NeurIPS 2023 - MedFM: Foundation Model Prompting for Medical Image Classification Challenge 2023

A naive baseline and submission demo for the Foundation Model Prompting for Medical Image Classification Challenge 2023 (MedFM).

✨ Notification

Please check out master branch. Third party implementation of MedFMC baseline is supported! It is based on the MMPreTrain, with backbone of ViT-cls, ViT-eva02, ViT-dinov2, Swin-cls and ViT-clip. More details could be found in its document. Thanks Ezra-Yu for this excellent work.

🛠️ Installation

Install requirements by

$ conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=10.1 -c pytorch
$ pip install mmcls==0.25.0 openmim scipy scikit-learn ftfy regex tqdm
$ mim install mmcv-full==1.6.0

We suggest you install PyTorch successfully first, then install OpenMMLab packages and their dependencies.

Moreover, you can use other Computer Vision or other foundation models such as EVA and CLIP.

📊 Results

The results of ChestDR, ColonPath and Endo in MedFMC dataset and their corresponding configs on each task are shown as below.

Few-shot Learning Results

We utilize Visual Prompt Tuning method as the few-shot learning baseline, whose backbone is Swin Transformer. The results are shown as below:

ChestDR

N Shot	Crop Size	Epoch	mAP	AUC	Config
1	384x384	20	13.14	56.49	config
5	384x384	20	17.05	64.86	config
10	384x384	20	19.01	66.68	config

ColonPath

N Shot	Crop Size	Epoch	Acc	AUC	Config
1	384x384	20	77.60	84.69	config
5	384x384	20	89.29	96.07	config
10	384x384	20	91.21	97.14	config

Endo

N Shot	Crop Size	Epoch	mAP	AUC	Config
1	384x384	20	19.70	62.18	config
5	384x384	20	23.88	67.48	config
10	384x384	20	25.62	71.41	config

Transfer Learning on 20% (Fully Supervised Task)

Noted that MedFMC mainly focuses on few-shot learning i.e., transfer learning task. Thus, fully supervised learning tasks below only use 20% training data to make corresponding comparisons.

ChestDR

Backbone	Crop Size	Epoch	mAP	AUC	Config
DenseNet121	384x384	20	24.48	75.25	config
EfficientNet-B5	384x384	20	29.08	77.21	config
Swin-B	384x384	20	31.07	78.56	config

ColonPath

Backbone	Crop Size	Epoch	Acc	AUC	Config
DenseNet121	384x384	20	92.73	98.27	config
EfficientNet-B5	384x384	20	94.04	98.58	config
Swin-B	384x384	20	94.68	98.35	config

Endo

Backbone	Crop Size	Epoch	mAP	AUC	Config
DenseNet121	384x384	20	41.13	80.19	config
EfficientNet-B5	384x384	20	36.95	78.23	config
Swin-B	384x384	20	41.38	79.42	config

🎫 License

This project is released under the Apache 2.0 license.

🙌 Usage

Data preparation

Prepare data following MMClassification. The data structure looks like below:

data/
├── MedFMC
│   ├── chest
│   │   ├── images
│   │   ├── chest_X-shot_train_expY.txt
│   │   ├── chest_X-shot_val_expY.txt
│   │   ├── train_20.txt
│   │   ├── val_20.txt
│   │   ├── trainval.txt
│   │   ├── test_WithLabel.txt
│   ├── colon
│   │   ├── images
│   │   ├── colon_X-shot_train_expY.txt
│   │   ├── colon_X-shot_val_expY.txt
│   │   ├── train_20.txt
│   │   ├── val_20.txt
│   │   ├── trainval.txt
│   │   ├── test_WithLabel.txt
│   ├── endo
│   │   ├── images
│   │   ├── endo_X-shot_train_expY.txt
│   │   ├── endo_X-shot_val_expY.txt
│   │   ├── train_20.txt
│   │   ├── val_20.txt
│   │   ├── trainval.txt
│   │   ├── test_WithLabel.txt

Noted that the .txt files includes data split information for fully supervised learning and few-shot learning tasks. The public dataset is splitted to trainval.txt and test_WithLabel.txt, and trainval.txt is also splitted to train_20.txt and val_20.txt where 20 means the training data makes up 20% of trainval.txt. And the test_WithoutLabel.txt of each dataset is validation set.

