Ulcerative Colitis Severity Assessment Powered by Vision Models

SITUATION

A large biopharma company with a strong immunology pipeline aimed to bring a new drug for Ulcerative Colitis (UC) to market. Accurate measurement of disease severity during clinical trials was critical to demonstrating drug efficacy and obtaining regulatory approval

CHALLENGE

Existing assessment methods, particularly the Mayo Endoscopic Subscore (MES), failed to capture the nuanced differences in disease severity across segments of the large intestine. MES scoring treated localized and widespread inflammation the same, overlooked partial healing, and was highly subjective, compromising the accuracy of trial outcomes and masking partial treatment responses.

APPROACH

Mu Sigma partnered with the company’s Clinical Analytics team to build machine learning and deep learning-based vision models that could analyze endoscopy videos with greater precision.

• Over 2,500 anonymized endoscopic videos were processed using OpenCV to extract frames.
• Frames were manually labeled using Mu Sigma’s in-house tool, muPercept, for features like colon segments, inflammation, forceps, and blood.
• Data pre-processing involved frame tagging, augmentation (via Sempler), and visual contrast techniques to manage variability in lighting and angles.
• 50+ neural network models were developed and tested using Mu Sigma’s proprietary classification modules to detect mucosal injury with improved granularity.
• Models delivered continuous-scale assessments rather than rigid 4-point scores, minimizing subjectivity and enhancing response measurement.

IMPACT

• Reduced physician workload by 90% through automated video timestamp detection and tagging.
• Enabled the client to publish findings and seek regulatory recognition for more sensitive and accurate UC assessment techniques.
• Positioned the company to better showcase treatment effectiveness in Irritable Bowel Disease (IBD) trials, improving the chances of drug approval.