Artificial Intelligence for Quantifying Endoscopic Mucosal Ulceration in Crohn's Disease

Area of Science:

• Gastroenterology and Medical Imaging
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Assessing Crohn's disease (CD) endoscopic activity is challenging due to patchy, variable mucosal damage.
• Standard scoring systems like the Simple Endoscopic Score for Crohn's Disease (SES-CD) have limitations in quantifying disease burden.
• Computer vision endoscopic (CVE) assessment offers a potential solution for objective disease quantification.

Purpose of the Study:

• To evaluate the performance of computer vision endoscopic (CVE) assessment in quantifying Crohn's disease (CD) mucosal ulceration and injury.
• To compare the accuracy and agreement of CVE measures with the established SES-CD.
• To determine if CVE can effectively differentiate between clinical remission and active disease states.

Main Methods:

• Colonoscopy videos from phase 3 clinical trials (STARDUST, SEAVUE) were analyzed using CVE.
• A segmentation model was trained on expert annotations to identify and quantify ulcer area, severity, and size.
• CVE-derived metrics were compared to SES-CD scores for quantification, localization, and correlation with clinical remission (CDAI <150).

Main Results:

• CVE ulcer segmentation models achieved performance comparable to human annotators (DICE 0.591 vs. 0.462).
• CVE measures strongly correlated with SES-CD scores (r=0.73-0.85), indicating good agreement.
• CVE measures demonstrated a greater effect size in separating clinical remission status than SES-CD (g=0.416 vs. g=0.290).

Conclusions:

• Computer vision endoscopic (CVE) assessment provides automated quantification of ulceration and mucosal injury in Crohn's disease.
• CVE shows conceptual agreement with SES-CD and offers enhanced granularity for endoscopic assessment.
• This technology has the potential to improve personalized disease management in CD.