Computer-aided automatic classification of tissue pathological images is crucial for the early diagnosis of breast cancer. When dealing with multiple features, a Convolutional Neural Network (CNN) typically prioritizes obtaining global semantic features, while overlooking local details, which results in insufficient feature information. In this paper, we propose a dual-channel deep feature adaptive fusion algorithm called DEADF-Net. DEADF-Net first uses the EfficientNetV2 architecture network with an efficient channel-attention ECA to obtain global semantic features, and then utilizes the MobileNetV2 architecture network to obtain local detail features. Then, combine the two normalized depth features into fused features. Finally, the fused features are inputted into the fully connected layers for classification of benign and malignant tumors. In addition, the method uses transfer learning to optimize the depth features, thus improving the performance and robustness of the model. The performance of DEADF-Net was evaluated using two commonly used benchmark datasets for breast cancer histopathological image classification, namely BreakHis and BHI. The results showed that the accuracy on BreakHis is 40 ×: 95.8%, 100 ×: 96.1%, 200 ×: 97.4%, 400 ×: 96.6%, and BHI is 96.3%, which is superior to most mainstream single feature classification models.
Paper
Full text
Deep Feature Adaptive Fusion Algorithm for Breast Cancer Histopathology Image Classification
OpenAlex · AI in cancer detection · 2024
Abstract
Computer-aided automatic classification of tissue pathological images is crucial for the early diagnosis of breast cancer. When dealing with multiple features, a Convolutional Neural Network (CNN) typically prioritizes obtaining global semantic features, while overlooking local details, which results in insufficient feature information. In this paper, we propose a dual-channel deep feature adaptive fusion algorithm called DEADF-Net. DEADF-Net first uses the EfficientNetV2 architecture network with an efficient channel-attention ECA to obtain global semantic features, and then utilizes the MobileNetV2 architecture network to obtain local detail features. Then, combine the two normalized depth features into fused features. Finally, the fused features are inputted into the fully connected layers for classification of benign and malignant tumors. In addition, the method uses transfer learning to optimize the depth features, thus improving the performance and robustness of the model. The performance of DEADF-Net was evaluated using two commonly used benchmark datasets for breast cancer histopathological image classification, namely BreakHis and BHI. The results showed that the accuracy on BreakHis is 40 ×: 95.8%, 100 ×: 96.1%, 200 ×: 97.4%, 400 ×: 96.6%, and BHI is 96.3%, which is superior to most mainstream single feature classification models.