This repository contains the implementation of fully convolutional neural networks for segmenting retinal vasculature from fundus images.
Four architecures/models were made keeping U-NET architecture as the base. The models used are:
- Simple U-NET
- Residual U-NET (Res-UNET)
- Attention U-NET
- Residual Attention U-NET (RA-UNET)
The performance metrics used for evaluation are accuracy and mean IoU.
Images from HRF, DRIVE and STARE datasets are used for training and testing. The following pre-processing steps are applied before training the models:
- Green channel selection
- Contrast-limited adaptive histogram equalization (CLAHE)
- Cropping into non-overlapping patches of size 512 x 512
10 images from DRIVE and STARE and 12 images from HRF was kept for testing the models. The training dataset was then split into 70:30 ratio for training and validation.
Adam optimizer with a learning rate of 0.001 was used as optimizer and IoU loss was used as the loss function. The models were trained for 150 epochs with a batch size of 16, using NVIDIA Tesla P100-PCIE GPU.
The performance of the models were evaluated using the test dataset. Out of all the models, Attention U-NET achieved a greater segmentation performance.
The following table compares the performance of various models
| Datasets | Models | Average Accuracy | Mean IoU |
|---|---|---|---|
| HRF | Simple U-NET | 0.965 | 0.854 |
| HRF | Res-UNET | 0.964 | 0.854 |
| HRF | Attention U-NET | 0.966 | 0.857 |
| HRF | RA-UNET | 0.963 | 0.85 |
| DRIVE | Simple U-NET | 0.9 | 0.736 |
| DRIVE | Res-UNET | 0.903 | 0.741 |
| DRIVE | Attention U-NET | 0.905 | 0.745 |
| DRIVE | RA-UNET | 0.9 | 0.735 |
| STARE | Simple U-NET | 0.882 | 0.719 |
| STARE | Res-UNET | 0.893 | 0.737 |
| STARE | Attention U-NET | 0.893 | 0.738 |
| STARE | RA-UNET | 0.891 | 0.733 |
The datasets of the fundus images can be acquired from:
The trained models are present in Trained models folder.
[1] Vengalil, Sunil Kumar & Sinha, Neelam & Kruthiventi, Srinivas & Babu, R. (2016). Customizing CNNs for blood vessel segmentation from fundus images. 1-4. 10.1109/SPCOM.2016.7746702..
[2] Ronneberger O., Fischer P., Brox T. U-Net: Convolutional networks for biomedical image segmentation International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer (2015), pp. 234-241
[3] Zhang, Zhengxin & Liu, Qingjie. (2017). Road Extraction by Deep Residual U-Net. IEEE Geoscience and Remote Sensing Letters. PP. 10.1109/LGRS.2018.2802944.
[4] Oktay, Ozan & Schlemper, Jo & Folgoc, Loic & Lee, Matthew & Heinrich, Mattias & Misawa, Kazunari & Mori, Kensaku & McDonagh, Steven & Hammerla, Nils & Kainz, Bernhard & Glocker, Ben & Rueckert, Daniel. (2018). Attention U-Net: Learning Where to Look for the Pancreas.
[5] Ni, Zhen-Liang & Bian, Gui-Bin & Zhou, Xiao-Hu & Hou, Zeng-Guang & Xie, Xiao-Liang & Wang, Chen & Zhou, Yan-Jie & Li, Rui-Qi & Li, Zhen. (2019). RAUNet: Residual Attention U-Net for Semantic Segmentation of Cataract Surgical Instruments.
[6] Jin, Qiangguo & Meng, Zhaopeng & Pham, Tuan & Chen, Qi & Wei, Leyi & Su, Ran. (2018). DUNet: A deformable network for retinal vessel segmentation.