Abstract: SA-PO0946
Prediction of Kidney Graft Inflammation Using Self-Supervised Learning
Session Information
- Pathology: Updates and Insights
November 08, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1800 Pathology and Lab Medicine
Authors
- Naglah, Ahmed, University of Florida, Gainesville, Florida, United States
- Jen, Kuang-Yu, University of California Davis, Davis, California, United States
- Rosenberg, Avi Z., The Johns Hopkins University School of Medicine, Baltimore, Maryland, United States
- Hodgin, Jeffrey B., University of Michigan, Ann Arbor, Michigan, United States
- Tomaszewski, John E., University at Buffalo, Buffalo, New York, United States
- Alquadan, Kawther, University of Florida, Gainesville, Florida, United States
- Clapp, William L., University of Florida, Gainesville, Florida, United States
- Sarder, Pinaki, University of Florida, Gainesville, Florida, United States
Background
Tissue inflammation is a key indicator of kidney transplant rejection. While pathology foundation models showed recent progress in computational tasks, their usability has been limited to routine H&E stained slides trained mostly on cancer data. In this work, we study how to expand these models to other stains using self-supervised learning aiming to improve the prediction of graft inflammation in indication biopsy.
Methods
We collected multi-modal brightfield images of indication biopsies, clinical history, and pathological evaluation of rejection factors (860 WSI from n=172 participants) with a wide spectrum of immune activity (interstitial inflammation distribution of i-0: 44, i-1: 61, i-2: 38, and i-3: 29 participants). To learn domain-relevant features from different histology stains (Fig.1C), we implemented a contrastive self-supervised learning framework based on the student–teacher paradigm. We used prov-gigapath tile encoder for network architecture and parameter initialization. We used cosine contrastive loss and exponential moving average of the student weights during the training phase. We fine-tuned the model on the prediction of interstitial inflammation from CD3 stained tiles.
Results
Our proposed method on interstitial inflammation prediction achieved an F-score of 0.76 compared to 0.51 for DINOv2 as a baseline. We also performed correlation between the Banff scores and the pathological assessment of rejection which shows risk of rejection can be linearly projected given these scores (Fig.1B).
Conclusion
We proposed a self-supervised learning approach to adapt foundation models to different stains to enhance the prediction graft inflammation in indication biopsy. Our proposed method can enable better risk stratifications for kidney patients towards better post-transplant management (Fig.1A).
Fig. 1 (A) Proposed framework. (B) Banff scores vs. findings correlation. (C) Proposed self-supervised learning concept.
Funding
- NIDDK Support