Awesome

MUMC

Masked Vision and Language Pre-training with Unimodal and Multimodal Contrastive Losses for Medical Visual Question Answering

This is the official implementation of MUMC for the medical visual question answering, which was accepted by MICCAI-2023. Our proposal achieves superior accuracy in comparison with other state-of-the-art (sota) methods on three public medical VQA datasets: VQA-RAD dataset, PathVQA dataset and Slake dataset. Paper link here.

This repository is based on our previous work and inspired by @Junnan Li's work. We sincerely thank for their sharing of the codes.

Requirements

Run the following command to install the required packages:

pip install -r requirements.txt

Training and Testing

1. Dataset Preparation

Please organize the datasets as the following structure:

+--clef2022
| +--train
| | +--ImageCLEFmedCaption_2022_train_000001.jpg
| | +--ImageCLEFmedCaption_2022_train_000002.jpg
| | +--...
| +--valid
| | +--ImageCLEFmedCaption_2022_valid_084258.jpg
| | +--ImageCLEFmedCaption_2022_valid_084259.jpg
| | +--...
| +--clef22022_train.json
| +--clef22022_valid.json

+--data_RAD
| +--images
| | +--synpic100132.jpg
| | +--synpic100176.jpg
| | +--...
| +--trainset.json
| +--testset.json
| +--answer_list.json

+--data_PathVQA
| +--images
| | +--train
| | | +--train_0000.jpg
| | | +--train_0001.jpg
| | | +--...
| | +--val
| | | +--val_0000.jpg
| | | +--val_0001.jpg
| | | +--...
| | +--test
| | | +--test_0000.jpg
| | | +--test_0001.jpg
| | | +--...
| +--pathvqa_test.json
| +--pathvqa_train.json
| +--pathvqa_val.json
| +--answer_trainval_list.json

+--data_Slake
| +--imgs
| | +--xmlab0
| | | +--source.jpg.jpg
| | | +--question.json
| | | +--...
| | +--....
| +--slake_test.json
| +--slake_train.json
| +--slake_val.json
| +--answer_list.json

2. Pre-training

python3 pretrain  --output_dir ./pretrain

3. Finetune on Medical VQA tasks

# choose medical vqa dataset(rad, pathvqa, slake)
python3 train_vqa.py --dataset_use rad --checkpoint ./pretrain/med_pretrain_29.pth  --output_dir ./output/rad

Comparison with the sota

Pretrained weights

You can download the pre-trained weights through the following link.

Citation:

@article{MUMC,
  title     = {Masked Vision and Language Pre-training with Unimodal and Multimodal Contrastive Losses for Medical Visual Question Answering},
  author    = {Pengfei Li, Gang Liu, Jinlong He, Zixu Zhao and Shenjun Zhong},
  booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2023},
  year      = {2023},
  pages     = {374--383},
  publisher = {Springer Nature Switzerland}
}

License

MIT License