TinyPath

TinyPath is a 103.6M-parameter Vision Transformer for H&E histopathology tiles. It was pruned from Prov-GigaPath (1.13B parameters) and then distilled from it.

• Architecture: ViT with 33 blocks, width 640, 10 heads, patch size 16 and a SwiGLU MLP
• Input: 224 × 224 RGB tiles
• Output: a 640-d CLS embedding and 196 patch tokens
• Teacher: Prov-GigaPath

Evaluation

TinyPath and Prov-GigaPath were run through the same evaluation pipeline, and the other three models' scores are published. H0-mini's come from the HEST and PLISM leaderboards and, for EVA, from its paper. H-optimus-0's come from the HEST, EVA 0.4.5 and PLISM leaderboards, and DistillPath-KS16-HOpt0's from its paper. Error bars are ±1 SD over repeated runs (±1 SEM over folds for PORPOISE); the published scores have none.

• Higher is better on every benchmark: Pearson r for HEST, balanced accuracy for the EVA classification tasks, Dice for EVA MoNuSAC, the leaderboard score and its components for PLISM, c-index for PORPOISE and AUROC for GTEx.
• A missing bar or a dash means that model has no published score on that benchmark.
• PLISM measures robustness to scanner and stain changes rather than accuracy. It is run two ways: on its 224 px tiles unresized, the setting of its public leaderboard and of H0-mini's own preprocessing, and with tiles resized to 256 px and center-cropped to 224, as PLISM's code now does for Prov-GigaPath. The PLISM figure uses 224 px, the only setting with scores for every model.
• H0-mini is Bioptimus's 86M-parameter ViT-B, distilled from H-optimus-0 (1.1B parameters) on the same 6,093 TCGA slides as TinyPath. All its scores here use the class token, like TinyPath's. Its EVA scores come from its paper's table, which has no BreakHis or Gleason. The paper's HEST score (0.4044) concatenates the class token with the mean patch token, so the HEST leaderboard's class-token score is used instead.
• On HEST, H0-mini scores higher than TinyPath but keeps less of its teacher: 95.4% of H-optimus-0's score, where TinyPath keeps 98.6% of Prov-GigaPath's. Under its paper's protocol it keeps 95.7% (0.4044 against H-optimus-0's 0.4224). The two students differ in more than their teachers, though: H0-mini is a ViT-B with 14 px patches and four register tokens, does not start from its teacher's weights, distills through its teacher's pretrained projection heads, which Prov-GigaPath does not ship, and saw about 215M tiles, 2.7 times TinyPath's 78.9M. Retention alone cannot say which of these differences matters.
• DistillPath-KS16-HOpt0 is kaiko's 22M-parameter ViT-S/16, further distilled from H-optimus-0's class and patch tokens on 6,000 TCGA slides. Its paper uses the same EVA, HEST and PLISM protocols but reports PLISM retrieval at top-5 rather than top-10, so only its PLISM score and median cosine are listed.
• TinyPath and H0-mini were trained without labels on the same 6,093 TCGA slides, which include slides from 18 of MoNuSAC's 25 test patients and 443 of PORPOISE's 444 patients. Prov-GigaPath never saw TCGA.

Exact numbers

Benchmark	Metric (split)	Better	TinyPath	Prov-GigaPath	H0-mini (published)	H-optimus-0 (published)	DistillPath-KS16-HOpt0 (published)
HEST, mean of 9 tasks	Pearson r	Higher	0.3822	0.3875	0.3958	0.4150	0.376
EVA PatchCamelyon	Balanced accuracy (test), 5 runs	Higher	0.9462 ± 0.0010	0.9436 ± 0.0015	0.942	0.943	0.921
EVA CRC	Balanced accuracy (val), 5 runs	Higher	0.9571 ± 0.0006	0.9519 ± 0.0014	0.961	0.955	0.943
EVA BACH	Balanced accuracy (val), 5 runs	Higher	0.7601 ± 0.0135	0.7637 ± 0.0025	0.774	0.759	0.742
EVA Gleason (Arvaniti)	Balanced accuracy (val), 5 runs	Higher	0.7346 ± 0.0092	0.7246 ± 0.0100	—	0.770	0.756
EVA BreakHis	Balanced accuracy (val), 5 runs	Higher	0.7875 ± 0.0098	0.8288 ± 0.0084	—	0.801	0.690
EVA Camelyon16-small	Balanced accuracy (test), 20 runs	Higher	0.8368 ± 0.0221	0.8036 ± 0.0308	0.842	0.827	—
EVA Camelyon16, 10,000 patches per slide	Balanced accuracy (test), 20 runs	Higher	0.9340 ± 0.0126	0.8897 ± 0.0316	—	—	—
EVA PatchCamelyon 10-shot	Balanced accuracy (test), 50 runs	Higher	0.8628 ± 0.0397	0.8477 ± 0.0492	—	—	—
EVA MoNuSAC	Dice (release test split), 5 runs	Higher	0.6443 ± 0.0065	0.6834 ± 0.0066	0.643	0.685	0.634
PLISM	Leaderboard score (8,139 positions, 256 px)	Higher	0.520	0.320	—	—	—
PLISM	Leaderboard score (8,139 positions, 224 px)	Higher	0.530	0.333	0.541	0.480	0.480
PORPOISE LUAD	Overall-survival c-index, fold mean ± SEM	Higher	0.5816 ± 0.0221	0.5534 ± 0.0235	—	—	—
GTEx findings (5 tasks)	Mean AUROC, ABMIL	Higher	0.7587	0.7699	—	—	—
GTEx findings (5 tasks)	Mean AUROC, mean-pool logistic regression	Higher	0.7426	0.7349	—	—	—
GTEx findings (5 tasks)	Mean AUROC, slide linear probe	Higher	0.7501	0.7470	—	—	—
GTEx findings (5 tasks)	Mean AUROC, slide k-NN probe	Higher	0.7259	0.7321	—	—	—

