AI-Enhanced Wide-Field Endoscopic Fluorescence Mapping of Gastrointestinal Mucosal Permeability in IBD - A Pilot Study in IBD Patients and Controls

• Feasibility of Procedure: The proportion of enrolled patients in whom the full fluorescence mapping procedure is successfully completed as intended. This will include metrics such as ability to intubate the colon and obtain fluorescence images from cecum to rectum; any significant technical difficulties encountered (and their nature); additional procedure time added by the imaging (in minutes); and whether the resulting fluorescence data was of analyzable quality. We will define success as achieving a usable permeability map of the colon surface. Feasibility will also be described by qualitative feedback from endoscopists on workflow integration.

Incidence and severity of adverse events related to the investigational aspects (fluorescein and AI system). This includes any allergic reactions to fluorescein, any hemodynamic instability during the procedure attributable to the dye, any complications from prolonged procedure time (e.g. hypoxia from sedation), or any device malfunctions leading to patient risk. Adverse events will be categorized as mild, moderate, severe, and their relationship to the study intervention will be adjudicated. The primary safety endpoint is the absence of any Serious Adverse Effect. We hypothesize that the rate of serious complications will be <5%, with the expectation of 0% severe allergic reactions in a 60 patient sample (based on fluorescein's known <1/1000 severe reaction rate). We will specifically report the number of patients with any fluorescein-related reaction (and details if so).

We will assess whether fluorescence permeability at each biopsy site correlates with endoscopic inflammation in that same anatomical region. Fluorescence metrics (e.g., normalized peak intensity, number of leak foci) from each biopsy site - high-fluorescence and low-fluorescence regions - will be compared with the corresponding local endoscopic severity score (Mayo endoscopic subscore for UC; the segmental SES-CD subscore for CD). Because biopsies are obtained from specific marked locations rather than from every colonic segment, analyses will be performed at the biopsy-site level rather than the segment level. Correlation coefficients (Spearman or Pearson) will quantify the association between fluorescence signal and endoscopic activity. We will also compare mean fluorescence values between biopsy sites taken from endoscopically inflamed vs non-inflamed regions. Controls will typically score as non-inflamed (Mayo = 0; SES-CD = 0).

Fluorescence permeability measurements at each biopsy site will be correlated with histologic inflammation scored using disease-appropriate indices (Nancy Index for UC; GHAS for CD and controls). Because each participant provides a paired set of biopsies (one high-fluorescence and one low-fluorescence site from the same anatomical region), comparisons will occur at the biopsy-site level. We will analyze:

• correlation between fluorescence intensity and histology score across all biopsy sites
• differences in histologic scores between high- and low-fluorescence sites within each participant
• exploratory pooled analyses using a unified ordinal inflammation scale Controls are expected to score as non-inflamed on GHAS.

For each participant, fluorescence metrics at the specific biopsy sites will be correlated with matched ex vivo ⁵¹Cr-EDTA flux from the same locations. Analyses will again be at the biopsy-site level. Additionally, we will perform within-patient gradient analysis, comparing the difference between high- and low-fluorescence sites:

• Δ fluorescence (high - low)
• Δ Ussing flux (high - low) A positive association between Δ fluorescence and Δ Ussing flux will indicate that higher in vivo leak corresponds to higher physiologic permeability.

Using the initial manual annotations and multidisciplinary review as the reference, we will calculate how accurately the AI algorithm identifies high-permeability regions. Metrics like sensitivity, specificity, and predictive values for AI-detected "leak sites" against human-identified or biopsy-proven inflammation sites will be reported. Including control participants (non-IBD) will help ensure that the AI model correctly identifies normal mucosa as low-permeability, improving specificity.

Although the AI provides an objective measure, we will check consistency by having multiple investigators review a subset of the fluorescence videos independently to mark obvious leak spots. The AI's output will be compared to each reviewer's marks (and reviewers to each other) to ensure the AI reduces inter-observer variability. High agreement would support reproducibility.