A group of researchers was able to leverage a pair of artificial intelligence (AI) models to better assess disease activity in patients with ulcerative colitis.
A team led by Kathleen Supipto of PathAI used AI-powered digital pathology to compare two different quantitative approaches to predict disease activity scores and histologic remission.
The data were presented at the 2023 Crohn’s Disease and Colitis Conference in Denver.
Microscopic inflammation is an important indicator of disease activity in patients with ulcerative colitis. However, manual histological scoring is semi-quantitative and is often subject to inter-observer variability. Additionally, artificial intelligence-based solutions often lack interpretability.
The two solutions compared in this study were Random Forest Classifiers (RFC) and Graph Neural Networks (GNN). Both of these models were expected to identify histological features that inform model predictions to provide explainability and biological insight.
In this study, researchers developed a convolutional neural network with >162k annotations at 820 WSI of H&E-stained colorectal biopsies for pixel-level identification of tissue regions and cell types. .
Each WSI was scored by five board-certified pathologists using the Nancy Histological Index (NHI) to establish consensus ground truth. On the one hand, a rich and quantitative human interpretable feature set was extracted by capturing CNN predictions of tissue regions and cell types across each WSI. We also predicted her NHII score at the slide level by training the RFC.
To predict the NHI score, the team trained another GNN using nodes defined by the output generated by the spatially resolved CNN model, which enabled the spatial relationships of tissue regions and cellular We tested the hypothesis that composition can inform AI-based predictions of disease activity.
Finally, researchers calculated feature importance for all combinations of RFCs and applied GNNExplained to identify significant interactions between domains within the organization and identify features that contribute significantly to GNN predictions. bottom.
The results show that both models predicted histologic remission with high accuracy (weighted kappa of 0.87 and 0.85, respectively) and identified histologic features associated with predicting disease activity. However, several features, including infiltrating epithelial or neutrophilic cells, were well established to distinguish cases with histologic remission.
The model also identified features beyond those assessed by the NHI, such as the area proportion of basal plasmacytosis associated with NHI 2 and 3 predictions.
There were several novel features not previously implicated in ulcerative colitis disease activity, such as intraepithelial lymphocytes that differentiate NHI 3 cases.
“We report a quantitative and interpretable AI-powered approach for UC histological assessment,” the authors wrote. “CNN identification of UC histology was used as input to two different disease activity classifiers that showed strong agreement with consensus pathologist scoring. It provides capabilities that could be used to inform biomarker selection and clinical development efforts.”
Study “A Quantitative and Explainable Artificial Intelligence (AI)-Based Approach for Predicting Ulcerative Colitis Disease Activity from Hematoxylin and Eosin (H&E) Stained Whole Slide Images (WSI)” teeth, Crohn’s Disease and Colitis Conference.