Artificial intelligence model for analyzing colonic endoscopy images to detect changes associated with irritable...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Irritable Bowel Syndrome (IBS) is often not considered an organic disease with no visible abnormalities on endoscopy.
• Recent findings suggest potential microscopic changes like biofilm, dysbiosis, and microinflammation in IBS patients.
• Current diagnostic methods for IBS may miss subtle endoscopic indicators.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) colorectal image model for detecting minute endoscopic changes indicative of IBS.
• To assess the AI model's ability to differentiate between IBS subtypes (IBS-C, IBS-D) and healthy controls.
• To investigate AI's potential in identifying IBS-associated abnormalities not typically visible to human endoscopists.

Main Methods:

• Utilized Google Cloud Platform AutoML Vision for single-label classification to build AI image models.
• Collected colonoscopy images from IBS patients (IBS, IBS-C, IBS-D) and asymptomatic healthy subjects (Group N).
• Calculated model performance metrics including sensitivity, specificity, predictive value, and Area Under the Curve (AUC).

Main Results:

• The AI model achieved an AUC of 0.95 in discriminating between IBS patients and healthy subjects.
• For IBS detection, the model showed 30.8% sensitivity, 97.6% specificity, 66.7% positive predictive value, and 90.2% negative predictive value.
• The overall AUC for discriminating between healthy controls and IBS subtypes (IBS-C, IBS-D) was 0.83.

Conclusions:

• An AI-powered colorectal image analysis model can effectively discriminate colonoscopy images of IBS patients from those of healthy individuals.
• The AI model demonstrates high specificity in identifying IBS, suggesting potential for objective diagnostic support.
• Further prospective studies are required to validate the AI model's diagnostic capabilities across different institutions and its utility in treatment efficacy assessment.