Detection of Polyps Suspected of Colorectal Cancer in Endoscopy Images of Using a Network Based on Transformer

Artificial intelligence (AI) has significantly impacted the healthcare sector, particularly in improving diagnostic and prognostic accuracy in medical imaging. This manuscript illustrates that AI integration can enhance polyp characterization accuracy, matching the proficiency of expert practitioners, and thereby reducing diagnostic errors and unnecessary interventions. By empowering general endoscopists with AI tools, underserved areas can ensure comprehensive patient care irrespective of medical expertise levels. The study focuses on developing an automated method for colon cancer categorization in colonoscopy images, leveraging a dataset from Kerman province and evaluating it against the publicly available ELSIVIER dataset. Employing deep learning techniques, specifically the Trans model, the research aims to detect and classify colonoscopy images into polyp-present and polyp-absent categories. The TransResU-Net architecture combines the strengths of residual networks, transformer blocks, and dilated convolutions, facilitating effective real-time polyp segmentation. Results show the model achieved a 98.90% accuracy on the private dataset, with a recall of 94.18% and an F1 score of 0.9641.