RenalPath

Two things change at once when a pathology foundation model is specialized to an organ: it sees domain-specific tissue, and its developers often redefine what a training sample is. RenalPath holds the second variable still. It is a ViT-Base encoder pretrained with a self-supervised objective on 1,017,879 conventional 224×224 patches cut at 0.42 µm/pixel from 14,049 renal biopsy whole-slide images — patches that fall on glomeruli, on interstitium, and on everything else a kidney biopsy contains, sampled the way pathology encoders are normally sampled.

That corpus and that recipe are shared exactly with GloPath, which was trained on the same slides with the same architecture and the same number of samples, but whose samples are individually detected glomeruli rather than grid tiles. The paper states the constraint plainly: the only difference between the two models is the definition of the pretraining unit and the organization of the data. RenalPath is therefore the conventional arm of a controlled experiment about what a pretraining unit is worth — and, separately, a usable renal-domain encoder in its own right, since nothing about it depends on a glomerular detector running first.

It was developed at the Institute of Biopharmaceutical and Health Engineering of Tsinghua University's Shenzhen International Graduate School with Fuzhou University, the Department of Pathology at Xijing Hospital (Fourth Military Medical University), and Oxford, and published in Advanced Science in 2026.

#Key Features

• Whole-biopsy coverage: patches span glomeruli, interstitial regions, and other renal structures across H&E, PAS, Masson's trichrome, and PASM staining, so the representation reflects the heterogeneity of a renal biopsy rather than one compartment of it.
• Detector-free featurization: because the learning unit is a grid tile, a slide can be embedded end to end without an upstream entity-detection step.
• Frozen-feature transfer: the CLS token is extracted and held fixed while SVM, logistic regression, MLP, random forest, or prototype heads are fitted on top, across labeled set sizes from one example per class up to 100.
• Matched-sample design: the patch count is set equal to GloPath's glomerulus count, so the comparison between the two isolates the pretraining unit from data volume, backbone, and objective.
• Segmentation backbone: paired with a Segmenter decoder, the frozen encoder delineates Bowman's capsule and the glomerular tuft for downstream morphometry.

#Technical Details

RenalPath is a ViT-Base — 12 transformer layers, 86.8 million parameters — trained with the DINO teacher–student objective, the same recipe used for GloPath. Its pretraining corpus, XJ-Light-2, comprises Xijing Hospital chronic kidney disease cases archived between 2022 and 2023; the downstream cohort XJ-Light-1 covers 2019–2021 and was built from strictly non-overlapping patients, so evaluation data is disjoint from pretraining data at the patient level. Across the paper's 27 lesion-recognition tasks, RenalPath placed behind UNI (p = 0.0362) and ahead of PLIP and CONCH, and on lesion grading it outperformed the general-purpose pathology models consistently — a result the authors read as evidence that large-scale renal pretraining pays off on its own. GloPath's remaining margin over it is comparatively narrow: 4.62%, 3.11%, and 3.57% on PAS, MT, and PASM lesion recognition, 1.80% on lesion grading, 2.86% and 3.14% on immunofluorescence deposition region and pattern classification, and 1.22–1.23% IoU on tuft segmentation.

#Applications

The tasks it was evaluated on are the working tasks of nephropathology: recognizing and grading glomerular lesions on the four routine stains, classifying immune-deposit region and pattern on immunofluorescence, and few-shot adaptation to external cohorts — the Spanish and Lithuanian AIDPATH-G collection and the US Kidney Precision Medicine Project. It was also one of three encoders, with GloPath and UNI, carried into the paper's clinicopathological correlation analysis, where segmentation-derived glomerular morphology was tested against clinical variables.

#Impact

RenalPath's value is as a measurement. It establishes that a ViT-Base pretrained on a single hospital's renal biopsies beats general-purpose pathology encoders trained on far larger and more diverse corpora, and it bounds how much of GloPath's advantage is attributable to entity-aligned pretraining rather than to renal data or model scale. Its practical availability is limited: no RenalPath checkpoint has been released, and nothing in the shared repository — which carries the DINO pretraining scripts, the architecture, and inference notebooks — is labelled RenalPath or configured for it, so reproducing it means re-running the published pipeline over patch-sampled slides. That code carries no license of the authors' own, and the pretraining slides are private Xijing Hospital material available only by research proposal.