Large Scale Screening of Antibody-Stained Immunohistochemistry Images: An Observational Study on Pancreatic Islets Promotes the Pre-eminence of the ResNet50 Model

Abstract

Background: The deep neural network (DNN) is a growing field of artificial intelligence that aims to mimic human intelligence. The modern health care system has encouraged this hybrid technology to work smartly and faster to improve efficiency, reduce human error, and support personalized medicine. Aim: This article presents an advanced method based on DNN algorithms (ResNet50, EfficientNet, and U-Net) used for the analysis of Immunohistochemistry (IHC) images of pancreatic islets coated with anti-synaptophysin, anti-insulin, and anti-glucagon antibodies. This research aimed to overcome the time-consuming factor, which is a drawn-out process in manual interpretation, and serves as a valuable tool for advancing modern healthcare. Methods: The IS-IHC dataset was divided into two groups, training (70%) and test (30%) sets. The models were trained at 500 iterations using a Graphics Processing Unit (GPU) parallel computing architecture and the TensorFlow 2.15 deep learning framework. Several other performance metrics, such as precision, recall, Jaccard, F1 score, and prediction probability, were evaluated to validate the model's accuracy. Results: The ResNet50 model automatically detects and quantifies brown-stained regions in IHC images of pancreatic islets, achieving 93.22% accuracy for insulin detection, 90.39% for synaptophysin, and 82.59% for glucagon staining. The model calculates area percentages with precision scores of 93.62% (insulin), 90.33% (synaptophysin), and 89.29% (glucagon), closely matching manual ImageJ calculations, suggesting its superiority over the other two models. The ResNet50 performance was also comparable to manual calculations performed using ImageJ software. Moreover, a high F1 score and prediction probability indicate that the model makes both precise and comprehensive predictions. Conclusion: This study found that a ResNet50-based deep learning model enables robust, automated quantification of immunohistochemical markers in pancreatic islets. It achieves high concordance with manual annotation and offering a scalable, objective, and time-efficient computational tool to advance diagnostic accuracy and digital pathology integration in clinical practice.