Corresponding .txt files are stored at ./data_backup/ folder, the few-shot learning data split files {dataset}_{N_shot}-shot_train/val_exp{N_exp}.txt could also be generated as below:

python tools/generate_few-shot_file.py

Where N_shot is 1,5 and 10, respectively, the shot is of patient(i.e., 1-shot means images of certain one patient are all counted as one), not number of images.

The images in each dataset folder contains its images, which could be achieved from original dataset.

Training and evaluation using OpenMMLab codebases.

In this repository we provided many config files for fully supervised task (only uses 20% of original traning set, please check out the .txt files which split dataset) and few-shot learning task.

The config files of fully supervised transfer learning task are stored at ./configs/densenet, ./configs/efficientnet, ./configs/vit-base and ./configs/swin_transformer folders, respectively. The config files of few-shot learning task are stored at ./configs/ablation_exp and ./configs/vit-b16_vpt folders.

For the training and testing, you can directly use commands below to train and test the model:

# you need to export path in terminal so the `custom_imports` in config would work
export PYTHONPATH=$PWD:$PYTHONPATH
# Training
# you can choose a config file like `configs/vit-b16_vpt/in21k-vitb16_vpt1_bs4_lr6e-4_1-shot_chest.py` to train its model
python tools/train.py $CONFIG

# Evaluation
# Endo and ChestDR utilize mAP as metric
python tools/test.py $CONFIG $CHECKPOINT --metrics mAP
python tools/test.py $CONFIG $CHECKPOINT --metrics AUC_multilabel
# Colon utilizes accuracy as metric
python tools/test.py $CONFIG $CHECKPOINT --metrics accuracy --metric-options topk=1
python tools/test.py $CONFIG $CHECKPOINT --metrics AUC_multiclass

The repository is built upon MMClassification/MMPretrain. More details could be found in its document.

Generating Submission results of Validation Phase

Noted:

• The order of filanames of all CSV files must follow the order of provided colon_val.csv, chest_val.csv and endo_val.csv! You can see files in ./data_backup/result_sample for more details.
• The name of CSV files in result.zip must be the same names xxx_N-shot_submission.csv below.

Run

python tools/test_prediction.py $DATASETPATH/test_WithoutLabel.txt $DATASETPATH/images/ $CONFIG $CHECKPOINT --output-prediction $DATASET_N-shot_submission.csv

For example:

python tools/test_prediction.py data/MedFMC/endo/test_WithoutLabel.txt data/MedFMC/endo/images/ $CONFIG $CHECKPOINT --output-prediction endo_10-shot_submission.csv

You can generate all prediction results of endo_N-shot_submission.csv, colon_N-shot_submission.csv and chest_N-shot_submission.csv and zip them into result.zip file. Then upload it to Grand Challenge website.

result/
├── endo_1-shot_submission.csv
├── endo_5-shot_submission.csv
├── endo_10-shot_submission.csv
├── colon_1-shot_submission.csv
├── colon_5-shot_submission.csv
├── colon_10-shot_submission.csv
├── chest_1-shot_submission.csv
├── chest_5-shot_submission.csv
├── chest_10-shot_submission.csv

Then using zip to make them as .zip file(i.e., result_sample.zip in ./data_backup folder) and upload it to submission site of Grand Challenge MedFMC Validation Phase.

🖊️ Citation

@article{wang2023real,
  title={A real-world dataset and benchmark for foundation model adaptation in medical image classification},
  author={Wang, Dequan and Wang, Xiaosong and Wang, Lilong and Li, Mengzhang and Da, Qian and Liu, Xiaoqiang and Gao, Xiangyu and Shen, Jun and He, Junjun and Shen, Tian and others},
  journal={Scientific Data},
  volume={10},
  number={1},
  pages={574},
  year={2023},
  publisher={Nature Publishing Group UK London}
}