Learning_AVSegmentation 👀

--Vessel segmentation, artery and vein, retinal image

--Code for MICCAI paper "Learning to Address Intra-segment Misclassification in Retinal Imaging"

Please contact ykzhoua@gmail.com or yukun.zhou.19@ucl.ac.uk if you have questions.

Brief Background

This repository aims at improving multi-class vessel segmentation performance in retinal fundus photograph by alleviating the intra-segment misclassification around the intersections. The research data sets in experiments include DRIVE-AV ^1,2, LES-AV ³, and HRF-AV ^4,5.

Advantages

There are a few strengths in this work:

We strictly evaluate the method performance in multi-class segmentation manner, instead of only considering the classification accuracy (previous evaluation). Mean value and standard deviation are calculated to show robust performance in test.
The GAN-based segmentation backbone is revised based on a SOTA vessel segmentation method ⁶.
The binary-to-multi fusion network avoids directly learning on the ambiguous pixel label brought by intersections, achieving SOTA performance on multi-class vessel segmentation.
The code and algorithm are easily transferred to other medical or natural linear segmentation fields.

Install

Requirements

Work on Linux and Windows, but Linux is preferred to replicate the reported performance.
This project is based on pytorch==1.6.0, torchvision==0.7.0, CUDAToolkit==10.1(10.2-11.3 is capable).
A GPU is essential. In our work, we utilise one Tesla T4 with 15 GB of DRAM. If with weaker GPU, we suggest to change the image size setting in scripts.utils.py

Packages installation:

pip install -r requirements.txt

Usage

Pretrained Model

The pretrained model are provided in Google_DRIVE. Download them and unzip them directly at the project folder.

Train

Start training, the dataset can be set as DRIVE_AV, LES-AV, or HRF-AV.

python train.py --e=500 \
                --batch-size=2 \
                --learning-rate=8e-4 \
                --v=10.0 \
                --alpha=0.5 \
                --beta=1.1 \
                --gama=0.08 \
                --dataset=DRIVE_AV \
                --discriminator=unet \
                --job_name=DRIVE_AV_randomseed_42 \
                --uniform=True \
                --seed_num=42

Test

Test the trained models.

python test.py --batch-size=1 \
               --dataset=DRIVE_AV \
               --job_name=DRIVE_AV_randomseed \
               --uniform=True

Performance

Switch final activation map from sigmoid to softmax

Test dataset	Sensitivity	AUC-ROC	F1-score	AUC-PR	MSE
DRIVE-AV	70.8 ± 0.1	84.7 ± 0.05	71.99 ± 0.04	73.06 ± 0.03	2.85 ± 0.01
LES-AV	64.41 ± 0.09	81.72 ± 0.04	67.22 ± 0.06	69.08 ± 0.06	2.22 ± 0.01
HRF-AV	71.85 ± 0.29	85.38 ± 0.13	71.92 ± 0.03	73.23 ± 0.03	2 ± 0.01

Reference

Staal J, Abràmoff M D, Niemeijer M, et al. Ridge-based vessel segmentation in color images of the retina[J]. IEEE transactions on medical imaging, 2004, 23(4): 501-509.
Hu Q, Abràmoff M D, Garvin M K. Automated separation of binary overlapping trees in low-contrast color retinal images[C]//International conference on medical image computing and computer-assisted intervention. Springer, Berlin, Heidelberg, 2013: 436-443.
Orlando J I, Breda J B, Van Keer K, et al. Towards a glaucoma risk index based on simulated hemodynamics from fundus images[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018: 65-73.
Budai A, Bock R, Maier A, et al. Robust vessel segmentation in fundus images[J]. International journal of biomedical imaging, 2013, 2013.
Hemelings R, Elen B, Stalmans I, et al. Artery–vein segmentation in fundus images using a fully convolutional network[J]. Computerized Medical Imaging and Graphics, 2019, 76: 101636.
Zhou Y, Chen Z, Shen H, et al. A refined equilibrium generative adversarial network for retinal vessel segmentation[J]. Neurocomputing, 2021, 437: 118-130.

Citation

@inproceedings{zhou2021learning,
  title={Learning to address intra-segment misclassification in retinal imaging},
  author={Zhou, Yukun and Xu, Moucheng and Hu, Yipeng and Lin, Hongxiang and Jacob, Joseph and Keane, Pearse A and Alexander, Daniel C},
  booktitle={Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th International Conference, Strasbourg, France, September 27--October 1, 2021, Proceedings, Part I 24},
  pages={482--492},
  year={2021},
  organization={Springer}
}


@article{zhou2022automorph,
  title={AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline},
  author={Zhou, Yukun and Wagner, Siegfried K and Chia, Mark A and Zhao, An and Xu, Moucheng and Struyven, Robbert and Alexander, Daniel C and Keane, Pearse A and others},
  journal={Translational vision science \& technology},
  volume={11},
  number={7},
  pages={12--12},
  year={2022},
  publisher={The Association for Research in Vision and Ophthalmology}
}

rmaphoh/Learning-AVSegmentation