metadata

license: mit
library_name: pytorch
tags:
  - Medical Vsion-Language Pre-Training
  - BenchX

REFERS Checkpoint Model Card

A retrained REFERS model for benchmarking medical vision-language pre-training methods within the BenchX framework.

Model Details

Model Type: REFERS
Architecture: Vit-Base image encoder and BERT text encoder
Original Papers: Generalized Radiograph Representation Learning via Cross-Supervision Between Images and Free-Text Radiology Reports
Benchmark Paper: BenchX: A Unified Benchmark Framework for Medical Vision-Language Pretraining on Chest X-Rays
Benchmark Framework: https://github.com/yangzhou12/BenchX

Intended Use

Primary Use Cases:
- Benchmarking performance for Medical Image Classification
- Benchmarking performance for Medical Image Segmentation
- Benchmarking performance for Medical Report Generation

Pre-Training Data

Dataset:
- Data source(s): MIMIC-CXR
- Types of medical images: Frontal chest X-rays
- Text data type: Associated radiology reports

Prerequisites

Please follow the instruction to install BenchX.

Training & Evaluation

1. Classification

To fine-tune REFERS for classification, run this command:

python bin/train.py config/classification/<dataset_name>/refers.yml

2. Segmentation

To fine-tune REFERS for segmentation, run this command:

python mmsegmentation/tools/train.py config/benchmark/<dataset_name>/refers.yml

3. Report Generation

To fine-tune REFERS for report generation, run this command:

python bin/train.py config/report_generation/<dataset_name>/refers.yml

4. Evaluation

To evaluate fine-tuned REFERS models, run:

# For classification and report generation
python bin/test.py config/<task_name>/<dataset_name>/refers.yml validator.splits=[test] ckpt_dir=<path_to_checkpoint>

# For segmentation
python mmsegmentation/tools/my_test.py mmsegmentation/config/<dataset_name>/refers.yml <path_to_checkpoint>

Citations

@article{zhou2022generalized,
  title={Generalized radiograph representation learning via cross-supervision between images and free-text radiology reports},
  author={Zhou, Hong-Yu and Chen, Xiaoyu and Zhang, Yinghao and Luo, Ruibang and Wang, Liansheng and Yu, Yizhou},
  journal={Nature Machine Intelligence},
  volume={4},
  number={1},
  pages={32--40},
  year={2022}
}

@inproceedings{zhou2024benchx,
  title={BenchX: A Unified Benchmark Framework for Medical Vision-Language Pretraining on Chest X-Rays},
  author={Yang Zhou, Tan Li Hui Faith, Yanyu Xu, Sicong Leng, Xinxing Xu, Yong Liu, Rick Siow Mong Goh},
  booktitle={Proceedings of NeurIPS},
  year={2024}
}