---
tags:
- model_hub_mixin
- pytorch_model_hub_mixin
- medical
license: apache-2.0
language:
- en
metrics:
- Dice
- Jaccard
- 95HD
- ASD
pipeline_tag: image-segmentation
library_name: pytorch
---

This model has been pushed to the Hub using the [PytorchModelHubMixin](https://huggingface.co/docs/huggingface_hub/package_reference/mixins#huggingface_hub.PyTorchModelHubMixin) integration:
- **Hugging Face Space** (available now):  
  https://huggingface.co/spaces/Tournesol-Saturday/railNet-tooth-segmentation-in-CBCT-image
- Code: https://github.com/Tournesol-Saturday/RAIL
- Paper: [RAIL: Region-Aware Instructive Learning for Semi-Supervised Tooth Segmentation in CBCT](https://huggingface.co/papers/2505.03538)

**Steps to use our model in this repository:**
1. Clone this repository with the following command:
   ```bash
   git clone https://huggingface.co/Tournesol-Saturday/railNet-tooth-segmentation-in-CBCT-image
   cd railNet-tooth-segmentation-in-CBCT-image
   ```
3. Create a virtual environment to experience our model using the following command:
   ```python
   conda create -n railnet python=3.10
   conda activate railnet
   pip install -r requirements.txt
   python gradio_app.py
   ```
4. In the current working directory, find the `example_input_file` folder.  
   Select an arbitrary `.h5` file in this folder and drag it into the `Gradio` interface for model inference.
5. Waiting for about 1min~2min30s, the model inference is completed and the segmentation result and 3D rendering visualization will be produced.  
   Both the original image and the segmentation result are saved in `.nii.gz` format in the `output` folder of the same directory.
6. Since `Gradio` performs 1/2 downsampling on the 3D segmentation visualization, the segmentation accuracy is degraded.  
   Users can drag the `.nii.gz` format files in the `output` folder into the `ITK-SNAP` software to view the accurate segmentation visualization.