Professionals

Authors: Priyadharshini Sivasubramaniam, Abdelrahman Shabaan, Rofyda Elhalaby, Bashar Hasan, Ameya A Patil, Saadiya Nazli, Adilson DaCosta, Byoung Uk Park, Lindsey Smith, Taofic Mounajjed, Stephen M Lagana, Chamil Codipilly, Puanani Hopson, Imad Absah, Christopher P Hartley, Rondell P Graham, Roger K Moreira

Journal: Journal of Pathology Informatics 2025

Paper of the Month – Selected and discussed by Luc Biedermann

🎯 Study Aim

The authors aimed to develop and validate an AI-based model on routine H&E slides to objectively quantify lamina propria fibrosis in EoE. They further sought to test whether this fibrosis score predicts fibrostenotic complications such as rings, strictures, and need for dilatation over longitudinal follow-up.

🔬 Methods

A supervised AI model was trained using a cloud-based platform (Aiforia Inc.) on 99 whole-slide images from multiple institutions and validated on 213 esophageal biopsies (100 adult, 113 pediatric) prospectively evaluated with standardized EoEHSS scoring. Clinical outcomes including rings, strictures, and need for dilatation were correlated with AI-derived fibrosis scores after a median follow-up of 31.4 months.

📊 Key Findings

• The AI fibrosis score correlated strongly with pathologist LPF assessments (Spearman R_s ≈ 0.64–0.69) but showed higher and more consistent predictive value for fibrostenotic outcomes than human readings.
• Higher AI fibrosis scores were significantly associated with subsequent rings, strictures, and need for dilatation, even in patients without prior strictures and in samples that pathologists had labeled as inadequate for LPF assessment.

Additional findings

• The model could extract clinically meaningful fibrosis information from much smaller lamina propria areas than pathologists usually deem sufficient, effectively rescuing a sizeable fraction of otherwise “inadequate” biopsies.
• LP area itself correlated with disease activity and fibrosis, and AI-derived adequacy thresholds (e.g., around 0.1–0.16 mm² LP per level) provided more objective criteria than traditional subjective adequacy judgments.
• A clinician survey indicated that gastroenterologists would likely alter monitoring or escalate therapy if provided with a validated AI-based fibrosis risk score.

🧠 Interpretation

This study demonstrates that AI-based quantification of lamina propria fibrosis provides an objective, reproducible, and clinically meaningful assessment that predicts fibrostenotic complications more consistently than conventional pathology evaluation. The model's ability to extract predictive information from limited biopsy samples addresses a major limitation in current EoE histological assessment. However, the predictive model needs external validation and currently requires proprietary software.

⚠️ Critical Appraisal

As someone who strongly believes that objective determination of subepithelial fibrosis will be key for accurate grading and staging of EoE, I highly welcome this work from the Mayo Clinic group. Having access to objective longitudinal follow-up data makes such AI-based tools particularly valuable for risk stratification and therapeutic decision-making. While this is undoubtedly important work that advances the field significantly, I believe tools like this should ideally be provided in an open-source fashion so that other scientific groups could validate, use, and potentially improve the model across diverse patient populations and clinical settings.

📝 Take-Home Message

Machine‑learning–based analysis of routine H&E slides according to that paper can (and potentially in the near future real world clinical practice WILL) detect and quantify lamina propria fibrosis and predict fibrostenotic complications in EoE more consistently than pathologist scoring. Integration of objective fibrosis measures into clinical practice could improve risk stratification and may prompt development of biopsy techniques and open‑source tools that enhance our ability to stage and monitor this chronic disease.