Abstract: TH-PO0734
Deep Learning-Based Classification of Proliferative vs. Nonproliferative Glomerular Lesions in Kidney Biopsies of Patients with Lupus Nephritis
Session Information
- Glomerular Innovations: Artificial Intelligence, Multiomics, and Biomarkers
November 06, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Glomerular Diseases
- 1402 Glomerular Diseases: Clinical, Outcomes, and Therapeutics
Authors
- Daouk, Mohammad, University of Houston, Houston, Texas, United States
- Becker, Jan U., Universitatsklinikum Koln Klinische Infektiologie, Cologne, NRW, Germany
- Kambham, Neeraja, Stanford University, Stanford, California, United States
- Chang, Anthony, The University of Chicago, Chicago, Illinois, United States
- Nguyen, Hien V, University of Houston, Houston, Texas, United States
- Mohan, Chandra, University of Houston, Houston, Texas, United States
Background
Automatic classification of glomerular lesions by machine learning (LN), could improve standardization in renal biopsy assessment. We developed and evaluated a comprehensive, uncertainty aware deep learning binary classifier to distinguish ISN/RPS proliferative versus non-proliferative LN.
Methods
We assembled 9 674 glomerular images (7 767 non-proliferative, 1 907 proliferative) from 363 H&E, PAS, trichrome and silver WSIs across three centers. Patches were extracted at three resolutions and z-normalized using ImageNet statistics. We trained convolutional backbones (ResNet-18/34/50/101/152; DenseNet-121/169/201/264d; EfficientNet-B0–B7; RegNetY008/016; ResNeXt101_32x8d) and a ViT-Base-Patch16-224 Transformer, all initialized on ImageNet. To quantify uncertainty and flag ambiguous cases for expert review, we applied Monte Carlo dropout with 50 inference passes. We held out 15 % of patches for final testing; the remaining 85 % underwent five-fold cross-validation (shuffled at each fold: 80 % training, 20 % validation).
Results
On the hold out set, the best ensemble achieved an accuracy of 0.94, an area under the ROC curve (AUC) of 0.97 and a mean uncertainty score of 0.02, for discriminating proliferative from non-proliferative LN. Uncertainty estimates clustered near decision boundaries and inversely correlated with predictive confidence, effectively highlighting cases for expert review.
Conclusion
Our multi center, multi stain deep learning framework delivers highly accurate classification of proliferative LN, with well calibrated uncertainty estimates in a binary LN glomerular lesion analysis. By providing rapid, uncertainty aware lesion classification in LN biopsies, this tool has the potential to standardize interpretation, and reduce inter observer variability in renal diagnostics of LN.
Funding
- Other NIH Support