Giters

jlcsilva / EfficientUNetPlusPlus

Decoder architecture based on the UNet++. Combining residual bottlenecks with depthwise convolutions and attention mechanisms, it outperforms the UNet++ in a coronary artery segmentation task, while being significantly more computationally efficient.

Home Page:https://arxiv.org/abs/2106.11447

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Drive evaluation result

loopdigga96 opened this issue · comments

commented

Hi @jlcsilva thanks for your repository and paper. It was really interesting to read it but I cannot find any evaluation of your approach on public datasets. You mentioned DRIVE in repository but what is actual performance on this dataset?

commented

Hey @loopdigga96, you're welcome! Unfortunately, I still haven't obtained approval for sharing the coronary artery dataset. Regarding DRIVE, I still haven't had time to adjust parameters and train it. Due to the lower amount of data (~20 images vs ~200 images), the augmentation policy and loss function hyperparameters will probably need to be fine-tuned for optimal performance. I'm currently finishing a some other work, but I hope to be able to train it soon. I will get back to you with results when I do!

However, taking into account the 0.8149 F1-score of the ResUNet, which the EfficientUNet++ significantly outperforms in the coronary artery segmentation task, I would expect the EfficientUNet++ to perform better or at least as well as the ResUNet (while being more computationally efficient!)

commented

Thanks for you response @jlcsilva! It would be great to hear the results from you.