search

SHI-Labs / SGL-Retinal-Vessel-Segmentation

[MICCAI 2021] Study Group Learning: Improving Retinal Vessel Segmentation Trained with Noisy Labels: New SOTA on both DRIVE and CHASE_DB1.
https://arxiv.org/abs/2103.03451
MIT License
102 stars 24 forks source link

Enhancement map #2

Closed Darcy103 closed 2 years ago

Darcy103 commented 3 years ago

Thank you very much for your research work. I have a question to ask you. Does the enhancement map have GT supervision? Or is it just an intermediate result of the three channels learned by the first network?

yzhouas commented 3 years ago

Thanks for your interests. There is no GT for the enhancement map, but ony the intermediate result in the middle bottleneck layers. It can be well learned to illustrate the vessel structure because it can also learn information from the segmentation loss in the final layer.