MFI-Net: A multi-resolution fusion input network for retinal vessel segmentation

Authors: Yun Jiang, Chao Wu, Ge Wang, Hui-Xia Yao, Wen-Huan Liu

0. Preface

• This repository provides code for "MFI-Net: A multi-resolution fusion input network for retinal vessel segmentation" PLOS ONE. (https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0253056)

1. MFI-Net

Figure 1. Overview of MFI-Net segmentation model for retinal vessel.

2. Dataset

Figure 2. Sample images of DRIVE, CHASE DB1 and STARE datasets.

• DRIVE DRIVE dataset files are available from http://www.isi.uu.nl/Research/Databases/DRIVE.
• CHASE_DB1 CHASE_DB1 dataset files are available from https://blogs.kingston.ac.uk/retinal/chasedb1/
• CHASE_DB1 STARE dataset files are available from http://cecas.clemson.edu/~ahoover/stare/.

3. Usage

1. Train

python 2_train_dynamic.py --device 0 --dataset DRIVE --data_path /home/data/dataset/ --model UNet --epoch 200 --batch_size 1024 --patch_num 10000 --logs_path /home/data/logs/ --lr 0.001

2. Test

python 3_test.py --device 0 --check_path YOUR_LOG_PATH

4. Result

Table 1. Ablation experiment results.

Figure 3. Segmentation result of ablation experiment.

Figure 4. Comparison of segmentation results of MFI-Net(ours), UNet++ and AA-UNet.

5. Our related work

A Multi-Scale Residual Attention Network for Retinal Vessel Segmentation. (https://www.mdpi.com/2073-8994/13/1/24)

Multi-Scale and Multi-Branch Convolutional Neural Network for Retinal Image Segmentation. (https://www.mdpi.com/2073-8994/13/3/365)

⬆ back to top