Histopathology foundation model pretrained with DINOv2 on tiles chosen by unsupervised hierarchical clustering over 350 million whole-slide tiles.
No providers recorded yet. Browse all providers
Vision foundation models for digital pathology are trained on tiles cropped from whole-slide images, and their quality depends heavily on which tiles they see. Until now, that selection has been made by pathologists working at the level of the whole slide — balancing cancer types, organs, or staining protocols — while the enormous variation within a slide has gone unexamined. Patho-SSL Curated ViT-L is the encoder that results from replacing this expert step with fully unsupervised curation at the tile level.
The work comes from ETH Zurich and EPFL with collaborators in Basel, at the Idiap Research Institute, and at the Icahn School of Medicine at Mount Sinai, and was presented at MICCAI 2025. The recipe builds a hierarchical clustering tree over embeddings of 350 million tiles, extracted with UNI, then samples a curated subset that covers that embedding space uniformly. A central finding is that curation alone is not enough: applied naively, it does not reliably improve the learned representations. The missing ingredient is how curated data is fed to the model, and the authors supply it in the form of stratified batches that draw equally from every top-level cluster.
The released encoder is used the way its peers are — as a frozen tile-level feature extractor whose embeddings feed linear probes on regions of interest and attention-based multiple-instance learning models on whole slides.
The pretraining corpus combines 11,666 whole-slide images from TCGA spanning 32 cancer types with 25,430 slides from GTEx spanning 40 healthy tissue sites, tiled at 20x magnification (0.5 µm/pixel) into roughly 350 million non-overlapping 224x224 pixel patches. Two four-level clustering trees were compared: T1, with cluster counts of {3.5M, 35k, 350, 62} from bottom to top, and T2, with {3.5M, 100k, 10k, 2048}. Curated subsets take 10% of the corpus. Training uses DINOv2 with a ViT-L backbone for 170,000 iterations at a global batch size of 2048, equal to one full pass over the uncurated data.
The best configuration — the narrower T1 tree with stratified batches — reaches 82.0% average balanced accuracy across the region-level benchmarks (LUAD, BRACS, BreakHis, BACH, PCAM, CRC, UniToPatho, Chaoyang), against 79.9% for training on the full corpus with random batches and 80.8% for expert-supervised curation with the same stratified sampling. On the whole-slide benchmarks — breast cancer detection, ER, PR, HER2 and HRD status, lung EGFR mutation and immunotherapy response, and inflammatory bowel disease detection — it averages 83.3% AUC against 82.6% and 82.8%. The margins are widest on the hardest tasks: seven-subtype breast carcinoma classification on BRACS improves from 66.1% to 69.3%, and BACH from 85.0% to 91.2%.
The encoder slots directly into computational pathology pipelines that consume frozen tile embeddings: tumour subtyping and grading on regions of interest, and weakly supervised biomarker prediction from whole slides, including receptor status, homologous-recombination deficiency, and driver mutations from H&E morphology alone. The curation recipe is the more portable contribution — any group assembling a pretraining corpus of histology tiles can apply the clustering and batch-stratification strategy on top of an existing encoder's embedding space.
The study is the first fully automated data curation scheme for foundation model training in digital pathology, and its central negative result is as useful as its positive one: clustering-based curation transplanted from natural images does not transfer to pathology without a matching change to the data feeding strategy. The clustering assignments, centroids, and curation code are public, and the encoder achieves its gains while drawing smaller proportions of breast, colon, and lung tiles than the baselines it beats, suggesting the improvement comes from diversity rather than from alignment with the evaluation cohorts. Trained model weights are not distributed, so reproducing the encoder requires rerunning the DINOv2 pretraining from the released curation artifacts.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.