNet network achieved 98.4% accuracy in eight classifications on the BreakHis dataset, which was expected to act a decision aid tool in breast tumor diagnostic research.

Key words: breast tumor, EfficientNet, image classification, convolutional neural network