BACKGROUND. Endoscopy-based disease severity assessment in Inflammatory Bowel Disease (IBD) clinical trials is typically done using human-read scoring systems such as the Mayo Endoscopic Score (MES) or Ulcerative Colitis Endoscopic Index of Severity (UCEIS). Computer vision and artificial intelligence (AI) have the potential to automate and improve upon these measurements. Here we present an R&D toolbox (Arges) designed to facilitate rapid prototyping and evaluation of AI-derived disease severity biomarkers for IBD.
METHODS. Arges (AI-reported gastrointestinal endoscopic scoring system) is a suite of AI models capable of rapid, accurate, and reproducible scoring of endoscopic disease severity in IBD. Current capabilities (Arges v1.0) include automation of MES and UCEIS scoring as well as endoscope localization throughout the colon and terminal ileum. Arges also can compute novel severity scores such as Arges-CMES, a continuous MES score that can be computed over the whole colon (Arges-CMES_ALL) or as an average over multiple left colon segments (Arges-CMES_SEG). Arges models were trained using a large collection of clinical trials (UNIFI: NCT02407236, JAK-UC: NCT01959282, SEAVUE: NCT03464136, and TRIDENT: NCT02877134) that account for over one thousand hours of endoscopy video (sixty million frames). Arges v1.0 is based on a self-supervised, pre-trained foundational vision transformer (ViT) network, which extracts features from videos, and a second transformer network, which uses these features to learn to perform tasks of interest such as disease severity estimation.
To facilitate video/frame annotation and model evaluation, we developed ArgesVIEW: an in-browser visualizer that enables collaboration and rapid evaluation of AI models while optimizing for flexibility, traceability, and minimal impact on clinician’s workflow.
RESULTS. Models trained using Arges have encouraging performance (AUC = 0.8 for MES, AUC=0.79 for UCEIS, AUC=0.93 for segment-wise endoscope location). By combining disease severity models and segment location, Arges can also provide measures of disease extension which, coupled with ArgesVIEW (Fig. 1), can complement traditional disease scoring approaches.
CONCLUSIONS. Arges allows for rapid training of models that can significantly impact clinical trial R&D. Because Arges is based on a foundational feature extraction method (pre-trained ViT), the training of entirely new models from scratch (e.g., UCEIS estimation) can be achieved within minutes. With ArgesVIEW, model evaluation as well as in-video annotations can be performed quickly, facilitating the incorporation of AI-derived scores into clinical trial workflows. Arges significantly boosts analysis, productivity and facilitates gathering of novel insights during the R&D phase of IBD clinical trial development.

Shreyas Fadnavis

Appearances