Quickly and accurately segments an optimally-aligned T1 and T2 pair with a deep neural network trained via nnU-Net and SynthSeg with a large 0 to 8 month old infant MRI brain dataset. Please note that this only runs BIBSnet, not pre- or post-BIBSnet. These will eventually be rolled into the present repository.
usage: run.py [-h] --input INPUT --output OUTPUT [--nnUNet NNUNET]
[--task TASK] [--model MODEL]
optional arguments:
-h, --help show this help message and exit
--input INPUT, -i INPUT
Valid path to existing input directory following valid
nnU-Net naming conventions (T1w files end with
_0000.nii.gz and T2w end with _0001.nii.gz). There
should be exactly 1 T1w file and exactly 1 T2w file in
this directory.
--output OUTPUT, -o OUTPUT
--nnUNet NNUNET, -n NNUNET
Valid path to existing executable file to run nnU-
Net_predict. By default, this script will assume that
nnU-Net_predict will be in the same directory as this
script:
--task TASK, -t TASK Task ID, which should be a 3-digit positive integer
starting with 5 (e.g. 512).
--model MODEL, -m MODEL
When running CABINET using a GPU, the job typically takes about 4 minutes, 2 tasks, and one node with 20 gb of memory to run effectively.
This has been primarily tested in Singularity. We are less able to provide technical support for Docker execution.
singularity pull docker://dcanumn/BIBSNet
singularity run --nv --cleanenv --no-home \
-B /path/to/input:/input \
-B /path/to/output:/output \
/path/to/BIBSNet.sif \
--input /input --output /output --task <task ID> --model 3d_fullres
For choosing the model, be sure to choose the model according to what anatomicals you are providing. The task ID you provide with the --task flag needs to correlate with the model number.
data/models.csv lists all available BIBSnet models to run. Below are the default BIBSnet models, all trained on manually-segmented 0- to 8-month-old BCP subjects' segmentations.
| Model | Description |
|---|---|
| 512 | Model Used within BIBSNet manuscript |
| 551 | Default T1w and T2w model |
| 514 | Default T1w-only model |
| 515 | Default T2w-only model |
Additionally, see the "location" column within data/models.csv to download the models. The models are publically available and can be pulled down with CLI tools such as wget. For example to pull down model 512 something like this work: wget https://s3.msi.umn.edu/CABINET_data/Task512_BCP_ABCD_Neonates_SynthSegDownsample.zip
If you provide an input directory with both a T1w and a T2w, BIBSnet expects them to be aligned. If they are not aligned, the outputted segmentation will be inaccurate. It is also recommended to crop the input images just below the head to ensure the best results. Input files must be identified with _0000 and _0001 at the end of the file names for T1w and T2w, respectively (e.g. a T1w image needs to be named something like sub-XXX_ses-baseline_T1w_0000.nii.gz). If there is only one anatomical image, regardless of it being T1w or T2w, the file will need to be named with _0000 at the end. Pre-BIBSnet (usage here) performs each of these preprocessing steps. You will need to average images of the same anatomical file type if you have more than one T1w or T2w image.
Outputs may need to be chirality corrected, which can be done with post-BIBSnet. The usage for post-BIBSnet can be found here. The outputs will be in the space that the anatomical was aligned with, i.e. if the T2w was aligned to the T1w, the output will be in T1w space and vice versa. If only one file type is inputted, the output will be in the space of that file.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent