A novel machine learning-based algorithm to identify and classify lesions and anatomical landmarks in colonoscopy...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Computer-aided diagnosis (CAD) with AI shows promise for colon polyp detection.
• AI application for identifying normal colon landmarks and differentiating various colon diseases is not yet established.
• Convolutional neural networks (CNNs) are powerful tools for image analysis.

Purpose of the Study:

• To develop a CNN-based algorithm (GUTAID) for recognizing colon lesions and anatomical landmarks.
• To evaluate the accuracy of AI in identifying normal colon structures and diverse colonic diseases.
• To establish AI classification methodology for multiple colon disease identification.

Main Methods:

• Development of the GUTAID algorithm using a 16-layer Visual Geometry Group (VGG16) architecture.
• Implementation of two sub-models: Normal, Polyp, Diverticulum, Cecum and CAncer (NPDCCA) and Narrow-Band Imaging for Adenomatous/Hyperplastic polyps (NBI-AH).
• Training and validation using 7838 colonoscopy images, with independent verification on 1273 images.

Main Results:

• GUTAID achieved high detection accuracy: 93.3% for polyps, 93.9% for diverticula, 91.7% for cecum, and 97.5% for cancer.
• Specific accuracy for adenomatous/hyperplastic polyps was 83.5%.
• The AI system demonstrated high performance in identifying various colonic abnormalities and landmarks.

Conclusions:

• A CNN-based algorithm (GUTAID) was successfully developed for identifying colonic abnormalities and landmarks with high accuracy.
• The GUTAID system can assist in characterizing polyps for optical diagnosis.
• AI classification is a feasible approach for identifying multiple and diverse colon diseases.