Skip to content

Latest commit

 

History

History
 
 

CLPI-Collaborative-Learning-for-Diabetic-Retinopathy-Grading

README.md

CLPI: Robust Collaborative Learning of Patch-level and Image-level Annotations for Diabetic Retinopathy Grading from Fundus Image

In this paper, we present a robust framework, which can collaboratively utilize both patch-level lesion and image-level grade annotations, for DR severity grading. By end-to-end optimizing the entire framework, the fine-grained lesion and image-level grade information can be bidirectionally exchanged to exploit more discriminative features for DR grading. Compared with the recent state-of-the-art algorithms and three over 9-years clinical experience ophthalmologists, the proposed algorithm shows effective performance. Testing on the datasets from totally different label distributions and scenarios, our algorithm is proved robust in facing distribution and low image quality problems that is commonly exists in real world practice. Extensive experimental ablation studies dissect the proposed framework into parts, and reveal the effective and indispensable of each component.

Arch

Requirements

- PaddlePaddle-GPU >= 1.6.3
- opencv-python 3.4.3.18
- pandas 1.0.5
- tabulate 0.8.7
- scikit-learn (optional, for kappa evaluation)

Preparation

For uncompressing model weights:

cd ./demo/
tar zxvf messidor_densenet_binary_full_best.tar.gz

cd ../

Inference

Create infer file (see demo/infer_file_demo.csv)

Write relative path of images start from ./demo/

demo/images/20051213_62648_0100_PP.png
demo/images/20051021_59136_0100_PP.png
demo/images/20051214_40719_0100_PP.png
demo/images/20051205_31994_0400_PP.png

Put images into ./demo/images/

Due to the limitation of the size of the images, we can not upload unmodified images from Messidor-2, please download Messidor-2 dataset and put the four images listed above into ./demo/images/.

Run inference

FLAGS_fraction_of_gpu_memory_to_use=0.1 \
FLAGS_eager_delete_tensor_gb=1 \
CUDA_VISIBLE_DEVICES=0 \
python3 -u inference.py \
        -a DenseNet121 \
        --resume-from demo/messidor_densenet_full_best/ \
        --infer-file='./demo/infer_file_demo.csv' --infer-classdim=5

A pandas dataframe will be displayed like this:

pred
demo/images/20051213_62648_0100_PP.png 4
demo/images/20051021_59136_0100_PP.png 2
demo/images/20051214_40719_0100_PP.png 1
demo/images/20051205_31994_0400_PP.png 2

Evaluation

You might modify main() @ inference.py to obtain the dataframe named prediction, then concatenate DR Label from Messidor-2 by file name.

And evaluate kappa score:

from sklearn.metrics import cohen_kappa_score
kappa_score = cohen_kappa_score(..., weights="quadratic")

CAM

For acquire CAM(class activation heatmaps):

FLAGS_fraction_of_gpu_memory_to_use=0.1 \
FLAGS_eager_delete_tensor_gb=1 \
CUDA_VISIBLE_DEVICES=0 \
python3 -u get_cam_heatmap.py \
        -a DenseNet121 \
        --resume-from demo/messidor_densenet_full_best/ \
        --infer-file='./demo/infer_file_demo.csv' \
        --infer-classdim=5

Visualization

CAM

Appendix

  • For reproducible, we provide our train / validation / test split on Messidor-2 following 6 / 2 / 2, because of the lack of official test set. Please refer to demo/partition/

Citation

If you use this code for your research, please cite our papers.

@misc{yang2020robust,
    title={Robust Collaborative Learning of Patch-level and Image-level Annotations for Diabetic Retinopathy Grading from Fundus Image},
    author={Yehui Yang and Fangxin Shang and Binghong Wu and Dalu Yang and Lei Wang and Yanwu Xu and Wensheng Zhang and Tianzhu Zhang},
    year={2020},
    eprint={2008.00610},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}