MO062: Pan-Disease Segmentation and Feature Extraction for Large-Scale Digital Nephropathology

Abstract

Abstract BACKGROUND AND AIMS Nephropathology is essential for the diagnosis of kidney diseases. Deep learning-based image analyses, including segmentation of kidney histology, open new possibilities for reproducible quantitative precision pathology. Current segmentation approaches in kidney histology focused on specific use cases. Here, we developed a framework for automated segmentation and quantification of a wide spectrum of non-neoplastic kidney diseases. METHOD We trained U-Net Convolutional Neural Networks (CNNs) for two streamlined tasks: (i) detection of kidney tissue and (ii) instance segmentation of relevant histological structures, i.e. glomeruli, tubules and arteries. These were used to segment 1103 Whole-Slide-Images (WSIs) from four cohorts, including two external datasets for multi-center validation. In total, 35 features from 51 445 glomeruli, 4016 792 tubules and 362 471 arteries were extracted, generating 30 million morphometric data points. Morphometry data were associated with clinical and pathology data. RESULTS Kidney tissue was precisely outlined by the tissue segmentation CNN. The structure segmentation CNN for histology provided accurate results on WSI-level despite difficult histopathological morphologies such as crescents, segmental sclerosis, tubular casts or arteriosclerosis. Quantitative analyses of morphological alterations of glomeruli revealed, for example, an increased glomerular area in membranous glomerulonephritis and all cases characterized with nephrotic range proteinuria independent of the underlying disease, and decrease in glomerular tuft circularity in pauci-immune glomerulonephritis, especially in cases with severe loss of kidney function. Principal component analysis of multiple glomerular, tubular and interstitial features stratified based on estimated glomerular filtration rate identified CKD-relevant morphological alterations of glomeruli and the tubulointerstitium. Image extraction of morphometric outliers enabled fast-track visual assessment of severe lesions. CONCLUSION Segmentation and large-scale quantitative feature extraction enable reproducible quantitative analysis of kidney morphology, opening new possibilities for digital precision nephropathology.