IGUANA

IGUANA applies an interpretable graph neural network to whole-slide images of colon biopsies to distinguish normal from abnormal large bowel endoscopic biopsies and prioritize cases for pathologist review.

Key Features:

• Graph Neural Network Architecture: Represents whole-slide images (WSIs) as graphs with glands as nodes and assigns each node interpretable features derived from clinically relevant histological characteristics.
• Interpretable and Explainable Outputs: Produces heatmap overlays and numerical feature attributions that link model predictions to glandular architecture, inflammatory cell density, and spatial relationships between tissue components.
• High Predictive Accuracy: Reports an internal AUC-ROC of 0.98 and AUC-PR of 0.98, with mean external AUC-ROC of 0.97 and mean external AUC-PR of 0.97 across external datasets.
• Reduction in Pathologist Workload: Identifies normal biopsies with 99% sensitivity and can reduce the number of slides requiring pathologist review by approximately 55%.

Scientific Applications:

• Colon biopsy screening and triage: Screens and triages large bowel endoscopic biopsies to separate normal from abnormal cases within colon cancer screening workflows.
• Decision support for abnormal cases: Provides feature-level explanations and highlights diagnostically important regions to support pathologist interpretation of abnormal biopsies.
• Histopathology resource prioritization: Enables prioritization of cases to optimize allocation of histopathology and pathologist review resources.

Methodology:

WSIs are represented as graphs with glands as nodes and interpretable histological feature vectors as node attributes; an interpretable graph neural network generates predictions, heatmap overlays, and numerical feature attributions; model training and internal validation used data from one UK NHS site with external validation on two additional NHS sites and one site in Portugal, totaling 6,591 WSIs from 3,291 patients.