Development of a real-time endoscopic image diagnosis support system using deep learning technology in colonoscopy

Area of Science:

• Medical technology
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Colonoscopy skills vary among endoscopists, leading to missed colorectal neoplasms.
• Early detection of colorectal cancer is crucial for improving patient outcomes.
• A need exists for advanced tools to assist in identifying subtle lesions.

Purpose of the Study:

• To develop a real-time artificial intelligence (AI) system for automatic detection of early colorectal cancer signs during colonoscopy.
• To enhance the detection of difficult-to-identify non-polypoid lesions.
• To reduce the rate of missed abnormalities during colonoscopy procedures.

Main Methods:

• Development of an AI system trained to automatically detect early colorectal cancer indicators.
• Evaluation of the AI system's performance using sensitivity, specificity, and area under the curve (AUC) metrics.
• Optimization of the AI model through tensor metric decomposition for accelerated image processing (21.9 ms/image).

Main Results:

• The AI system demonstrated high performance in the validation set: 97.3% sensitivity and 99.0% specificity (AUC 0.975).
• Sensitivity was particularly high for polypoid (98.0%) and non-polypoid (93.7%) subgroups.
• The system provides real-time predictions, aiding endoscopists in identifying lesions quickly.

Conclusions:

• The developed AI system is effective in supporting endoscopists for high detection rates, especially for non-polypoid lesions often missed.
• Real-time alerts from the AI system can prevent the omission of abnormalities during colonoscopy.
• This technology holds significant potential for improving the early detection of colorectal cancer.