Deep Learning Model Using Stool Pictures for Predicting Endoscopic Mucosal Inflammation in Patients With Ulcerative...

Area of Science:

• Gastroenterology
• Artificial Intelligence in Medicine
• Medical Imaging

Background:

• Ulcerative colitis (UC) endoscopic activity can influence stool characteristics.
• A novel deep learning model, DLSUC, was developed to analyze stool photographs for predicting UC mucosal inflammation.

Purpose of the Study:

• To develop and validate a deep learning model (DLSUC) for predicting endoscopic activity in ulcerative colitis patients using stool photographs.
• To compare the performance of DLSUC against fecal calprotectin (Fcal) in assessing endoscopic inflammation.

Main Methods:

• A prospective, multicenter study involving 6 tertiary hospitals.
• Patients provided stool photographs within a week before endoscopy.
• DLSUC was trained on 2,161 images from 306 patients and validated on 1,047 images from 126 patients, with performance compared to Fcal.

Main Results:

• DLSUC achieved an AUC of 0.801 for predicting endoscopic activity, comparable to Fcal's AUC of 0.837.
• Excluding rectal-sparing cases improved DLSUC's AUC to 0.849, with higher accuracy, sensitivity, and specificity.
• Patients classified as active by DLSUC showed a higher likelihood of disease relapse (P = 0.002).

Conclusions:

• DLSUC demonstrates significant discriminating power for predicting endoscopic activity in UC, comparable to Fcal.
• The model's accuracy improved in patients without rectal sparing.
• Stool photography analyzed by DLSUC presents a promising, non-invasive tool for monitoring UC activity and potentially predicting relapse.