Automatic Scoring Method for Tumor IHC Images Based on Deep Learning and Its Application on P53 Protein

Immunohistochemical (IHC) assay is a commonly used auxiliary technique in pathological diagnosis. Compared to the conventional manual scoring methods that are complicated and time-consuming, automated scoring methods have been playing a more and more important role in the development of digital medicine due to their adaptability and consistency. This study proposes an automatic scoring model for tumor IHC images, which mainly consists of a module for extracting the regions of interest (ROI) and a feature fusion scoring network. The former module extracts the effective tissue regions and the nuclear regions as prior knowledge to exclude cytoplasmic staining interference. The feature fusion network includes two branches. The main branch network combines the structure of cross-block stitching feature maps and the frequency channel attention networks (FcaNet) to extract the features of the effective tissue region images. The other branch network extracts the color representation vector of the cell nucleus region images. The fully-connected layers combine the features from both branches to give a comprehensive final score as the result. We performed experiments on IHC images of P53 protein in colorectal cancer. The results show that the proposed P53Net achieves better classification results than the commonly used classification models, with 94.21% accuracy, 89.24% F1-Score, and 0.9136 kappa coefficient.