jiajinuiuc / regression-metric-loss

A loss function originated for regression tasks to learn a representation manifold that is isometric to the label space.

RMLoss: Regression Metric Loss

RMLoss explores the structure of the continuous label space and regularizes the model to learn a better representation space which is a semantically meaningful manifold that is isometric to the label space. The paper has been accepted by MICCAI 2022.

Prerequisites

• Python 3.8
• PyTorch 1.8.2+
• A computing device with GPU

Getting started

Installation

• (Not necessary) Install Anaconda3
• Install CUDA 11.0+
• Install PyTorch 1.8.2+

Noted that our code is tested based on PyTorch 1.8.2

Dataset & Preparation

The original RSNA Pediatric Bone Age Dataset contains various noises. In our experiments, we used preprocessed data from this repository. All images are resized into 400x520.

• The trained model is at ./work/checkpoints
• The data splition used in our experiments is at ./work/data/data_info.csv
• Before running the code, please put the preprocessed images into ./work/data/img

Train

Train a model by

python train_main.py

Evaluation

Evaluate the trained model by

python test_main.py

• --iter iteration of the checkpoint to load. #Default: 14500
• --batch_size batch size of the parallel test. #Default: 64

Citation

Please cite these papers in your publications if it helps your research:

@article{chao2022regression,
  title={Regression Metric Loss: Learning a Semantic Representation Space for Medical Images},
  author={Chao, Hanqing and Zhang, Jiajin and Yan, Pingkun},
  journal={arXiv preprint arXiv:2207.05231},
  year={2022}
}

Link to paper:

• Regression Metric Loss: Learning a Semantic Representation Space for Medical Images

License

The source code of RMLoss is licensed under a MIT-style license, as found in the LICENSE file. This code is only freely available for non-commercial use, and may be redistributed under these conditions. For commercial queries, please contact Dr. Pingkun Yan.