IHCNet: Advancing Multi-Stage HER2 Status Detection in Breast Cancer Using Interpretable Robust Transfer Learning Model from IHC Images

Human Epidermal Growth Factor Receptor (HER2) has become one the most dangerous subtype of Breast Cancer (BC). According to the guidance of CAP/ASCO, HER2 expression can be classified between 0 and 3+ range in terms of its severity level. it can be detected from immunohistochemical (IHC) of several classes such as 0, 1+, 2+, and 3+. Early diagnosis can alleviate and help pathologists to take early treatment. Utilizing artificial intelligence (AI) in digital pathology, a robust transfer learning (TL) model has been utilized known as IHCNet based on DenseNet201 to diagnose BC from the distinct expression of HER2 subtypes. This model is introduced with a fine-tuned combination of Global Average Pooling (GAP), Batch Normalization (BN), Dense Layers with Swish activation functions, Dropout layers, and a Dense classifier with a softmax function. IHCNet exceeded all other state-of-art models with an Accuracy of 93.45%, Precision of 94.01%, and Recall of 93.14%, on the BCI dataset, maintaining promising improvement. Finally, Grad-CAM has been incorporated into the proposed model to explain the model more intuitively. Grad-CAM explains IHCNet performances in different convolution layers by generating heat maps.