Loading…
Renal pathology foundation model whose learning unit is a whole detected glomerulus, pretrained on over a million of them across four biopsy stains.
No providers recorded yet. Browse all providers
A renal biopsy is read one glomerulus at a time. The pathologist locates each glomerular tuft in the section and scores it for mesangial hypercellularity, segmental sclerosis, crescents, and a dozen other lesions. Computational pathology models, by contrast, are almost always pretrained on square tiles cut from a whole-slide image on a fixed grid, so one training sample may contain half a glomerulus, two of them, or none at all. The diagnostic entity and the learning unit do not line up, and what the model represents is tissue in general rather than the structure the diagnosis rests on.
GloPath removes that mismatch by making the glomerulus itself the training sample. Every biopsy slide is passed through a glomerular detector, each detected glomerulus is cropped at 20× magnification (≈0.42 µm/pixel), and self-supervised pretraining runs over those entities rather than over grid tiles. The model was developed by the Institute of Biopharmaceutical and Health Engineering at 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.
The paper isolates the effect of that design choice with a controlled counterpart, RenalPath, built from the same ViT-Base backbone and the identical pretraining recipe on the same 14,049 biopsy slides — but sampling conventional 224×224 patches instead of detected entities. RenalPath beats the general-purpose pathology models it is compared against; GloPath beats RenalPath. The gain therefore comes from aligning the pretraining unit with the pathological entity, not from architecture or scale.
GloPath is a ViT-Base encoder — 12 transformer layers, 86.8 million parameters — trained with the DINO teacher–student objective, the teacher updated by exponential moving average (momentum 0.996, temperature 0.04). Pretraining used AdamW at learning rate 5e-4 with a 10-epoch warm-up, batch size 64 per GPU, and 200 epochs on eight A100 GPUs. Across 52 downstream tasks spanning three cohorts, it led on 42 (80.8%), with mean F1 of 0.938 on PAS, 0.965 on Masson's trichrome, and 0.951 on PASM lesion recognition, and 88.66% ROC-AUC for immunofluorescence deposition-region classification. On XJ-CLI, an unfiltered stream of 13,749 glomeruli from routine clinical practice, mean ROC-AUC was 0.913. It reaches this with roughly a quarter of the parameters of UNI, the ViT-Large baseline it is measured against, and well under the pretraining volume of UNI, CONCH, or PLIP. For Bowman's capsule and tuft segmentation — annotations seeded with Glo-In-One-v2 and refined by hand — the frozen backbone approaches the fully supervised nnU-Net upper bound of 0.851 and 0.905 IoU.
The immediate use is nephropathology triage and quantification: recognizing and grading glomerular lesions across the four routine stains, classifying immune-deposit patterns on immunofluorescence, and adapting to a new cohort or lesion definition from a handful of labeled examples. Beyond diagnosis, the segmentation head turns each glomerulus into a vector of morphological measurements, which the authors correlated against clinical variables — finding significant morphology–clinical associations in 37 of 98 tested pairs internally and 43 of 126 in a Kidney Precision Medicine Project cohort, including links between tuft geometry and serum creatinine that survive adjustment for sex, age, and disease type.
GloPath is evidence that in a domain organized around a discrete anatomical unit, matching the pretraining sample to that unit buys more than scaling a general-purpose encoder. The caveats are real: pretraining and most internal evaluation data come from a single tertiary hospital and are available only by research proposal, rare glomerulonephritis subtypes are underrepresented, and scanner harmonization was not explored. The public repository ships the architecture, the DINO pretraining scripts, and an inference notebook, but no checkpoint file; the authors state that most weights cannot be distributed because pretraining used private patient data and offer them on request. Reproducing the reported numbers therefore depends on that request path, not on a downloadable release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.