New dataset hc-lumbar-shanghai

[Nilser3]

Description

I would like to push a new dataset hc-lumbar-shanghai to git-annex.

Is a Open Dataset available here, from this paper: An open-access lumbosacral spine MRI dataset with enhanced spinal nerve root structure resolution

Content:

anat: using constructive interference in steady state (CISS) sequence and double echo steady state (DESS) sequence for lumbar images, and T2-weight turbo spin echo (T2-TSE) sequence for whole-spine.

Tagging @NathanMolinier

GT: .stl and .json files for lumbar-rootlets annotations

[mguaypaq]

I created the dataset, and gave write access to @Nilser3:
https://data.neuro.polymtl.ca/datasets/hc-lumbar-shanghai

[valosekj]

@Nilser3, thank you for pushing the dataset to git-annex!

GT: .stl and .json files for lumbar-rootlets annotations

In the long term, it would be great if we find a way to convert stl GTs to nii so that they can used for DL models training.

[Nilser3]

The .stl files was converted to .nii.gz files @valosekj

PR was created ! @mguaypaq

[valosekj]

Note

This comment was originally posted on data.neuro.polymtl.ca. I'm copy-pasting it here for more visibility.

---

Hey @nilaia, thanks for pushing the dataset and converting .stl files to .nii! 💯

It seems that the labels under derivatives/labels correspond to acq-axCISS3d_T2w.nii.gz images, right?

screenshot

Perhaps we could add this BIDS tag to labels names. For example:
sub-02_label-canal_seg.nii.gz --> sub-02_acq-axCISS3d_T2w_label-canal_seg.nii.gz

What do you think?

---

I also noticed that label files have different dimensions than the source acq-axCISS3d_T2w.nii.gz files. Might it caused by the .stl to .nii conversion?

GIF

---

Also, the rootlets masks are now binary. But based on Fig 1 in the dataset publication, it seems that .stl files are "rootlets-specific" (i.e., each rootlet level has its own annotation). Do you know if it is possible to keep this "rootlets-specific" information when converting .stl to .nii?

[Nilser3]

Note

This comment was originally posted by @Nilser3 on data.neuro.polymtl.ca. I'm copy-pasting it here for more visibility.

---

Hi @valosekj

I apologize for my late response,

Perhaps we could add this BIDS tag to labels names. For example:
sub-02_label-canal_seg.nii.gz --> sub-02_acq-axCISS3d_T2w_label-canal_seg.nii.gz

What do you think?

The acq-axCISS3d_T2w.nii.gz image has a resolution of 0.3x0.3x2.0 and label-canal_seg.nii.gz 0.1945x0.1945x0.1945x0.1945, so I wanted to keep a segmentation at a fine resolution, so I didnt use the anat file suffix for not assume that they are at the same resolution and dimension (same with the other masks)
What do you think?

I also noticed that label files have different dimensions than the source acq-axCISS3d_T2w.nii.gz files. Might it caused by the .stl to .nii conversion?

It is possible, since when converting .stl to .nii.gz (using Slicer), I didn't need any .nii.gz image reference, so it was converted to the same resolution saved in the .stl header.

Also, the rootlets masks are now binary. But based on Fig 1 in the dataset publication, it seems that .stl files are "rootlets-specific" (i.e., each rootlet level has its own annotation). Do you know if it is possible to keep this "rootlets-specific" information when converting .stl to .nii?

The shared .stl belong to a single class (binaries), but I think we could make a rootlets-specific with the .json information that we have in derivatives/markers/sub-02/sub-02/sub-02_nerveroots_L1_L.json.

For example here I have loaded the raw .stl (as model) and the .json (as markerks) using Slicer and I manage to have the specific rootlet.

screenshot

I think I can make rootlet-specific masks using this information.
However what intensity do I give to each rootlet?

L1 left = 1 , L1 right = 2 ,L2 left = 3 , L2 right = 4 ?

[valosekj]

The acq-axCISS3d_T2w.nii.gz image has a resolution of 0.3x0.3x2.0 and label-canal_seg.nii.gz 0.1945x0.1945x0.1945x0.1945, so I wanted to keep a segmentation at a fine resolution, so I didnt use the anat file suffix for not assume that they are at the same resolution and dimension (same with the other masks)
What do you think?

I see! Indeed, you're right -- although derivatives/labels/sub-02/anat/sub-02_label-canal_seg.nii.gz visually corresponds to sub-02/anat/sub-02_acq-axCISS3d_T2w.nii.gz, their resolution (and also image size) is different:

$ header derivatives/labels/sub-02/anat/sub-02_label-canal_seg.nii.gz | grep pixdim
pixdim		[-1.0, 0.194471, 0.194471, 0.194471, 1.0, 1.0, 1.0, 1.0]
$ header sub-02/anat/sub-02_acq-axCISS3d_T2w.nii.gz | grep pixdim
pixdim		[-1.0, 0.3, 0.3, 1.999984, 1.0, 1.0, 1.0, 1.0]

There seems to be no anatomical image with pixdim of 0.194471, 0.194471, 0.194471 corresponding to sub-02_label-canal_seg.nii.gz, so I'm wondering whether we will actually be able to effectively use this "fine" resolution as we do not any corresponding image for the label file. Maybe we could try to resample sub-02_acq-axCISS3d_T2w.nii.gz to the "fine" resolution when training a model.

anatomical images mixdim

$ hc-lumbar-shanghai/sub-02/anat$ for f in *.nii.gz; do echo $f;pixdim $f;done
sub-02_acq-axCISS3d_T2w.nii.gz
pixdim		[-1.0, 0.3, 0.3, 1.999984, 1.0, 1.0, 1.0, 1.0]
sub-02_acq-corDE3d_T2w.nii.gz
pixdim		[-1.0, 1.265625, 1.27, 1.265625, 1.0, 1.0, 1.0, 1.0]
sub-02_acq-sag_T2w.nii.gz
pixdim		[-1.0, 3.300003, 0.625, 0.625, 1.0, 1.0, 1.0, 1.0]

---

It is possible, since when converting .stl to .nii.gz (using Slicer), I didn't need any .nii.gz image reference, so it was converted to the same resolution saved in the .stl header.

I see! But again, this is not ideal for future analysis, model training, etc., as most of the tools expect corresponding dim and pixdim between the image and the label file.

---

For example here I have loaded the raw .stl (as model) and the .json (as markerks) using Slicer and I manage to have the specific rootlet.

I think I can make rootlet-specific masks using this information.

This would be great!

However what intensity do I give to each rootlet?
L1 left = 1 , L1 right = 2 ,L2 left = 3 , L2 right = 4 ?

I would suggest the following to be consistent with lumbar-marseille and cervical rootlets.

{"C1": 1, "C2": 2, "C3": 3, "C4": 4, "C5": 5, "C6": 6, "C7": 7, "C8": 8, 
"T1": 9, "T2": 10, "T3": 11, "T4": 12, "T5": 13, "T6": 14, "T7": 15, "T8": 16, "T9": 17, "T10": 18, "T11": 19, "T12": 20, 
"L1": 21, "L2": 22, "L3": 23, "L4": 24,"L5": 25, "S1": 26, "S2": 27, "S3": 28, "S4": 29,"S5": 30}

[Nilser3]

Thanks for your feedback @valosekj
I am working on it:

[valosekj]

This is amazing, Nilser! Would you mind recording your screen when doing this encoding in Slicer? (we can then put the recording to our G Drive)