HAHNet: A convolutional neural network for HER2 status classification of breast cancer

Abstract

Background: Breast cancer is a major health problem for women. Human epidermal growth factor receptor-2 (HER2) is a very important diagnostic and prognostic factor for breast cancer, and HER2 status classification is essential for the development of treatment plans for breast cancer. Generally speaking, pathologists will adopt immunohistochemistry (IHC) to assess HER2 status, which requires additional economic costs. Furthermore, the manual assessment of HER2 status is time-consuming and error-prone. In recent years, deep learning has been widely used in medical field and has attained great achievements. However, the existing deep learning methods for HER2 status classification of conventional hematoxylin and eosin (H&E) stained images are not accurate enough. Results: To address these problems, a neural network model named HAHNet is proposed in this paper. HAHNet combines multi-scale features with attention mechanisms, which is able to directly classify HER2 status of H&E stained histological images of breast cancer. Typically, the HAHNet network mainly includes convolution preprocessing, attention mechanism, downsampling, and multi-scale feature extraction. The experimental results show that HAHNet outperforms other existing methods with regard to six metrics of Accuracy, Sensitivity, Precision, F-score, MCC, and AUC. Conclusions: Collectively, the above experiments demonstrate that our proposed HAHNet achieves high performance in classifying the HER2 status of breast cancer using only H&E stained samples, which can be used in case classification and helps to reduce the cost required for diagnosis.