HEST, Pearson r per task (higher is better):

Encoder	IDC	PRAD	PAAD	SKCM	COAD	READ	CCRCC	LUNG	LYMPH_IDC	Mean
TinyPath	0.5679	0.3759	0.4722	0.5474	0.2688	0.1672	0.2395	0.5454	0.2553	0.3822
Prov-GigaPath	0.5515	0.3698	0.4746	0.5618	0.2992	0.1960	0.2429	0.5411	0.2500	0.3875
H0-mini (published)	0.5862	0.3687	0.4919	0.6012	0.2494	0.1863	0.2670	0.5482	0.2629	0.3958
H-optimus-0 (published)	0.5976	0.3848	0.4911	0.6454	0.3086	0.2216	0.2676	0.5590	0.2591	0.4150
DistillPath-KS16-HOpt0 (published)	0.554	0.349	0.458	0.554	0.261	0.177	0.228	0.546	0.256	0.376

PLISM by component over 8,139 positions (higher is better in every column). TinyPath and Prov-GigaPath ran in float16:

Encoder	Input	Median cosine, all pairs	Top-10, scanner only	Top-10, staining only	Top-10, both differ	Leaderboard	Mean cosine
TinyPath	256 px	0.7460	0.9082	0.2681	0.1581	0.520	0.7192
Prov-GigaPath	256 px	0.5533	0.5659	0.1098	0.0489	0.320	0.5388
TinyPath	224 px	0.7556	0.9218	0.2737	0.1686	0.530	0.7283
Prov-GigaPath	224 px	0.5704	0.5920	0.1178	0.0539	0.333	0.5547
H0-mini (published)	224 px	0.800	0.864	0.318	0.183	0.541	—
H-optimus-0 (published)	224 px	0.685	0.744	0.327	0.166	0.480	—
DistillPath-KS16-HOpt0 (published)	224 px	0.656	—	—	—	0.480	—

Usage

pip install -r requirements.txt

Tiles are resized to 256 px on the shorter side (bicubic), center-cropped to 224 and normalized with ImageNet statistics. AutoImageProcessor and the timm transform do this for you.

transformers

import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

processor = AutoImageProcessor.from_pretrained("<org>/tinypath")
model = AutoModel.from_pretrained("<org>/tinypath", trust_remote_code=True).eval()

image = Image.open("patch.png").convert("RGB")
with torch.no_grad():
    outputs = model(processor(image, return_tensors="pt")["pixel_values"], pool_mode="cls")
embedding = outputs.pooler_output   # [1, 640]
tokens = outputs.last_hidden_state  # [1, 197, 640]

Without trust_remote_code, the same weights load into transformers' built-in TimmWrapperModel.

timm

import timm

model = timm.create_model("hf-hub:<org>/tinypath", pretrained=True).eval()
transform = timm.data.create_transform(**timm.data.resolve_data_config({}, model=model))
with torch.no_grad():
    embedding = model(transform(image).unsqueeze(0))  # [1, 640]

uint8 encoder

From a local copy of this repository:

from modeling_tinypath import build_encoder, eval_transform

encoder = build_encoder(".")
with torch.no_grad():
    embedding = encoder(eval_transform()(image).unsqueeze(0))  # [1, 640]

python run_inference.py checks that every interface loads all weights and returns identical outputs.