Loading…
Five DINO and DINOv2 tile encoders pretrained on 29k open-access TCGA slides, sampled online at four magnifications instead of a fixed patch set.
No providers recorded yet. Browse all providers
Self-supervised pretraining on whole-slide images (WSIs) normally starts with a preprocessing job: crop the gigapixel slides into a fixed set of tiles and write them to disk. That set then constrains everything downstream — changing the tile size or the magnification means regenerating the entire corpus, and the storage overhead makes dynamic sampling impractical. The answer from kaiko.ai is Online Patching, a library that exposes a whole slide archive as a single Zarr data source through a virtual in-memory filesystem and crops tiles at arbitrary coordinates, sizes and magnifications while training runs. Because patches are drawn from random positions rather than a precomputed grid, nearly every tile the model sees is new — a single 10^5 x 10^5 pixel slide admits on the order of 10^10 distinct crops.
Built on that pipeline, kaiko.ai released five pathology tile encoders in 2024: four trained with DINO (ViT-S/16, ViT-S/8, ViT-B/16 and ViT-B/8) and one with DINOv2 (ViT-L/14). All were pretrained on roughly 29,000 open-access H&E slides from The Cancer Genome Atlas covering 32 cancer types, and all are used as frozen feature extractors — a downstream head is fit on the embeddings while the backbone stays fixed.
The release sits between early TCGA-only encoders such as Phikon, trained on 6,000 slides, and later models built on proprietary archives one to two orders of magnitude larger, including UNI, Virchow and Prov-GigaPath. It is a separate, earlier line of work from kaiko.ai's Midnight models. Alongside the encoders the group published eva, the open evaluation framework that produced every number in the paper.
eva runs the full linear-probing protocol across tile-level classification and segmentation datasets, so a new encoder can be compared on identical splits.The backbones range from roughly 22M parameters (ViT-S) through 86M (ViT-B) to 304M for the DINOv2 ViT-L/14. Pretraining runs in two phases of 100 ImageNet-style epochs each, where an epoch is 1,280,000 patches: the first phase uses FFPE slides only, the second adds fresh-frozen slides at half the peak learning rate. Compute ranged from 4 H100 GPUs for most variants to 8 for ViT-B/8 and 16 for ViT-L/14, sampling 256x256 patches.
Evaluation is linear probing on frozen embeddings, reported as balanced accuracy over five runs, across BACH (breast subtyping), CRC (colorectal tissue typing), MHIST (colorectal polyps), PatchCamelyon (lymph node metastasis), a TCGA TP53 mutation-status task, and CoNSeP segmentation by DICE. ViT-B/8 records 0.865 / 0.956 / 0.809 / 0.921 / 0.659 / 0.741 across those six, and ViT-L/14 leads on BACH at 0.870; both beat the contemporaneous TCGA-trained Lunit (21k WSIs) and Phikon (6k WSIs) encoders on most of the set. A data-scaling study found that 1% of TCGA — 108 slides — already gives reasonable accuracy on the out-of-distribution tasks, while only the in-distribution TP53 task keeps improving with more slides, which the authors read as evidence that TCGA's diversity is exhausted well before its volume is.
The encoders drop into histopathology pipelines wherever a frozen tile-level feature extractor is needed: cancer subtyping and tissue classification, metastasis detection, biomarker prediction from morphology, and nuclear or tissue segmentation with a lightweight trainable decoder. Loading is a single torch.hub.load call against the public repository, and the /8 variants suit dense prediction rather than slide-level classification. The weights carry the Kaiko Non-Commercial Public License, which permits research use but rules out commercial deployment without a separate agreement.
The release gave the field a reusable evaluation harness at a moment when published comparisons used incompatible protocols, and eva has since benchmarked encoders well beyond kaiko.ai's own. In a later independent study of 19 pathology foundation models across 31 weakly supervised clinical tasks, the ViT-L/14 encoder placed mid-field on average AUROC, reflecting its small TCGA-only corpus, though segmentation-focused benchmarks rank it among the stronger encoders. The paper's own conclusion points the same way: limited returns from scaling model size on TCGA, and hospital-scale multi-institution data as the necessary next step — the direction kaiko.ai's later work took.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.