Multi-classification of breast cancer pathology images based on a two-stage hybrid network

In current clinical medicine, pathological image diagnosis is the gold standard for cancer diagnosis. After pathologists determine whether breast lesions are malignant or benign, further sub-type classification is often necessary.

For this task, this study designed a multi-classification model for breast cancer pathological images based on a two-stage hybrid network. Due to limited sample size for breast sub-type data, this study selected the ResNet34 network as the base network and improved it as the first-level convolutional network, using transfer learning to assist network training. In order to compensate for the lack of long-distance dependencies in the convolutional network, the second-level network was designed to use Long Short-Term Memory (LSTM) to capture contextual information in the images for predictive classification.

For the 8 sub-types of breast cancer classification on the BreakHis (40×, 100×, 200×, 400×) dataset, the ensemble model achieved accuracy rates of 93.67%, 97.08%, 98.01%, and 94.73% respectively. For the 4 sub-types of breast cancer classification on the ICIAR2018 (200×) dataset, the ensemble model achieved accuracy, precision, recall, and F1 Score rates of 93.75%, 92.5%, 92.5%, and 92.5% respectively.

The results show that the multi-classification model proposed in this study outperforms other methods in terms of classification performance, and further demonstrate that the proposed RFSAM module is beneficial for improving model performance.