DOP095AI inflammation and architectural model differentiates active and remission Ulcerative Colitis states across histological scoring systems

Abstract Background Histological assessment of ulcerative colitis (UC) is used in trials to gain a deeper insight into mucosal healing than endoscopy alone. Many different and overlapping scoring systems are used, creating complexity and inconsistency across trials. In addition, there is a high degree of inter-rater variability in pathologist central reading. Our aim was to objectively measure histological remission by quantifying both standard and novel characteristics of UC. Methods Convolutional neural networks were developed with two analytical branches. The first quantified lymphocytes, neutrophils, eosinophils, and plasma cells. The second segmented crypts, lamina propria (LP), muscularis mucosa (MM) and derived crypt morphometrics (mucin depletion, solidity, roughness, branching, and abscess formation) (Fig.1). Model training used 62,000 annotated image pairs (80% training, 20% validation) and independently tested on 30,000 annotation pairs from 26 held-out whole slide images (WSIs). The AI-assisted model was applied to quantify inflammatory cell densities and crypt morphometrics across 223 WSIs from 57 patients. These were associated with a universal scoring system, designed for conversion to Simplified Geboes Score (SGS), Nancy Index (NI), and Robarts Histopathology Index (RHI) provided by two pathologists with correlations assessed using Spearman’s rank (rs). Results Compared with pathologist-assessed active (Nancy ≥1) and remission disease, the AI model detected increases in neutrophil density in the LP (3.5-fold) and crypts (15-fold). In the LP, increases were also observed in eosinophils (2.2-fold), lymphocytes (1.4-fold), and plasma cells (1.2-fold). Architectural changes were evident in active disease, including an increase in mucin depletion (1.2-fold), crypt branching (1.4-fold), crypt abscess formation (10-fold), crypt roughness (1.1-fold), and reduced crypt solidity (0.9-fold) (p 0.0001 for all) (Fig.2). AI-derived metrics correlated with 12 pathologist-derived universal scoring system criteria and overall NI. For example, neutrophil quantification correlated with both pathologists in the LP (rs = 0.49-0.50) and crypts (rs=0.37-0.38), while mucin depletion also showed strong correlation (rs = 0.58-0.62) (p 0.0001 for all). Similar trends were observed for SGS and RHI. Conclusion This model provides new objective insight into ulcerative colitis histology by quantifying established (e.g. crypt neutrophils) and novel (e.g. mucin depletion) features. As a pathologist support tool, this model has the potential to reduce inter-rater variability while providing quantitative insights into histological features that may reveal new aspects of drug effect. Conflict of interest: Windell, Dylan: I disclose the following financial relationship(s) with a commercial interest​: Perspectum Ltd.​ Magness, Alastair: I am an employee of Perspectum Ltd. Li, Rongxue: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Davis, Tom: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Beyer, Cayden: Employee of Perspectum Thomaides Brears, Helena: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Larkin, Sarah: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Aljabar, Paul: Employee of Perspectum Ltd. Kainth, Reema: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Wakefield, Phil: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Langford, Caitlin: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Fryer, Eve: No conflict of interest Goldin, Rob: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Landy, Jonathan: No conflict of interest