Lesion protocol section

[plbenveniste]

Opening this issue to discuss the creation of a "Lesion protocol" section.

This paper was mentionned in the discussion : (Carass et al. 2024)

This issue is also relevant : #73

Tagging @Nilser3 and @jcohenadad for further discussion

[plbenveniste]

Adding reference to both of these articles, as they don't have the same criteria for lesion segmentation:

• https://pubmed.ncbi.nlm.nih.gov/30679225/
• https://pubmed.ncbi.nlm.nih.gov/38289376/

[plbenveniste]

Here is a first draft of a lesion segmentation/detection protocol: Feedback is welcome !

Lesion segmentation protocol:
The following details the protocol for Multiple Sclerosis (MS) lesion segmentation in the spinal cord.
Imaging the spatial cord is often essential to confirm the diagnosis of MS. That is because the lesions of the spinal cord are included in the McDonald diagnosis criteria, which studies dissemination in space and in time (Thompson et al. 2018). While the MAGNIMS-CMSC-NAIMS working group recommends to use at least two sagittal images for MS diagnosis, still, axial imaging is mentioned as optional in international imaging guidelines (Wattjes et al. 2021).
For detecting MS lesions in the spinal cord, two main contrasts emerge: PSIR and STIR contrasts. New studies (Peters et al. 2024)(Fechner et al. 2019) showed that using PSIR contrasts improved MS lesion detection in the spinal cord. (Fechner et al. 2019) showed that the PSIR contrast showed a higher signal-to-noise (SNR) ratio compared to the STIR contrast.

Criteria to segment MS lesions in the spinal cord:
Do not segment lesions in images with too many artifacts (such as this one : #53 (comment)). Preferably, add the image to the exclude file so that it isn’t used for model training…
When segmenting lesions on thick slices, always look at the above/below slices to build the volume of the lesion (this can minimize partial volume effect).
Do not segment lesions above the first vertebrae (because here we focus only on MS lesions in the spinal cord).
For lesions segmentations which you are not 100% sure, flag the subject and report it for external validation of the segmentation.

How to manually segment lesions:

• MS spinal cord lesions can be manually corrected from the prediction of a segmentation model or manually segmented from scratch. In the first case, make sure to build the json file associated with the segmentation prediction such as detailed in this comment (Create protocol for manual lesion segmentation #73 (comment)).
• For manual correction of the segmentation file use the manual-correction (https://github.com/spinalcordtoolbox/manual-correction) repository. The command can be inspired from this:

python manual_correction.py -path-img ~/data/canproco -config ~/config_seg.yml  -path-label ~/data/canproco/derivatives/labels  -suffix-files-lesion _lesion-manual -fsleyes-dr="-40,70"

• Then a QC should be produced (prefarably using SCT) and added to a Github issue for further validation by other investigators.

• If you are not sure of a subject, it should be flagged on Github for a more open discussion.

More details:
The following section details the different types of errors which occur during lesion segmentation. It is based on the condensed Nascimento Taxonomy:

[jcohenadad]

spatial --> spinal

[jcohenadad]

@plbenveniste maybe it would be easier to do this edit within a PR-- otherwise, for each modification i'm going to create a post, you have to figure out where the issue is, etc. this is sub-efficient

[plbenveniste]

Great idea !
A PR was created in the ms-lesion-agnostic repo : PR 9.