Surgical Insight-guided Deep Learning for Colorectal Lesion Management

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Colonoscopy is crucial for diagnosing gastrointestinal diseases, including cancer.
• Accurate lesion identification during colonoscopy remains a challenge.
• AI and machine learning offer potential for enhanced medical image analysis.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) model, ColoNet, for detecting lesions in colonoscopic images.
• To assess the diagnostic accuracy of the DL model using established metrics.

Main Methods:

• A DL model, ColoNet, was developed using the YOLOv8 architecture.
• 1760 colonoscopic images from 306 patients were used for training and validation.
• Data augmentation techniques were employed to enhance model performance.

Main Results:

• The model achieved a precision of 0.796 and recall of 0.781 on the validation dataset.
• On a real-time dataset, ColoNet demonstrated 70.73% sensitivity and 92.00% specificity.
• The model achieved an overall accuracy of 82.42% with a positive predictive value of 87.88%.

Conclusions:

• ColoNet shows potential as an assistive tool for surgeons during colonoscopies.
• The model's ability to detect suspicious lesions aids in early colorectal cancer diagnosis.
• Further multicentric, prospective research is recommended for clinical validation.