ALPA: Taming the Tail

This repository contains the implementation of the paper "Taming the Tail: Leveraging Asymmetric Loss and Padé Approximation to Overcome Medical Image Long-Tailed Class Imbalance". The code addresses challenges in medical imaging caused by long-tailed class distributions, introducing an innovative approach based on asymmetric loss functions and Padé approximations to improve classification performance on imbalanced datasets.

---

Features

• Asymmetric Loss (AL): Reduces the impact of over-represented classes while improving performance for under-represented ones.
• Padé Approximation: Enhances numerical stability for gradient computation, addressing the skewness of class distributions.
• K-Fold Cross-Validation: Supports robust evaluation across multiple folds.
• Scalable Design: Supports ConvNeXt and other backbones for experiments with configurable parameters.

---

Loss function code

• ALPA code is there in models/get_optimizer.py

---

Installation

1. Clone the repository:

   git clone https://github.com/your-repo/alpa.git
   cd alpa

2. Install dependencies:

   pip install -r requirements.txt

3. Ensure GPU compatibility with PyTorch:

   conda install pytorch torchvision torchaudio pytorch-cuda=11.x -c pytorch -c nvidia

---

Training

To train the model with ALPA, use the following command:

python3 main.py \
    --is_master=1 \
    --train=1 \
    --gpu_ids="cuda:3" \
    --criterion="Pade" \
    --seed=0 \
    --model=convnext \
    --batchsize=128 \
    --epochs=50 \
    --img_size=256 \
    --model_path="result" \
    --save_model=1 \
    --store_name="_kfold"

Key Arguments:

• --is_master: Specifies if the process is the main training process (1 for master).
• --train: Set to 1 to enable training mode.
• --gpu_ids: Specify the GPU(s) for training (e.g., "cuda:0").
• --criterion: Set to "Pade" for the Padé Approximation loss.
• --seed: Random seed for reproducibility.
• --model: Backbone model for training (e.g., convnext).
• --batchsize: Batch size for training.
• --epochs: Total number of epochs.
• --img_size: Input image resolution (e.g., 256x256).
• --model_path: Path to save the trained model.
• --save_model: Set to 1 to save the best-performing model.
• --store_name: Suffix for the model filename to track experiments.

---

Evaluation

You can evaluate the trained model use test.py

---

Citation

If you find this code helpful for your research, please cite.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Contact

For questions or feedback, feel free to reach out to pankhikashyap.research@gmail.com.