---
language: en
---

# Refined RadImageNet - Conversion Tools

The RadImageNet dataset is available upon request at [https://www.radimagenet.com/](https://www.radimagenet.com/).

This repository provides tools to process the RadImageNet dataset, converting it into a refined and stratified organization suitable for various medical imaging applications.

For detailed information, refer to our preprint paper: [Policy Gradient-Driven Noise Mask](https://arxiv.org/abs/2406.14568).

If you use this code in your research, please cite our paper:

```bibtex
@inproceedings{yavuz2025policy,
  title={Policy Gradient-Driven Noise Mask},
  author={Yavuz, Mehmet Can and Yang, Yang},
  booktitle={International Conference on Pattern Recognition},
  pages={414--431},
  year={2025},
  organization={Springer}
}
```

## Performance Comparison of ResNet Models

This table compares the performance of ResNet models pretrained on 2D RadImageNet using regular and Two2Three convolution techniques across various metrics:

| Model     | Precision (macro) | Recall (macro) | F1 Score (macro) | Balanced Accuracy | Average Accuracy |
|-----------|-------------------|----------------|------------------|-------------------|------------------|
| ResNet10t | 0.4720            | 0.3848         | 0.3998           | 0.3848            | 0.7981           |
| ResNet18  | 0.5150            | 0.4383         | 0.4545           | 0.4383            | 0.8177           |
| ResNet50  | 0.5563            | 0.4934         | 0.5097           | 0.4934            | 0.8352           |

We recommend adapting the code for benchmarking other models, which can be found here: [https://github.com/pytorch/vision/tree/main/references/classification](https://github.com/pytorch/vision/tree/main/references/classification).

## ResNet Models and Weights

The model codes are shared through [https://github.com/convergedmachine/Refined-RadImagenet/](https://github.com/convergedmachine/Refined-RadImagenet/).

To create a model using the `timm` library:

```python
import timm
model = timm.create_model('resnet10t', num_classes=165)
```

Replace `'resnet10t'` with `'resnet18'` or `'resnet50'` as needed.

## Folder Structure

```
correction_masks/
data/
weights/
output/
source/
    correction_masks.tar.gz
    radimagenet.tar.gz
    RadiologyAI_test.csv
    RadiologyAI_train.csv
    RadiologyAI_val.csv
process.py
measure_acc_metrics.py
```

### Files & Directories

- **correction_masks/**: Contains correction masks for the images.
- **data/**: Contains the extracted radiology images.
- **weights/**: Directory for model weights.
- **output/**: Directory for output files.
- **source/**: Contains source files and datasets.
  - **correction_masks.tar.gz**: Compressed file containing correction masks.
  - **radimagenet.tar.gz**: Compressed RadImageNet dataset.
  - **RadiologyAI_test.csv**: CSV file for the test dataset.
  - **RadiologyAI_train.csv**: CSV file for the training dataset.
  - **RadiologyAI_val.csv**: CSV file for the validation dataset.
- **process.py**: Main script to process and organize the RadImageNet files.
- **measure_acc_metrics.py**: Script to measure accuracy metrics.

## Download Processing Files

This repository contains files from the Hugging Face repository `convergedmachine/Refined-RadImagenet`. Follow the instructions below to clone the repository using Git.

### Prerequisites

Ensure that Git LFS (Large File Storage) is installed:

```bash
git lfs install
```

### Cloning the Repository

To clone the entire repository to your local machine:

```bash
git clone https://huggingface.co/convergedmachine/Refined-RadImagenet source/
```

This command clones all files from the repository into a directory named `source`.

### Notes

- Ensure you have sufficient storage space for large files.
- For more information about this dataset, visit the [Github page](https://github.com/convergedmachine/Refined-RadImagenet/).

Feel free to contribute or raise issues if you encounter any problems.

## Usage

1. **Extract the Dataset**:

   ```bash
   python process.py
   ```

   Ensure the dataset tar file is located in the `source/` directory. The script will automatically extract it to the `data/` directory.

2. **Process the Images**:

   The script will read the CSV files, refine the images, and organize them accordingly.

## Dependencies

- Python 3.9+
- pandas
- OpenCV
- tarfile
- tqdm
- numpy

Install the required packages using pip:

```bash
pip install pandas opencv-python tarfile tqdm numpy
```

---
### LICENSE
---
This project is licensed under the MIT License.