Screening of normal endoscopic large bowel biopsies with interpretable graph learning: a retrospective study

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Computational pathology

Background:

• Pathologist workload is increasing, necessitating tools to optimize resource allocation.
• Accurate and efficient review of large bowel endoscopic biopsies is crucial for timely diagnosis and patient management.

Purpose of the Study:

• To develop an interpretable artificial intelligence (AI) algorithm for classifying large bowel endoscopic biopsies as normal or abnormal.
• To reduce the burden on pathologists by automating the screening of normal biopsies.
• To aid in the early diagnosis of gastrointestinal conditions.

Main Methods:

• A graph neural network, incorporating domain knowledge, was developed to analyze 6591 whole-slide images (WSIs) from 3291 patients.
• The model, named Interpretable Gland-Graphs using a Neural Aggregator (IGUANA), was trained and validated on data from multiple UK and Portuguese sites.
• Clinically driven interpretable features were used for classification.

Main Results:

• The IGUANA model achieved high performance with an area under the curve-receiver operating characteristic (AUC-ROC) of 0.98 and AUC-precision-recall (PR) of 0.98 in internal validation.
• Consistent performance was observed in external validation datasets (mean AUC-ROC=0.97, AUC-PR=0.97).
• At 99% sensitivity, the model can reduce the review of normal slides by approximately 55% and provides explainable heatmaps and feature-based predictions.

Conclusions:

• The AI model demonstrates high accuracy and consistency, showing potential for optimizing pathologist resources.
• Explainable predictions can enhance pathologist confidence and guide diagnostic decisions.
• The algorithm is poised for clinical adoption to improve efficiency and diagnostic accuracy in digital pathology.