Search

BACKGROUND Microscopic inflammation has been shown to be an important indicator of disease activity in ulcerative colitis (UC). However, manual histologic scoring is semi-quantitative and subject to interobserver variation, and AI-based solutions often lack interpretability. Here we report two distinct quantitative approaches to predict disease activity scores and histological remission using AI-powered digital pathology. Both the random forest classifier (RFC) and graph neural network (GNN) further provide explainability and biological insight by identifying histological features informing model predictions. METHODS Convolutional neural networks (CNNs) were developed using >162k annotations on 820 WSI of H&E-stained colorectal biopsies for pixel-level identification of tissue regions (e.g. crypt abscesses, erosion/ulceration) and cell types (e.g. neutrophils, plasma cells). All WSI were scored by 5 board-certified pathologists using the Nancy Histological Index (NHI) to establish consensus ground truth. A rich, quantitative set of human interpretable features that capture CNN predictions of the tissue region and cell type across each WSI was extracted and used to train a RFC to predict slide-level NHI score. To test the hypothesis that tissue region spatial relationships and cellular composition can inform AI-based predictions of disease activity, a separate GNN was trained, using nodes defined by spatially-resolved CNN model-generated outputs, to predict NHI score. The RFC and GNN also predicted histologic remission (NHI RESULTS The RFC and GNN both predicted histologic remission with high accuracy (weighted kappa 0.87 and 0.85, respectively). Both models also identified histologic features relevant to disease activity predictions. Some features are well established, e.g. infiltrated epithelium or neutrophil cell features distinguish cases with histologic remission. The models also identified features beyond those assessed by the NHI, e.g. area proportion of basal plasmacytosis associated with predictions of NHI 2 and 3. Other features not previously implicated in UC disease activity were also identified, e.g. intraepithelial lymphocytes differentiate cases with NHI 3. CONCLUSIONS We report quantitative and interpretable AI-powered approaches for UC histological assessment. CNN identification of UC histology was used as input to two distinct disease activity classifiers that showed strong concordance with consensus pathologist scoring. Both approaches provide interpretable features that explain model predictions and that may be used to inform biomarker selection and clinical development efforts.