Abstract: TH-PO1067
Single-Stain vs. Multi-Stain Foundation Model Prediction of Kidney Biopsy Morphology Across CureGN Glomerular Diseases
Session Information
- Pathology and Lab Medicine
October 22, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1700 Pathology and Lab Medicine
Authors
- Lee, Tsung-Hua, Harvard Medical School Department of Biomedical Informatics, Boston, Massachusetts, United States
- Chang, David R., China Medical University Hospital Department of Internal Medicine, Taichung City, Taiwan
- Vremenko, Dmytro, Harvard Medical School Department of Biomedical Informatics, Boston, Massachusetts, United States
- Kuo, Chin-Chi, China Medical University Hospital Department of Internal Medicine, Taichung City, Taiwan
- Yu, Kun-Hsing, Harvard Medical School Department of Biomedical Informatics, Boston, Massachusetts, United States
Background
Kidney biopsy interpretation in glomerular disease integrates morphologic findings across routine stains. Although foundation models extract high-dimensional whole slide image features, whether single-stain inputs capture biopsy morphology or multi-stain integration improves automated phenotyping remains unclear. We compared Virchow2-derived hematoxylin and eosin (H&E), periodic acid-Schiff (PAS), silver (SIL), and trichrome (TRI) features for predicting pathology morphology across CureGN glomerular diseases.
Methods
We analyzed CureGN kidney biopsy whole slide images from patients with minimal change disease (MCD), focal segmental glomerulosclerosis (FSGS), membranous nephropathy (MN), and IgA nephropathy (IgAN) with core annotations for mesangial hypercellularity, endocapillary hypercellularity, segmental sclerosis, interstitial fibrosis/tubular atrophy, and crescents. Virchow2 was used as a frozen feature extractor. For each component, we trained single-stain, pooled multi-stain, and ensemble models. Performance was evaluated by held-out area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, and specificity.
Results
The cohort included 220 MCD, 252 FSGS, 241 MN, and 376 IgAN patients, with 10,952 whole slide images: 3,810 H&E, 3,091 PAS, 1,738 SIL, and 2,313 TRI. Lesion prevalence was M1 21%, E1 18%, S1 52%, T1/T2 20%, and C1/C2 18%. Best single stains were SIL for M (AUROC 0.701) and H&E for E, S, T, and C (0.750, 0.782, 0.923, and 0.723). Pooled four-stain models achieved AUROCs of 0.680, 0.720, 0.767, 0.913, and 0.697 for M, E, S, T, and C. Ensemble models achieved AUROCs of 0.696, 0.754, 0.796, 0.931, and 0.736, improving 4 of 5 components over the best single-stain model. Balanced accuracy ranged from 0.410 to 0.718.
Conclusion
In a large CureGN WSI cohort, Virchow2 features predicted MEST-C morphology across glomerular diseases, with particularly strong discrimination for sclerosis and cinterstitial fibrosis/tubular atrophy. H&E alone captured substantial morphologic signal and outperformed simple feature pooling, whereas stain-level ensembles improved 4 of 5 components, supporting complementary information across routine stains. These findings establish a benchmark for foundation model-based kidney biopsy phenotyping and motivate stain-aware fusion architectures and multi-stain pathology foundation models.
Acknowledgment
We thank the CureGN Pathology Committee and Digital Pathology Working Group for supporting access to pathology annotations and whole slide image resources.
Funding
- Other NIH Support