Development of multi-class computer-aided diagnostic systems using the NICE/JNET classifications for colorectal...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Endoscopic diagnosis of colorectal lesions can vary significantly among practitioners.
• Artificial intelligence (AI) offers a potential solution to standardize diagnostic accuracy.
• This study focuses on AI-driven diagnostic support utilizing NBI (Narrow Band Imaging) classification systems.

Purpose of the Study:

• To develop and evaluate computer-aided diagnosis (CADx) systems for colorectal lesions.
• To assess the performance of CADx systems based on the NICE and JNET classifications.
• To determine the potential of AI in equalizing endoscopic diagnostic capabilities.

Main Methods:

• Developed two CADx systems using Residual Network classifiers and NBI images.
• One CADx system (CADx-N) was based on the NICE classification.
• The second CADx system (CADx-J) was based on the JNET classification, with validation using specific image datasets and magnification levels.

Main Results:

• CADx-N achieved high accuracy for NICE Types 1, 2, and 3 (97.5%, 91.2%, 93.8%).
• CADx-J demonstrated strong performance across JNET types, with high sensitivity and specificity, particularly for Type 1 (100% sensitivity, 96.3% specificity) and Type 3 (62.5% sensitivity, 99.6% specificity).
• Diagnostic performance was consistent across different magnification levels for CADx-N.

Conclusions:

• Multi-class CADx systems utilizing both NICE and JNET classifications exhibit robust diagnostic performance.
• These AI tools can assist in educating less experienced endoscopists.
• The developed CADx systems show potential to aid in the accurate diagnosis of colorectal lesions.