An Optimized Deep Learning method for Gastric Cancer Image Classification using Histopathological Images

Abstract

This study investigates deep learning–based approaches for automated gastric cancer detection using histopathological images from the GasHisSDB dataset. Traditional diagnostic methods for gastric cancer are labor-intensive and reliant on expert interpretation, motivating the development of efficient AI-driven systems for early detection. Multiple architectures were evaluated, including Neural Networks, VGG19, MobileNet, EfficientNet_B0, and EfficientNet_B1, to identify the most effective model for classification. To ensure robust model selection and unbiased performance assessment, a nested cross-validation framework was employed, combined with both randomized and grid-based hyperparameter tuning for each architecture. Within this rigorous evaluation strategy, EfficientNet_B1 achieved the highest accuracy of 99.46%, outperforming Neural Networks (91.36%), MobileNet (92.69%), VGG19 (76.96%), and EfficientNet_B0 (79.25%). The results highlight the superior scalability and feature-learning capacity of EfficientNet_B1, underscoring its potential for reliable and rapid gastric cancer diagnosis in clinical workflows.