Enhancing Gastric Cancer Diagnosis through Deep Learning and Ensemble Techniques: A Comprehensive Analysis of Sub-database Performance

Abstract

Gastric cancer is the fifth most prevalent cancer worldwide and the fourth most fatal. Early detection is critical for effective treatment. Histopathological examination is the established diagnostic method for gastric cancer. Recent advancements in computer technology have accelerated the use of digital tools to aid pathologists in diagnosing gastric cancer from pathological images. Ensemble learning is employed to enhance algorithm precision, involving the integration of multiple complementary learning models.The experimental platform focused on three subdatabases within GasHisSDB. Four deep learning classifiers, specifically VGG19, Inception-V4, ResNeXt, and ResNet152, were employed for classification experiments on the GasHisSDB database. The evaluation encompassed performance metrics, including accuracy, precision, recall, specificity, and F1 score.In the examination of the 160 x 160 pixel sub-database, ResNet152 stood out by delivering exceptional results in both categories. It achieved a remarkable accuracy of 98.02% in the "Normal" category and 87.9% in the "Abnormal" category. In the 120 x 120 pixel sub-database, ResNeXt displayed strong performance with a 96.98% accuracy in the "Normal" category, while its accuracy dropped to 89.09% in the "Abnormal" category. Notably, in the 80 x 80 pixel sub-database, ResNet152 emerged as the top performer with a remarkable 98.67% accuracy in the "Normal" category and 95.12% in the "Abnormal" category. Across these diverse sub-databases, ResNet152 consistently outperformed other models, maintaining high accuracy and precision while ensuring balanced performance in both categories.