Development of a deep learning model for glomerulus detection and segmentation in kidney histopathology images
Journal:
medRxiv
Published Date:
Oct 22, 2026
Abstract
Background: Estimation of nephron number, an indicator of renal reserve, requires quantitative assessment of glomerular number and size in kidney biopsy specimens. We aimed to develop an automated system for quantitative measurement of glomerular areas from clinical kidney biopsy images, across diverse kidney diseases. Methods: We developed a two-stage deep learning system combining Faster R-CNN for glomerular detection in whole-slide images (WSIs) and DeepLabV3+ for segmentation. Of 114 kidney biopsies (228 WSIs), 104 were used for model development and 10 for independent final evaluation. Agreement was assessed using the F1-score for detection, the Dice coefficient for segmentation, and the intraclass correlation coefficient (ICC) for glomerular area measurement. Analysis was performed at 40x total magnification (2.331 m/pixel) and validated against 400x images. Results: In the independent evaluation set, the artificial intelligence (AI) model achieved a mean F1-score of 0.851 for detection, compared with 0.888 among experts. For segmentation, the Dice coefficient was 0.959 {+/-} 0.011 and the per-pair ICC (2,1) for area measurement was 0.989 {+/-} 0.009; both were comparable to inter-expert agreement. Bland-Altman analysis showed a mean bias of -344 m2 (95% limits of agreement, -1904 to 1216 m2) between 40x and 400x measurements. Conclusion: The two-stage model achieved glomerular segmentation and area measurement comparable to expert agreement in independent clinical data, with detection performance approaching inter-expert agreement, and provided reliable measurements at low magnification. This automated approach may facilitate large-scale quantitative kidney pathology and nephron number estimation from clinical kidney biopsies.
