Kidney Week

Abstract: SA-PO0011

Interpretable Vision Transformer-Based Artificial Intelligence Models Predict Glomerulonephritis Subtypes from Multistain Histopathology Whole-Slide Images

Session Information

• Intelligent Imaging and Omics: Phenotyping and Risk Stratification
  November 08, 2025 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Vremenko, Dmytro, Harvard Medical School Department of Biomedical Informatics, Boston, Massachusetts, United States

Dmytro Vremenko

• Chang, David R., Harvard Medical School Department of Biomedical Informatics, Boston, Massachusetts, United States

David R. Chang, MD

• Kuo, Chin-Chi, China Medical University Hospital, Taichung, Taichung City, Taiwan

Chin-Chi Kuo, MD, PhD, MPH

• Yu, Kun-Hsing, Harvard Medical School Department of Biomedical Informatics, Boston, Massachusetts, United States

Kun-Hsing Yu, MD, PhD

Background

Accurate diagnosis of glomerulonephritis (GN) requires analyzing large, multi-stain whole slide images (WSIs) from kidney biopsies. The size and heterogeneity of this data pose challenges for current deep learning models, which are primarily optimized for cancer and H&E-stained images. To address this, we aim to develop a deep learning approach capable of integrating multiple stains to perform patient-level GN subtype classification.

Methods

We analyzed 5,140 WSIs from 290 GN patients at China Medical University Hospital, Taiwan, covering minimal change disease (MCD), focal segmental glomerulosclerosis (FSGS), membranous nephropathy (MN), and IgA nephropathy (IgAN), across hematoxylin and eosin (H&E), periodic acid–Schiff (PAS), Masson trichrome (TRI), and Jones silver (SIL) stains. Tissue tiles (224×224 px at 10×) were extracted and encoded using the Virchow2 vision transformer. An attention-based multiple instance learning model was trained on patient-level tile sets. Data was split by patient (70% training, 10% validation, 20% testing) with 5-fold cross-validation. We trained stain-specific and multi-stain models, selected top performers by one-vs-rest AUC, ensembled them, and visualized model attention regions for interpretability.

Results

Multi-stain models outperformed single-stain models on most pairwise AUC comparisons (MCD vs FSGS: 0.802; MCD vs MN: 0.730, average MCD vs others: 0.744) (Table 1). The SIL stain performed best for MCD vs IgAN (0.744 AUC). Multi-stain model achieved the highest average one-vs-rest AUC (0.624), improving further with ensembling (0.660). Attention maps showed reliance on glomeruli and tubules.

Conclusion

GN subtype classification improves with multi-stain input, highlighting the need for multi-stain models. Our study establishes the feasibility of deep learning for GN diagnosis from histopathology WSIs and highlights the importance of training foundation models tailored to benign kidney pathology, beyond cancer and H&E.

Table 1. AUROC of Trained Models

Funding

• Other NIH Support

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025djq99ae8