Open NathanMolinier opened 3 months ago
Here is a segmentation applying SynthSeg 2.0 model on T1w:
Command line:
mri_synthseg --i whole-spine/sub-amuAL/anat/sub-amuAL_T1w.nii.gz --o brain-seg/sub-amuAL_T1w_brain.nii.gz --robust
Inference time: ~19min per subject
So, do you agree that I apply this model on all whole-spine
dataset (on T1w contrast) and store it in derivatives?
may be with this name?
sub-amuAL_T1w_label-brain_dseg.nii.gz
from SynthSeg 2.0
feedback pls @NathanMolinier @jcohenadad
Great! We also need the skull, fat, skin, muscles, cartilage, eyes-- is that possible?
sub-amuAL_T1w_label-brain_dseg.nii.gz
The name proposed seems to be fine, because i don't think we plan to split the labels into different files.
fat, muscles, cartilage is that possible?
For those, I think we should use the totalseg MRI, but I'm not sure for the rest.
Here a test for the same subject:
total_mr
classes : here
tissue_types_mr
(available with a license)
classes : skeletal_muscle.nii.gz
, subcutaneous_fat.nii.gz
, torso_fat.nii.gz
face_mr
(available with a license)
classes : face.nii.gz
skull
and eyes
:We can explore mideface tool, here a result (cropping with dilation around brain_dseg.nii.gz
) but we need to improve this processing
hum... this is not very encouraging unfortunately 😞.
I'm wondering if we should manually segment what we need in a few subjects, and then fine-tune the total-segmentation-mri model. Do you think it would produce satisfactory results?
With Samseg model we can also segment the skull
and the face
(partially), the only detail is that it also segments MS lesions (we will not take them into account)
Inference time: ~15min per subject
Hum, this is not very convincing... A lot of important tissue is not segmented.
I'm wondering whether we should do some manual segmentation of the skull, eyes and muscles in a few subjects (using threshold-based in-painting in Slicer, that should take a few hours) and then fine-tune the model. In the long term, that will probably save us time. Especially given that all our T1w images have the same contrast (ie: model "easy" to train).
Agree, for threshold-based approach, with ExtractSkin.py script it is possible to do everything in Slicer.
@CharlesPageot do you want to give it a try? @Nilser3 can help you navigate 3D Slicer segmentation modules
Yes for sure, I'll get familiar with the tool and give it a try!
Considering that the bottom of the images is very noisy, I focused on the head-neck region. Also, instead of a threshold-based approach, which requires a lot of manual correction, I used the 'Grow from seeds' tool to segment the body and the different air regions.
These are the general steps that I used for this first subject (sub-amuAL_T1w):
Maybe we group all the labels except the body to make a single air region?
The whole process is about 20-30 min. If this seems correct, I'll proceed with a couple more subjects.
Hi @CharlesPageot
I am currently segmenting all whole-spine
dataset with samseg
the names-intinsities of the output labels can be known with FreeSurferColorLUT.txt file.
I will add these masks to the git-annex.
Perfect, thank you!
I would be interested in helping adding tissue conductivity and permittivity to each compartment, to allow these models to be used for EM simulations such as pTx array and pulse design.
I would be interested in helping adding tissue conductivity and permittivity to each compartment, to allow these models to be used for EM simulations such as pTx array and pulse design.
@jaystock these information should already be available in this other repos: https://github.com/shimming-toolbox/tissue-to-MRproperty. If not, they should be added there. Looping in @sriosq
Hello @jaystock , I can add this 2 options as part of the repo mentioned by prof. Julien. I would only need a list of tissue types and their respective conductivity and permitivity.
Description
The objective would be to generate susceptibility maps of the spine and the brain using segmentations generated for the dataset
whole-spine
.Current state
Currently, segmentations of the:
Are already available in our git-annexed version of the dataset. Additional segmentations of the brain will be created using the model
wmh-synthseg
.