Our team in MICCAI MMAC 2023 - Myopic Maculopathy Analysis Challenge - Task 1

This repository contains our team's training lists and scripts used in MICCAI MMAC 2023. An effective computer-aided system is essential to help ophthalmologists analyze myopic maculopathy, further providing accurate diagnosis and reasonable intervention for this disease. Given a training dataset, the task of the challenge was to classify the five categories of myopic macular degeneration. We trained several deep learning models based on the Swin-base. Our best submission achieved a Mean score (Quadratic-weighted Kappa (QWK), F1 score, Specificity) score of 0.8660 (Rank 2) in the Test Phase and 0.8823 (Rank 1) in the Futher Test Phase. The official results can be found here.

Methodology

Our all models are based on the Swin-Transformer, which is a vision transformer model. We used the Swin-Base model with the input size of 384x384. We add ArcFace Loss to the model to improve the classification performance, and we use a weak label strategy to train the model.

Getting Started

Step 1 : Configure the mmpretrain environment

git clone https://github.com/qingtian5/miccai23_mmac_task1.git
cd miccai23_mmac_task1
conda create -n miccai_mmpre python=3.8 pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.3 -c pytorch -y
conda activate miccai_mmpre
pip3 install openmim
git clone https://github.com/open-mmlab/mmpretrain.git
cd mmpretrain
pip3 install openmim && mim install -e .
cd ..

Step 2 : Place the images in the data directory

1. The dataset can be download by the Google Drive. And the directory under data should be as follows.

data
├── classification_train_val # This means the training set and the validation set are put together 
├── classification_train
└── classification_val

2. For each folder, the ground truth label file (meta/train.txt) needs to be generated in the following way. If it is a validation set, change the path to meta/val.txt.

import pandas as pd

data = pd.read_csv("MICCAI_TASK1/data/classification_train/Groundtruths/train_labels.csv")

with open("MICCAI_TASK1/data/classification_train/meta/train.txt","w") as f:
    for idx, d in data.iterrows():
        f.write(d["image"] + " " + str(d["myopic_maculopathy_grade"]) + "\n")

3. The following format is required for each folder under data. If is the validation dataset, the label file should be meta/val.txt.

classification_train
├── Groundtruths
│   └── train_labels.csv
├── Images
│   └── train
│       ├── mmac_task_1_train_0001.png
│       ├── mmac_task_1_train_0002.png
│       ├── ...
│       └── mmac_task_1_train_1143.png
├── LICENSE.txt
└── meta
    └── train.txt

classification_val
├── Groundtruths
│   └── val_labels.csv
├── Images
│   └── val
│       ├── mmac_task_1_val_0001.png
│       ├── mmac_task_1_val_0002.png
│       ├── ...
│       └── mmac_task_1_val_0248.png
├── LICENSE.txt
└── meta
    └── val.txt

Step 3 : Place pre-trained weights in the pretrained_ckpt catalog.

Run the following command on the terminal

cd pretrained_ckpt
wget https://download.openmmlab.com/mmclassification/v0/swin-transformer/convert/swin_base_patch4_window7_224_22kto1k-f967f799.pth
cd ..

Step 4 : Run the training script

Run the following command on the terminal

cd mmpretrain
python tools/train.py ../projects/submission/my_swin_base_in1k_384.py

Step 5 : submit the results

Run the following command on the terminal

cd ../projects/submission
mv ../working/epoch_36.pth ./epoch_36.pth
zip -r submission.zip .

Then submit the submission.zip file to the competition website Link.

Citation

If you find our work helpful, please consider citing the following papers

@inproceedings{lu2023swin,
  title={Swin-MMC: Swin-Based Model for Myopic Maculopathy Classification in Fundus Images},
  author={Lu, Li and Pan, Xuhao and Jin, Panji and Ding, Ye},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={18--30},
  year={2023},
  organization={Springer}
}