OpenPath — Checkpoints
Teacher checkpoints of OpenPath, a ViT-g/14 pathology foundation model pre-trained with self-supervision (DINOv2 + gram anchoring) on public-only whole-slide histopathology tiles (OpenPath corpus).
Headline result. On AMC-HCC-ST — a contamination-free in-house Asan Medical Center hepatocellular-carcinoma spatial-transcriptomics cohort, the least leakage-prone benchmark since no public foundation model was trained on it — OpenPath ranks #1 among seven foundation models.
Checkpoints
- 61 teacher checkpoints:
training_0…training_345000, every 5,750 iters (≈ 1 native epoch total). - Each is
training_<iter>/teacher_checkpoint.pth(ViT-g/14 reg4, 1536-dim CLS embedding). - Released model =
training_316250— selected by the clean AMC-HCC-ST benchmark. (OpenPath's HEST-1K peaks earlier, ~training_23000at ~0.38; pick a checkpoint to match your task.)
Load & extract embeddings
Requires the OpenPath / DINOv2 code (taejoon89/openpath).
import torch, dinov2.models.vision_transformer as vits
ck = torch.load("training_316250/teacher_checkpoint.pth", map_location="cpu", weights_only=False)
sd = {k[len("backbone."):]: v for k, v in ck["teacher"].items() if k.startswith("backbone.")}
m = vits.vit_giant2(patch_size=14, img_size=224, block_chunks=4, num_register_tokens=4,
ffn_layer="swiglufused", init_values=1e-5,
interpolate_antialias=True, interpolate_offset=0.0)
m.load_state_dict(sd, strict=True); m.eval()
cls = m.forward_features(x)["x_norm_clstoken"] # (B, 1536), ImageNet-normalized 224x224 input
Evaluation
Frozen-encoder linear/ridge probing, all models under one protocol (sorted by the clean AMC-HCC-ST cohort). AMC-HCC-ST is our headline because public benchmarks (HEST-1K, CRC, BACH) derive from repositories these FMs were pre-trained on and are confounded by train/test leakage.
| Model | AMC-HCC-ST (clean) ↓ | HEST-1K (public) | NCT-CRC-HE (9-cls acc) | BACH (4-cls acc) |
|---|---|---|---|---|
OpenPath (training_316250) | 0.323 | 0.372 | 0.954 | 0.761 |
| UNI2-h | 0.301 | 0.414 | 0.966 | 0.908 |
| OpenMidnight | 0.300 | 0.390 | 0.967 | 0.906 |
| Virchow2 | 0.292 | 0.398 | 0.964 | 0.875 |
| prov-gigapath | 0.286 | 0.393 | 0.953 | 0.752 |
| Phikon-v2 | 0.274 | 0.375 | 0.937 | 0.708 |
| UNI | 0.257 | 0.386 | 0.946 | 0.777 |
Intended use & limitations
Frozen feature extractor for H&E histopathology tiles (native ~40× / 0.5 µm-per-pixel, ImageNet normalization) → 1536-dim CLS embedding for linear/ridge probing, k-NN, MIL, retrieval. Not a medical device; not for diagnosis. Public-benchmark numbers are leakage-confounded; it is a patch-level encoder (slide-level context needs a separate aggregator). See the model / code card for details.
Related artifacts
| Artifact | Hugging Face repo | Notes |
|---|---|---|
| Corpus | taejoon89/openpath-corpus | Native 40× pathology tiles, 33,991 WebDataset shards / ~17 TB |
| Checkpoints | taejoon89/openpath-checkpoints | This repository |
| Code | taejoon89/openpath | training & evaluation code (also on GitHub) |
Citation
@misc{openpath2026,
title = {OpenPath: Public-Data Pathology Foundation Models and Leakage-Free Evaluation},
author = {Tae Joon Jun},
year = {2026},
note = {https://huggingface.co/taejoon89/openpath}
}
Acknowledgements
This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: HR21C0198); the Advanced GPU Utilization Support Program funded by the Government of the Republic of Korea, Ministry of Science and ICT; and the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (grant number: RS-2026-25522634).
License
Weights — Apache-2.0 (warm-started from Meta DINOv2 ViT-g/14-reg, itself Apache-2.0). Training data: public pathology datasets under CC-BY / CC0 / NIH-open terms.