Retinal blood vessel segmentation based on U-Net. The model is a U-Net, We replace Residual Double Convolution Unit with Multi-Fiber Unit, which reduces the model parameters by 76.1%, while performance remains basically unchanged.
Two open Dataset are used: DRIVE and CHASE_DB, and calssical preprocessing are applied: Gray-scale Conversion, Stadardization, CLAHE and γ-adjustment. We extract 64 * 64 pix size patches from DRIVE, 128 * 128 for CHASE_DB, then rotate patch to get 4 times more data, then normalize input data for model.
Models can be found in src/U_Net2.py and src/MF_UNet2.py, proposed model is located in the latter file.
Multi-Fiber Unit is defined in src/MF_UNet2.py.
It is based on grouped convolution, multiplexers are applied to ficilitate information exchange between fibers.
Trained model is given in /ParameterFile.
Proposed model is trained on both dataset, classical U-Net is trained on DRIVE dataset.
This an image from test set of DRIVE and CHASE_DB respectively.
Various metrics are used to compare the performence of two models, the most advantage of MF-U-Net is parameter size.
| Model | Parameter File | IOU | F1-Score |
|---|---|---|---|
| U-Net | 32.6MB | 0.6850 | 0.8131 |
| MF-U-Net | 7.6MB | 0.6837 | 0.8121 |