---
license: mit
library_name: pytorch
tags:
  - medical-imaging
  - ct
  - trauma-detection
  - abdominal-trauma
  - dinov3
  - vision-transformer
  - pytorch
pipeline_tag: image-classification
---

<div align="center">

# TraumaNet DINOv3 ViT-Large Backbone

[![Checkpoint Repo](https://img.shields.io/badge/Hugging%20Face-Checkpoint-orange?logo=huggingface)](https://huggingface.co/frankzhang/Traumanet_ViT_DINOv3)
[![Source Code](https://img.shields.io/badge/GitHub-Source%20Code-black?logo=github)](https://github.com/FrankZhangRp/TraumaNet)
[![PyTorch](https://img.shields.io/badge/PyTorch-2.x-EE4C2C?logo=pytorch&logoColor=white)](https://github.com/FrankZhangRp/TraumaNet)
[![License](https://img.shields.io/badge/License-MIT-green.svg)](https://github.com/FrankZhangRp/TraumaNet)

</div>

---

## Overview

This repository hosts the **TraumaNet DINOv3 ViT-Large backbone checkpoint** used for downstream multi-task abdominal trauma detection on contrast-enhanced CT.

The checkpoint stored here is the pretrained backbone initialization used before downstream TraumaNet fine-tuning.

---

## Project Links

- **Hosted checkpoint repository:** https://huggingface.co/frankzhang/Traumanet_ViT_DINOv3
- **Source code repository:** https://github.com/FrankZhangRp/TraumaNet

---

## File

- `traumanet_dinov3_pretrain_backbone.pth`

---

## Intended Use

This checkpoint is intended to be used as the `dinov3_pretrained` initialization file in the TraumaNet downstream pipeline.

It is **not** a standalone end-to-end prediction package. To reproduce the downstream task, users should combine this checkpoint with the TraumaNet source code repository.

---

## Expected Downstream Setting

The downstream TraumaNet pipeline uses:

- contrast-enhanced abdominal CT
- HU soft-tissue windowing
  - window center = `40`
  - window width = `350`
- depth standardization to `240`
- 2.5D grouping with 3 adjacent slices per group
- DINOv3 ViT-Large backbone loading from this checkpoint

---

## Limitations

- This repository provides the backbone checkpoint only.
- It does not provide the trauma dataset.
- It does not provide train / validation / test labels.
- It does not provide external evaluation data.
- It does not provide a standalone inference API.

---

## Acknowledgments

We acknowledge the upstream **DINOv3** project:

- https://github.com/facebookresearch/dinov3

We also acknowledge the **RSNA 2023 Abdominal Trauma Detection AI Challenge** and its public challenge setting:

- https://www.rsna.org/rsnai/ai-image-challenge/abdominal-trauma-detection-ai-challenge
- https://www.kaggle.com/c/rsna-2023-abdominal-trauma-detection

---

## License

This checkpoint repository is released under the MIT License.