Artificial-intelligence-based decision support tools for the differential diagnosis of colitis

Area of Science:

• Gastroenterology
• Medical Artificial Intelligence
• Computational Pathology

Background:

• Artificial Intelligence (AI) tools are emerging in gastroenterology, but their application in Inflammatory Bowel Disease (IBD) diagnosis is nascent.
• Developing AI to differentiate IBD from infectious and ischemic colitis using endoscopic visuals and patient data is crucial.

Purpose of the Study:

• To establish and evaluate AI-based algorithms for distinguishing IBD from infectious and ischemic colitis.
• To compare the performance of different AI approaches (CNN, GBDT, hybrid) and expert endoscopists.

Main Methods:

• A Convolutional Neural Network (CNN) was trained on 1796 endoscopic images from 494 patients across three colitis types.
• A Gradient Boosted Decision Trees (GBDT) algorithm utilized five clinical parameters, and a hybrid CNN+GBDT approach was also evaluated.
• All methods were benchmarked against each other and expert endoscopists on independent test datasets.

Main Results:

• The GBDT algorithm achieved the highest accuracy (.792) and area under the ROC curve (.888), outperforming the CNN (.709) and hybrid approach (.766).
• The GBDT algorithm, using clinical parameters, demonstrated superior performance compared to the CNN and expert endoscopists.
• While CNN accuracy (.721) matched endoscopists, it was surpassed by the GBDT model.

Conclusions:

• AI decision support systems solely based on endoscopic images are not yet ready for widespread clinical use in colitis diagnosis.
• Further development with more diverse image datasets is needed to enhance AI performance.
• The clinical utility of the GBDT algorithm based on clinical parameters warrants validation in prospective studies.