BrainMaGe
BrainMaGe copied to clipboard
CBICA
Are you required to pad images to dimension (240, 240, 160)?
All of my images are unit volume. RAS orientation. BTW: I've tried RAS per documentation and also RAI and see no difference.
I'm getting reasonably good masks running
python ../BrainMaGe/brain_mage_single_run -i $input_brain_path -o 'output_mask.nii.gz' -m 'output_brain.nii.gz' -dev 0
However, when I run batch mode:
python ../BrainMaGe/brain_mage_run -params test_params_multi_4_2020.cfg -test True -mode Multi-4 -dev 0
I receive the following error: `Weight file used : /home/BrainMaGe/weights/resunet_multi_4.pt ../BrainMaGe/brain_mage_run
Hostname :None
Start Time :Tue Sep 21 11:26:16 2021
Start Stamp:1632223576.6366262
Generating Test csv
100%|███████████████████████████████████████████| 61/61 [11:05<00:00, 10.91s/it]
Done with running the model.
You chose to save the brain. We are now saving it with the masks.
0%| | 0/61 [00:00<?, ?it/s]
Traceback (most recent call last):
File "../BrainMaGe/brain_mage_run", line 215, in
Here is line 157:
image_data[mask_data == 0] = 0
https://github.com/CBICA/BrainMaGe/blob/master/BrainMaGe/tester/test_multi_4.py#:~:text=image_data%5Bmask_data%20%3D%3D%200%5D%20%3D%200
It is dependent on line 126. It looks like the images are hard-code to be (240, 240, 160)
to_save = interpolate_image(output, (240, 240, 160))
https://github.com/CBICA/BrainMaGe/blob/master/BrainMaGe/tester/test_multi_4.py#:~:text=to_save%20%3D%20interpolate_image(output%2C%20(240%2C%20240%2C%20160))
Am I required to pad images to dimension (240, 240, 160) to use multi_4 mode?
Any guidance would be appreciated.
Thanks, Jay
Have you passed the images through these steps https://github.com/CBICA/BrainMaGe/#steps-to-run-application? For the multi-modality segmentation, the images do need to pass through these steps.
However, we do suggest using 'brain_mage_single_run' when possible.
Yes, I followed the steps. Note: I did not use the Brain Imaging Toolkit. We have an existing pipeline where all of the images are registered.
I also ensured the orientation is RAI and normalize the images to unit volume. The dimensions of the preprocessed images will vary.
As I indicated earlier, 'brain_mage_single_run' did work for me. But I have a lot of images to skull strip.
The index error seems to indicate a sizing problem. So my basic question remains. Does multi_4 mode require images of a specific dimensionality?
Thanks.
The dimensions of the preprocessed images will vary.
This is the issue that step 1 in https://github.com/CBICA/BrainMaGe/#steps-to-run-application tackles. The output of our preprocessing pipeline ensures that the images are in the space the model expects.
Cheers.
"1. Co-registration within patient to the SRI-24 atlas in the LPS/RAI space."
Yes, I have done this.
Are there any additional specifics like the dimension of the image?
@jayurbain , The dimensions of the image should also match [240, 240, 160] as mentioned in the paper. The brain_mage_single_run should also work for all the cases individually.
Hey @jayurbain,
The padding should happen automatically by our preprocessing function, but we are debugging this and will get back to you.
Cheers, Sarthak
Thanks. I did find a problem with how our images are standardized and corrected it. For now, the single run agnostic approach is working well and I haven't found it to be that slow. ~10sec with GPU and I use the same mask on all 4 registered series. When I get a chance I will retest for Multi-4.
Hey Jay,
To be honest, we use the modality-agnostic model for all our pipelines, so you should be fine! In any case, @Geeks-Sid is debugging and he should reach out here once he is ready.
Cheers, Sarthak