UNI

Chen et al., "A General-Purpose Self-Supervised Model for Computational Pathology", Nature Medicine, 2024. DOI: 10.1038/s41591-024-02857-3

What they do

UNI uses offline CLAM-based tissue segmentation followed by non-overlapping patch extraction at 20× magnification (0.5 µm/px).

• Patch sizes: 256×256 (75.8M patches) and 512×512 (24.3M patches) at 20× (0.5 µm/px)
• Tissue detection: CLAM toolkit — saturation-channel thresholding in HSV, median blur, morphological closing, contour filtering with area thresholds. Exact CLAM parameters are not specified in the paper.
• Sampling: non-overlapping tissue patches extracted offline; "sampling approximately 800 histology tissue patches per WSI in Mass-100K"; 100,130,900 patches from 100,426 WSIs
• Normalization: ImageNet — mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)
• DINOv2 multi-crop: extracted tiles (256 or 512) are fed to DINOv2 using the "default configuration." Specific crop sizes, scale ranges, and counts are not stated in the paper. Values shown in the approximation (global 224, local 96, scale 0.32–1.0 / 0.05–0.32, 2+8 crops) are DINOv2 defaults inferred from ViT-L/16 (96 = 6×16).
• Architecture: ViT-L/16 (0.3B params)
• Training: DINOv2 (DINO self-distillation, iBOT masked-image modeling, KoLeo regularization), 125K iterations, batch size 3072, fp16 with PyTorch-FSDP on 32× A100 GPUs
• Data: Mass-100K — 100,130,900 patches from 100,426 H&E WSIs across 20 tissue types (MGH, BWH, GTEx; no TCGA)

wsistream approximation

UNI extracts all non-overlapping tissue patches at 20× offline. This is fundamentally a GridSampler workflow. However, GridSampler yields patches in fixed scan order, so with a finite patches_per_slide it always selects the same top-left-biased subset. RandomSampler avoids that spatial bias but samples with replacement and does not guarantee full coverage. Neither is a perfect online substitute for offline full-grid extraction.

from wsistream.pipeline import PatchPipeline
from wsistream.backends import OpenSlideBackend
from wsistream.tissue import CLAMTissueDetector
from wsistream.sampling import RandomSampler

pipeline = PatchPipeline(
    slide_paths=slide_paths,
    backend=OpenSlideBackend(),
    tissue_detector=CLAMTissueDetector(),
    sampler=RandomSampler(
        patch_size=256,  # 256x256 at 20x (paper also uses 512x512 for some slides)
        num_patches=-1,
        target_mpp=0.5,  # 20x magnification (0.5 µm/px)
    ),
    pool_size=8,
    patches_per_slide=100,
    cycle=True,
)

With multi-crop views

UNI feeds 256×256 tiles to DINOv2's multi-crop using the "default configuration". The paper does not enumerate crop sizes, scale ranges, or counts. The configuration below uses DINOv2 defaults; local size 96px is inferred from ViT-L/16's 16px patch stride (6 × 16 = 96).

from wsistream.views import ViewConfig, RandomResizedCrop

pipeline = PatchPipeline(
    slide_paths=slide_paths,
    backend=OpenSlideBackend(),
    tissue_detector=CLAMTissueDetector(),
    sampler=RandomSampler(patch_size=256, num_patches=-1, target_mpp=0.5),
    views=[
        ViewConfig(
            name="global",
            crop=RandomResizedCrop(size=224, scale=(0.32, 1.0)),
            count=2,  # global_0, global_1 — DINOv2 default: 2 global crops
        ),
        ViewConfig(
            name="local",
            crop=RandomResizedCrop(size=96, scale=(0.05, 0.32)),  # 6×16 for ViT-L/16
            count=8,  # local_0 … local_7 — DINOv2 default: 8 local crops
        ),
    ],
    pool_size=8,
    patches_per_slide=100,
    cycle=True,
)

Per-crop augmentations

UNI uses standard DINOv2 photometric augmentations (color jitter, Gaussian blur, grayscale, solarization, horizontal flip) applied per crop. To add them with view-asymmetric probabilities matching DINOv2 defaults, see the DINOv2-style multi-crop example.

Deviations from paper

Step	Paper	wsistream	Match
Tissue detection	CLAM (exact parameters not specified)	CLAMTissueDetector (wsistream defaults)	Approximate — same algorithm, but paper does not specify CLAM parameters to verify parity
Patch sizes	256×256 and 512×512	patch_size=256 only	Partial — paper uses both sizes; wsistream uses one
Magnification	20× (0.5 µm/px)	target_mpp=0.5	Exact
Extraction	Offline non-overlapping grid (all tissue patches)	Online random sampling (with replacement)	Different — random sampling does not guarantee full coverage
Normalization	ImageNet mean/std	Training code	Exact (not part of wsistream)
DINOv2 crop sizes / scales	"Default configuration" — specific sizes/scales not enumerated in the paper	DINOv2 defaults (global 224, scale 0.32–1.0; local 96 = 6×16, scale 0.05–0.32)	Inferred
DINOv2 augmentations	Standard DINOv2 training objective documented in the model card; exact augmentation chain not enumerated in the paper text	Training code	Inferred

UNI2-h (Chen et al., 2025)

UNI2-h scales up the original UNI approach to 200M+ image tiles sampled from 350K+ diverse H&E and IHC slides (Mass General Brigham), using a ViT-H/14 with 8 register tokens (681M params). Training uses the same DINOv2 recipe (DINO + iBOT + KoLeo) with bf16 mixed-precision on A100 GPUs. Inference uses 224×224 input with ImageNet normalization, producing 1536-dimensional embeddings.

Detailed preprocessing methodology for UNI2-h (tissue detection parameters, patch extraction specifics, tissue thresholds) is not publicly documented at the time of writing. The key known differences from UNI are the larger model (ViT-H/14 vs ViT-L/16), the addition of IHC slides alongside H&E, register tokens, and the substantially larger training set (350K+ vs 100K WSIs). The same wsistream approximation above applies, with the caveat that UNI2-h's exact preprocessing details cannot be verified.