luozhengrong / hdc-net Goto Github PK

Dear Mr. Luo,
Recently, when reproducing your HDCNet, the ET area has been very low, and the overall effect is not up to the accuracy of the paper. How do you deal with the accuracy of the paper? Have you used other post-processing operations? In addition, when reading your paper, I found that the ET area effect of HDC + is very high. I don't seem to find HDC + in the code. Would you like to provide the code of HDC +? Is the parameter setting for cross-validation training the same as that for ordinary training?

Hi!

Thank you for making the source code of your work available. I noticed that the testing routine does not work without Ground Truth segmentations due to issue in BraTSDataset class of data/datasets.py (even with mode=2). I can proceed to make changes in the code to enable it, but I wanted to have your version of the code to reproduce the results in the paper.

Best Regards,
Talha

Hello, Mr. Luo, I find some questions to ask you when I repeat your work.

1. The optimal epoch mentioned in the paper is 800, and 900 is set in the code, because the code does not set to save the optimal model, but only saves the last model. I would like to ask you how to determine the optimal model for training and test it? Furthermore, is the last model saved by cross-training the loss-optimal model?
2. According to the segmentation results, there is a big difference between the ET region and the paper, because there are many segmentation results that are 0. How did you solve this problem at that time?
3. Finally, I don't know if you are willing to upload HDC of other structures for reference only.

Looking forward to your reply, best wishes!