PathoJev / README.md
Mikezcy's picture
Add arXiv link and citation
da08197 verified
|
Raw History Blame Contribute Delete
3.75 kB
---
license: cc-by-nc-4.0
base_model:
- Qwen/Qwen3.5-9B
- autotrust/JEV-9B
- autotrust/JEV-27B-VL
language:
- en
tags:
- pathology
- histopathology
- whole-slide-image
- vision-language
- system-one
- medical
extra_gated_prompt: >-
PathoJev is released for non-commercial research use under CC BY-NC 4.0. It is not a medical device and must not be
used for diagnosis or clinical decision-making. Access requests are reviewed manually.
extra_gated_fields:
Full name: text
Affiliation: text
Country: country
Intended use: text
I will use PathoJev for non-commercial research only: checkbox
I will not use PathoJev for clinical decision-making: checkbox
---
# PathoJev
PathoJev answers pathology questions directly as probabilities over the answer options, for whole-slide images
(WSIs) and regions of interest (ROIs). It builds on the System One model JEV with a Qwen3.5 hybrid backbone. A WSI
enters as a 32 x 32 grid of pooled CONCH v1.5 patch features and is encoded once; every question is then read out as an
isolated row on the shared slide context by a typed decision head, without generating text. ROI images go through the
backbone's own vision tower.
Paper: [arXiv:2610.40752](https://arxiv.org/abs/2610.40752)
Code: https://github.com/ChyaZhang/PathoJev
## Files
| Path | Content |
|---|---|
| `9b/` | PathoJev-9B: LoRA adapter, decision head and slide projection (`pathojev.safetensors`) and `config.json`; backbone [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) |
| `27b/` | PathoJev-27B, same layout; backbone [autotrust/JEV-27B-VL](https://huggingface.co/autotrust/JEV-27B-VL) |
| `pathojev/`, `demo.py` | minimal inference code (same as the GitHub repository) |
Access is gated: request it on this page, then log in with `hf auth login`. The backbone is downloaded on first use. PathoJev-9B needs about 20 GB of GPU memory and PathoJev-27B about 60 GB.
## Usage
```bash
hf download Mikezcy/PathoJev --exclude "*.safetensors" --local-dir PathoJev && cd PathoJev
pip install -r requirements.txt
```
```python
from pathojev import PathoJev, make_grid
model = PathoJev.from_pretrained("9b") # or "27b"; weights are fetched from this repository
grid = make_grid("slide.h5") # TRIDENT CONCH v1.5 features (20x, 512 x 512 px patches)
questions = [
{"question": "What is the histologic grade of the tumor?", "options": ["Low-grade", "High-grade"]},
{"question": "Lymphovascular invasion is present.", "kind": "noul"}, # true / false statement
]
print(model.answer_slide(grid, questions)) # [{"answer": ..., "probs": {option: probability}}, ...]
```
`model.answer_roi(images, questions)` answers one question per ROI image. `demo.py` runs the eight SlideBench-TCGA
questions of slide TCGA-XF-AAMJ; see the GitHub repository for the expected output.
## Results
Accuracy (%):
| Model | WSI-VQA | WSI-Bench | SlideBench-TCGA | PathMMU-tiny | PathVQA | Quilt-VQA | OmniMedVQA | MedXpertQA |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| PathoJev-9B | 65.4 | 90.4 | 75.9 | 71.9 | 94.6 | 70.8 | 75.3 | 25.6 |
| PathoJev-27B | 66.4 | 91.1 | 76.9 | 73.8 | 95.1 | 67.3 | 69.9 | 50.0 |
## Citation
```bibtex
@article{zhang2026pathojev,
title = {PathoJev: Efficient Pathology Decisions from Shared Visual Context},
author = {Zhang, Chengyang and Zhang, Wenchuan and Li, Bo and Liu, Xinyu and Yang, Jiaming and Li, Mengran and Chen, Jie and Ju, Wei and Yi, Yuhao and Bu, Hong and Lv, Jiancheng},
journal = {arXiv preprint arXiv:2610.40752},
year = {2026}
}
```
## License
CC BY-NC 4.0, for non-commercial research use. CONCH v1.5 and pre-extracted TCGA features have their own licenses.
PathoJev is not a medical device and must not be used for clinical decision-making.