Artificial intelligence-assisted colonic endocytoscopy for cancer recognition: a multicenter study.

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Large adenomas are often misdiagnosed as cancer during colonoscopy, leading to unnecessary surgeries.
• An artificial intelligence (AI) tool was developed to accurately identify cancerous pathology in vivo.
• The AI tool underwent evaluation for regulatory approval.

Purpose of the Study:

• To evaluate the performance of a novel AI tool for in vivo pathological classification of colorectal lesions.
• To assess the AI tool's specificity and accuracy in identifying cancerous tissue during colonoscopy.

Main Methods:

• The AI tool was trained on 68,082 endocytoscopic images from six academic centers.
• A validation set of 500 endocytoscopic images from 50 large colorectal lesions was used.
• The AI tool predicted pathology (cancer, adenoma, non-neoplastic) for images at 520-fold magnification.

Main Results:

• The AI tool analyzed 83.6% of the 500 validation images.
• Overall accuracy for the analyzable images was 91.9%.
• The AI tool achieved 91.8% sensitivity and 97.3% specificity in identifying cancer.

Conclusions:

• The developed AI system for endocytoscopy shows high specificity in detecting colorectal cancer.
• This AI tool has the potential to reduce overtreatment of benign colorectal lesions.