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

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Gastric cancer is a leading cause of mortality worldwide.
• Early diagnosis is crucial for effective gastric cancer treatment.
• Manual histopathological image analysis is time-consuming and prone to errors, necessitating automated solutions.

Purpose of the Study:

• To develop and evaluate a multi-stage AI model for automated gastric cancer detection in histopathological images.
• To explore the efficacy of state-of-the-art vision transformer models for feature extraction.
• To optimize classification performance through advanced feature selection and deep neural network training.

Main Methods:

• Feature extraction using 11 vision transformer models.
• Feature selection via ANOVA F-Test, Recursive Feature Elimination, and Ridge regression.
• Deep neural network classification optimized with Particle Swarm Optimization.

Main Results:

• The best performing model achieved 97.96% accuracy, 96.95% sensitivity, 98.61% specificity, 97.85% precision, and 97.40% F1-score.
• Optimal configuration involved the DPT model, union-based feature selection, and 160x160 image resolution.
• Other configurations also demonstrated high performance, with accuracies up to 97.21% (DPT, 120x120) and 95.78% (BEiT, 80x80).

Conclusions:

• Transformer-based feature extraction combined with effective feature selection significantly enhances diagnostic performance in gastric cancer.
• The proposed AI model offers a promising tool for accurate and rapid analysis of gastric histopathological images.
• This approach can support early diagnosis, potentially improving patient outcomes for gastric cancer.