Table 5 Results of ablation experiments on DRIVE and STARE dataset, where Res-UNet represents UNet that replaces all the original convolution modules with structured residual convolution modules, Att-Res UNet represents Attention residual UNet, SDR represents scale-aware dense residual module and ML represents multi-output weighted loss
From: Scale-aware dense residual retinal vessel segmentation network with multi-output weighted loss
| Â | DRIVE dataset | STARE dataset | ||||||
|---|---|---|---|---|---|---|---|---|
Module | dice | acc | mIoU | recall | dice | acc | mIoU | recall |
Attention U-Net [10] | 0.7851 | 0.9638 | 0.8041 | 0.8728 | 0.8130 | 0.9717 | 0.8293 | 0.8737 |
Res-UNet | 0.7871 | 0.9645 | 0.8055 | 0.8748 | 0.8184 | 0.9721 | 0.8319 | 0.8789 |
Att-Res UNet | 0.7959 | 0.9650 | 0.8119 | 0.8805 | 0.8256 | 0.9733 | 0.8376 | 0.8809 |
Att-Res UNet+SDR | 0.7998 | 0.9651 | 0.8147 | 0.8851 | 0.8299 | 0.9736 | 0.8409 | 0.8877 |
Att-Res UNet+ML | 0.8024 | 0.9660 | 0.8170 | 0.8810 | 0.8316 | 0.9739 | 0.8423 | 0.8864 |
Att-Res UNet+SDR+ML | 0.8041 | 0.9667 | 0.8189 | 0.8893 | 0.8350 | 0.9744 | 0.8445 | 0.8920 |