Deep learning-based quantitative analysis of glomerular morphology in IgA nephropathy whole slide images and its prognostic implications

Abstract Kidney pathology of immunoglobulin A nephropathy (IgAN), which is the key finding of both diagnosis and risk stratification, involves labor-intensive manual interpretation as well as unavoidable interpreter-dependent variabilities. We propose artificial intelligence-based frameworks for quantitatively analyzing glomerular histologic features that can predict kidney progression in IgAN. A deep learning model, based on DeepLabV3Plus and EfficientNet-B3, was developed for segmenting glomeruli and quantifying the morphological features by using digitized whole slide images from seven tertiary hospitals. Subsequently, it was used for machine learning-based risk prediction of IgAN progression. Its predictability was compared with the conventional clinicopathologic feature-based model to demonstrate its comparable performance. In total, 1,241 whole slide images were obtained. The weighted averages of average precision and dice similarity coefficient were 0.795 and 0.721 in internal validation and 0.818 and 0.743 in external validation, respectively. Interestingly, image features-only-based kidney outcome prediction models showed similar predictability compared with clinical features-only-based models. In addition, incorporating an image-based deep learning model into the clinical features-based models enhanced predictabilities, although insignificant. These results show that quantitative glomerular histologic features are comparable to clinical data, suggesting that they may offer additional prognostic insights not covered by Oxford classification or other clinical parameters.