An Improved ResNet Architecture for Accurate Patch-Level HER2 Overexpression Classification in Breast Cancer Tissue Images
Shubhangi Joshi, Pallavi Sachin Chaudhari, Anupkumar M. Bongale, Deepak Sudhakar Dharrao, Vivek Dugad, Anand Bhosale · Engineering Technology & Applied Science Research · 2025
With the advent of artificial intelligence, various computer-aided diagnostic systems are being developed to assist medical professionals. Deep learning techniques powered by convolutional neural networks seem promising for obtaining new insights into the onco-histopathology domain. Breast cancer is confirmed by histopathological analysis of Hematoxylin and Eosin (H&E)-stained breast tissue images. Finding the molecular subtype of breast cancer using Immunohistochemistry (IHC)-stained breast tissue is essential to decide on the correct treatment plan for a breast cancer patient. IHC staining is an expensive process that is very time-consuming and involves intra- and inter-observer subjectivity. This work attempts to find the Human Epidermal growth factor Receptor Two (HER2) molecular subtype from H&E-stained tissue images instead of using IHC-stained tissues. H&E-stained tissue image data from two separate sources are used to predict HER2 status. This study aimed to improve the accuracy of HER2 overexpression classification by modifying the architecture of the ResNet50 model by cascading it with a squeeze and excitation block and a depth-wise separable convolutional layer. The dataset comprises a combination of tissue image patches from a publicly available Warwick dataset and a real-world dataset collected from a hospital in Pune, India. The dataset is preprocessed and split into 60% train, 20% validation, and 20% test subsets. The proposed architecture with a modified ResNet50 network achieves the best patch-level HER2 classification accuracy of 98.04%, with class-specific test accuracy results for HER2 negative, HER2 equivocal, and HER2 positive scores being 97.73%, 99.70%, and 98.93%, respectively.