Automated Lesion Segmentation in Fundus Images with Many-to-Many Reassembly of Features

This repo is the official implementation of paper "Automated Lesion Segmentation in Fundus Images with Many-to-Many Reassembly of Features".

Environment

This code is based on mmsegmentation.

• pytorch=1.6.0
• mmsegmentation=0.8.0
• mmcv=1.2.0

``` conda create -n m2mrf python=3.7 -y conda activate m2mrf

conda install pytorch=1.6.0 torchvision cudatoolkit=10.2 -c pytorch -y pip install mmcv-full==1.2.0 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.6.0/index.html -i https://pypi.douban.com/simple/ pip install opencv-python pip install scipy pip install tensorboard tensorboardX pip install sklearn pip install terminaltables pip install matplotlib

cd M2MRF-Lesion-Segmentation chmod u+x tools/* pip install -e . ```

Training and testing

```

prepare dataset

python tools/prepare_labels.py python tools/augment.py

train

CUDAVISIBLEDEVICES=0,1,2,3 PORT=12345 tools/disttrain.sh configs/m2mrf/fcnhr48-M2MRF-C40kidrid_bdice.py 4

test

CUDAVISIBLEDEVICES=0,1,2,3 PORT=12345 tools/disttest.sh configs/m2mrf/fcnhr48-M2MRF-C40kidridbdice.py /path/to/fcnhr48-M2MRF-C40kidridbdiceiter_40000.pth 4 --eval mIoU ```

Results and models

We evaluate our method on IDRiD and DDR.

IDRiD

| method |   mIOU   | mAUPR | download | | ------- | :--------------------------: | :---: | -------------------------------------------------------------------------------------------------------------------------------------------------------------- | | M2MRF-A | 49.86 | 67.15 | config | model | | M2MRF-B | 49.33 | 66.71 | config | model | | M2MRF-C | 50.17 | 67.55 | config | model | | M2MRF-D | 49.96 | 67.32 | config | model |

DDR

| method |   mIOU   | mAUPR | download | | ------- | :--------------------------: | :---: | ------------------------------------------------------------------------------------------------------------------------------------------------------------ | | M2MRF-A | 31.47 | 49.56 | config | model | | M2MRF-B | 30.56 | 49.86 | config | model | | M2MRF-C | 30.39 | 49.20 | config | model | | M2MRF-D | 30.76 | 49.47 | config | model |

In the paper, we reported average performance over three repetitions, but our code only reported the best one among them.

Citation

If you find this code useful in your research, please consider citing:

latex @article{liu2023m2mrf, title = {Automated Lesion Segmentation in Fundus Images with Many-to-Many Reassembly of Features}, author = {Qing Liu and Haotian Liu and Wei Ke and Yixiong Liang}, journal = {Pattern Recognition}, volume = {136}, pages = {109191}, year = {2023}, issn = {0031-3203}, doi = {https://doi.org/10.1016/j.patcog.2022.109191}, }