The link of pretrained model is dead

[zhhsf]

Hi,

Would you please update the download link of multi_site_2d_transformer_Unet_mit_b5_0.81.pth and pmb_2d_transformer_Unet_mit_b5.pth?

Thanks

Hua

[jinghangli98]

I am working on a pypi package as well. Should be released with the model by end of the week

[jinghangli98]

The PyPI package installation and the command line installation should be good now. Let me know if you encounter any issues

[zhhsf]

Hi,

Thank you for the quick update. The code runs smoothly, but the segmentation results are poor in the corpus callosum, basal ganglia, and external capsule areas.

Hua

[jinghangli98]

During training we did not include the lesions on our 7T FLAIR images in the corpus callosum in order to match the segmentation masks provided by the WMH segmentation challenge. For other regions, perhaps you can try turning the fast flag off, which is running inference axially, sagittally, and coronally. For command line usage, I believe the default is doing the prediction in these three views. In python you can set fast to False like so.

from wmh_seg import wmh_seg
import nibabel as nib
img = nib.load('FLAIR.nii').get_fdata()
out = wmh_seg(img,  batch=1, fast=False)

We have another model that can include the hyperintensities in the corpus callosum area. I can update the code to allow the flexibility choosing models.

[zhhsf]

Thanks for your answer! Would you please provide the download link of the model including corpus callosum?