Transformer-Based Feature Extraction and Optimized Deep Neural Network for Gastric Cancer Detection

Gastric cancer is among the most common diseases worldwide and can lead to fatal outcomes. Early diagnosis significantly increases the success of treatment, and accurate and rapid analysis of histopathological images is of enormous importance. However, since manual evaluation of these images is time-consuming and open to observational errors, the need for automatic diagnosis systems supported by artificial intelligence is increasing. In this study, a multi-stage artificial intelligence-based model that performs cancer detection on gastric histopathological images is proposed. In the first stage, features were extracted from the images using 11 different state-of-the-art vision transformer models. Then, the most significant features were determined by using feature selection methods such as ANOVA F-Test, Recursive Feature Elimination, and Ridge regression, and separate feature sets consisting of the intersections and unions of these features were created. The obtained feature sets were trained with a deep neural network model optimized with the Particle Swarm Optimization algorithm to increase the classification performance, and the detection of gastric tissues was achieved. Among the tested configurations, the highest classification performance was obtained using 160 × 160 image resolution, the DPT model, and union-based feature selection. This configuration achieved 97.96% accuracy, 96.95% sensitivity, 98.61% specificity, 97.85% precision, and a 97.40% F1-score. Additionally, strong results were observed with other configurations, such as 97.21% accuracy using the DPT model with 120 × 120 images, and 95.78% accuracy with the BEiT model at 80 × 80 resolution. These findings demonstrate that transformer-based feature extraction methods, when combined with effective feature selection strategies, can significantly enhance diagnostic performance.