AWCDL: Automatic weight calibration deep learning for detecting HER2 status in whole-slide breast cancer image

Defining an ERBB2 ( HER2/neu ) gene amplification status is critical to guiding human epidermal growth factor receptor 2 (HER2)-targeted therapy in breast cancer. Up to 40% of breast cancer patients are reported as having an immunohistochemistry (IHC) of HER2 2+ and requiring additional testing using fluorescence in situ hybridization to confirm the results. This paper aims to establish an automatically weighted calibration deep learning (AWCDL) algorithm to predict ERBB2 amplification based on IHC images. In this study, we applied IHC HER2 2+ images from 1,073 breast cancer patients at three cancer centers in China and extracted 376,099 tiles. Among these, 269,664 tiles were used for internal and external validation. The designed AWCDL consists of two steps. In Step 1, the internal validation achieved an accuracy of 89%, with a specificity of 0.89 and a sensitivity of 0.89. The external validation in the two other centers showed an average accuracy of 85%, with a specificity of 0.86 and a sensitivity of 0.82. In Step 2, the model achieved higher accuracy for the slides predicted as negative in Step 1 by automatically calibrating the weight. Collectively, these results suggest that this AWCDL model has successfully proved useful as an alternative method to fluorescence in situ hybridization for assessing the ERBB2 amplification status in breast cancer.