Abstract: SA-PO0043
Whole Slide Imaging (WSI)-Based Supervised Learning Differentiates Idiopathic Nodular Glomerulosclerosis from Diabetic Nephropathy
Session Information
- Top Trainee Posters - 3
October 24, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 01:00 PM - 01:06 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Ansari, Ishaq, University of Florida Division of Nephrology Hypertension & Renal Transplantation, Gainesville, Florida, United States
- Afsari, Fatemeh, University of Florida Division of Nephrology Hypertension & Renal Transplantation, Gainesville, Florida, United States
- Srivastava, Dhiraj, University of Florida Division of Nephrology Hypertension & Renal Transplantation, Gainesville, Florida, United States
- Reeves, Nathan, University of Florida Division of Nephrology Hypertension & Renal Transplantation, Gainesville, Florida, United States
- Rosenberg, Avi Z., Johns Hopkins University Department of Pathology, Baltimore, Maryland, United States
- Clapp, William L., University of Florida Department of Pathology Immunology and Laboratory Medicine, Gainesville, Florida, United States
- Phillips, Carrie L., Indiana University Department of Pathology and Laboratory Medicine, Indianapolis, Indiana, United States
- Sarder, Pinaki, University of Florida Division of Nephrology Hypertension & Renal Transplantation, Gainesville, Florida, United States
- Eadon, Michael T., Indiana University Department of Medical and Molecular Genetics, Indianapolis, Indiana, United States
Background
Idiopathic Nodular Glomerulosclerosis (ING) is a rare non-diabetic kidney disease characterized by nodular mesangial expansion, the prototypical histopathological feature of the Diabetic Nephropathy (DN), making morphologic distinction between the two highly challenging. Accurate differentiation is clinically essential, as misclassification carries significant consequences. We developed an AI-based framework that employs a lesion-specific lesion-aware feature extractor, followed by an attention-based aggregation classifier integrating embeddings across all glomeruli per biopsy to distinguish DN from ING.
Methods
Building on our previously developed HistoFormer-DKD model, we developed an attention-based multiple instance learning (MIL) classifier which takes glomerulus embeddings extracted from HistoFormer-DKD, learns attention weights for individual embeddings, and aggregates the embeddings into a patient-level representation. We validated this approach on 171 PAS-stained whole-slide images (149 DN, 22 ING) from 168 patients across two institutions, none of which were used during HistoFormer-DKD training.
Results
The model correctly identified 21 of 22 ING and 146 of 149 DN cases. Statistical tests showed that 603 of 768 embedding dimensions (78%) separated ING from DN, indicating the disease signal is broadly distributed. The dimensions with the strongest separation cleanly split DN and ING glomeruli to opposite ends of each dimension. These dimensions act as morphologic axes the model learned without disease labels.
Conclusion
This work shows that HistoFormer-DKD embeddings, combined with attention-based MIL, can separate ING from DN with high accuracy and without clinical metadata. The fact that the diagnostic signal is spread across most of the embedding space, rather than tied to a few features, suggests the model is capturing real morphologic differences rather than spurious correlations. Beyond this specific overlap, the same framework could be applied to other renal diseases that are histologically similar but clinically distinct.
Fig 1. (A) Violin plot of top discriminating dimensions, ING vs DN. (B) Differential lesion attention between DN and ING (correctly classified cases) (* p<0.05, ** p<0.01).