hongweilibran / wmh_ibbmtum Goto Github PK

Hi, Very much appreciate your beautiful work!

While runing the train_leave_one_out.py without changing anything, I found the dice loss is always .99x.
Also, I saw the code sinnipet
{if history.history['loss'][-1] > 0.99:
model = get_unet(img_shape, first5)
current_epoch = 1}(line 189 - line 191)
May I ask what's the reason for that ?
Also, would you share your insight about the layer initialization for this taks? (I saw the random initialization is used according to the paper)

Hi!

I have found that in pretrained_FLAIR_only and pretrained_FLAIR_T1 folders there are three h5 files on each. I haven't found any information about them in the README file.

Why are there three different files?

Thanks!

Hi Hongwei, could you advise how your wining approach can apply to other datasets for training/inference, regardless of the testing data used for the challenge? Also python2.7 is no longer supported. Any plan to update this amazing application with python3? Thanks.

Hi!

When I try to download your npy files to run the leave-one-subject-out experiments, the link https://drive.google.com/open?id=1m0H9vbFV8yijvuTsAqRAUQGGitanNw_k doesn't work. It said that the file has been moved to the trash.

Could you please restore it?

Thanks!

Could you please upload all the weights, when I'm trying to make the decompress it says is missing one file z03. Please help me with the other files

Hello, I want to reproduce your results on WMH dataset, but I have waited the responses from the official email of MICCAI2017 challenge for two weeks. I wonder if you could share the download link?

Hi
Could you please tell me the exact versions of python 3 and tensorflow and keras in new updates?
Thanks

Hi,

I've been trying to run the leave one out tests for this tool, but I'm not getting the same level of performance that you report in the paper. I think this is because as it's set up at the moment, the leave one out tests only use a single network rather than the ensemble (i.e. I only get one set of weights per subject).

In line 253 of test_leave_one_out you have the comment #if you want to test three models ensemble, just do like this: pred = (pred_1+pred_2+pred_3)/3 but it's not clear to me how I get pred_1, 2 and 3. Do I have to run the training stage three times and modify the testing script to load in the three separate models and predict each one or have I missed something?

Thank you for doing so much work, demo Google cloud file is invalid, could you please send me a copy

Hey @hongweilibran ,

Just wanted to express my gratitude for making this available---this is great. I incorporated it into our ANTsRNet and ANTsPyNet packages and have been running some of our data through the pipeline. Looks like everything's working as expected.

Thanks again,
Nick