Robust automated prediction of the revised Vienna Classification in colonoscopy using deep learning: development and...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Need for accessible optical diagnostic technology outside expert centers.
• Development of an AI system for robust pathological diagnosis prediction.
• Utilizes standard colonoscopy images and the revised Vienna Classification.

Purpose of the Study:

• To develop and validate an AI system for automated pathological diagnosis of colonoscopy images.
• To assess the AI system's diagnostic performance against expert endoscopists.
• To improve the accessibility and accuracy of colorectal neoplasia diagnosis.

Main Methods:

• Trained deep learning algorithms (ResNet152) on a large dataset of colonoscopy images with pathologically proven lesions.
• Classified lesions based on the revised Vienna Classification (categories 1, 3, 4/5) and normal images.
• Validated the AI system's performance through internal and external datasets, comparing it with endoscopist performance.

Main Results:

• Internal validation showed high sensitivity (84.6%) and specificity (99.7%) for adenoma.
• External validation demonstrated strong performance for neoplastic lesions (sensitivity 88.3%, specificity 90.3%).
• The AI system's diagnostic performance surpassed that of expert endoscopists, with an AUC of 0.903.

Conclusions:

• The AI system provides reliable differential diagnoses for colorectal neoplasia during colonoscopy.
• It empowers non-expert endoscopists to achieve diagnostic accuracy similar to experts.
• This technology enhances the quality of care in diverse clinical settings.