Clinical application of deep learning-based measurement of glomerular morphometry in IgAN.

Background: Accurate glomerular pathology assessment is crucial for predicting outcomes and guiding treatment in IgA nephropathy (IgAN). Although the Oxford classification provides a standardized histological assessment, it is semi-quantitative and subjective. Identifying objective and interpretable renal pathological features could enhance prognostic prediction. Methods: We developed and validated a deep learning-based computational pathology pipeline for glomerular analysis in IgAN. A glomerular segmentation and classification model was developed based on deep convolutional neural networks in 183 periodic acid-Schiff (PAS)-stained whole slide images (WSIs) of renal pathology from IgAN patients, and was further applied to WSIs of 886 IgAN patients from a prospective cohort. Associations of glomerular morphometry with clinical phenotypes and prognosis were determined. Results: The models demonstrated glomerular segmentation performance with a Dice coefficient of 0.91 and classification performance with an AUC of 0.94. Through a comprehensive correlation analysis of quantitative morphometric characteristics and clinical parameters, we found that age, gender, and body mass index significantly influenced glomerular morphology. No association was found between proteinuria and glomerular size. Notably, a nonlinear relationship was observed between glomerular size and estimated glomerular filtration rate (eGFR), with patients having an eGFR between 30 and 60 ml/min/1.73m Conclusions: In conclusion, we developed an end-to-end pipeline for glomerular segmentation, classification, and morphometry in WSIs. These objective and highly reproducible quantitative characteristics of glomeruli may serve as a valuable tool for therapeutic guidance and prognostic assessment in IgAN.