MS lesions masks outside the SC mask in basel-mp2rage dataset

[Nilser3]

Description

139 subjects of the 180 MS subjects in the basel-mp2rage database have lesion segmentation pixels (lesion-manualNeuroPoly mask) that are outside the SC segmentation

Here an example:

sub-P121

List of subjects with this issue: sub-P001, sub-P003, sub-P004, sub-P005, sub-P006, sub-P007, sub-P010, sub-P011, sub-P012, sub-P013, sub-P014, sub-P016, sub-P019, sub-P021, sub-P022, sub-P023, sub-P024, sub-P025, sub-P026, sub-P027, sub-P028, sub-P030, sub-P031, sub-P033, sub-P034, sub-P035, sub-P036, sub-P037, sub-P038, sub-P039, sub-P040, sub-P041, sub-P042, sub-P044, sub-P045, sub-P047, sub-P050, sub-P051, sub-P053, sub-P055, sub-P056, sub-P057, sub-P058, sub-P059, sub-P061, sub-P062, sub-P063, sub-P065, sub-P066, sub-P067, sub-P068, sub-P069, sub-P070, sub-P071, sub-P072, sub-P075, sub-P078, sub-P079, sub-P082, sub-P084, sub-P085, sub-P086, sub-P087, sub-P088, sub-P094, sub-P095, sub-P096, sub-P097, sub-P098, sub-P099, sub-P100, sub-P101, sub-P103, sub-P104, sub-P106, sub-P108, sub-P109, sub-P110, sub-P111, sub-P112, sub-P114, sub-P116, sub-P117, sub-P119, sub-P120, sub-P121, sub-P122, sub-P124, sub-P125, sub-P129, sub-P130, sub-P132, sub-P134, sub-P138, sub-P139, sub-P141, sub-P142, sub-P144, sub-P145, sub-P146, sub-P148, sub-P149, sub-P151, sub-P153, sub-P156, sub-P157, sub-P160, sub-P161, sub-P165, sub-P167, sub-P169, sub-P170, sub-P173, sub-P174, sub-P175, sub-P176, sub-P178, sub-P179, sub-P180, sub-P181, sub-P182, sub-P183, sub-P185, sub-P187, sub-P188, sub-P190, sub-P191, sub-P192, sub-P194, sub-P197, sub-P199, sub-P200, sub-P241, sub-P242, sub-P243, sub-P244, sub-P246, sub-P249, sub-P250

Maybe we should redefine the SC masks by making a union of the MS lesions and current SC masks

[jcohenadad]

Thank you for spotting this @Nilser3 🙏

In any case, I think that our plan was to re-create all SC segmentations using the contrast-agnostic model. So I would suggest to:

• run the contrast-agnostic model
• manually fix the segmentations is needed
• update the repository with the new segmentations
• the JSON files would include the agnostic model and the manual curation (when it exists)

Does that make sense?

[Nilser3]

Hi @jcohenadad

Thanks to the help of @naga-karthik and @sandrinebedard
I have executed the contrast-agnostic model on the basel-mp2rage and marseille-3T-mp2rage datas

Running inference

Based on: https://github.com/sct-pipeline/contrast-agnostic-softseg-spinalcord/blob/nk/monai/monai/inference_instructions.md#method-1-running-inference-on-a-single-image

python monai/run_inference_single_image.py --path-img /mnt/nvme/nilaia/basel-mp2rage/sub-P121/anat/sub-P121_UNIT1.nii.gz --chkp-path  /mnt/duke/temp/muena/contrast-agnostic/final_monai_model/nnunet_nf\=32_DS\=1_opt\=adam_lr\=0.001_AdapW_CCrop_bs\=2_64x192x320_20230918-2253/ --path-out test_contrast_agnostic  --device cpu

Here the same patient reported above sub-P121:

I have binarized the soft masks with a threshold of 0.5, and my issue persists in 137 patients out of 180 subjects.

I think we should retrain a new SC soft model with these soft masks in mp2rage datas.

[jcohenadad]

 I have binarized the soft masks with a threshold of 0.5, and my issue persists in 137 patients out of 180 subjects. I think we should retrain a new SC soft model with these soft masks in mp2rage datas.

Agreed. Could you please manually correct these 137 patients? Thanks! Good job!

[Nilser3]

To fix the MS lesion masks outside the SC,
I first made a manual correction based on binarized contrast-agnostic SC masks.
Then I made a union of these corrected SC masks with MS lesions (lesion-manualNeuroPoly.nii.gz) masks.

Here is the QC for basel-mp2rage dataset from sub-P001 to sub-P086

Legend of QC maks

• SC_seg.nii.gz -> sct_deepseg_sc
• pred_bin.nii.gz -> contrast-agnostic (version v2.0), binarized with threshold 0.5001
• SC_seg_rater2.nii.gz -> pred_bin.nii.gz corrected by rater 2 (Nilser) + union with MS lesion masks.

[jcohenadad]

I suggest we put a hold on the correction of the masks until this issue is settled: sct-pipeline/contrast-agnostic-softseg-spinalcord#99

[Nilser3]

Exploring the latest version (2024-02-11) of contrast-agnostic model on MP2RAGE data

Model available in : duke/temp/muena/contrast-agnostic/monai_bin_model

For basel-mp2rage dataset (here the QC), I observe that the masks generated by this latest version are closer to those of the GT (manual correction based on version v2.0) for sub-P001 to sub-P086.

Same observations on marseille-3T-mp2rage dataset (here the QC), where even in the presence of large MS lesions, the latest model performs slightly better than the previous one.

Legend of QC:

• contrast-agnostic.V2.0/.../...pred_bin.nii.gz -> contrast-agnostic (version v2.0), binarized with threshold 0.5001
• contrast-agnostic_monai_bin_model/.../...pred.nii.gz -> contrast-agnostic latest version (2024-02-11) using run_inference_single_image.py
• SC_seg_rater2.nii.gz -> contrast-agnostic.V2.0 segmentation corrected by rater 2 (Nilser)

@jcohenadad I still have to manually correct 203 subjects from basel-mp2rageand all nih-ms-mp2rage, maybe I should finish these correctiones based on the latest version instead of version 2.0 ?

[jcohenadad]

@jcohenadad I still have to manually correct 203 subjects from basel-mp2rageand all nih-ms-mp2rage, maybe I should finish these correctiones based on the latest version instead of version 2.0 ?

what do you mean by "finish these corrections based on the latest version instead of version 2.0"? did you mean:

1. run the latest version, and then perform manual correction, or
2. finish manual correction, and then run the latest version (but then, I don't see the logic here)

if you meant "1", then yes, I agree. Also, only do 20 subj and show me for validation before moving forward

[Nilser3]

I meant 1, run the latest version and then perform the manual correction.
perfect, I will do so

[Nilser3]

Hi @naga-karthik

Here the SC manual corrections for subjects (sub-P001 to sub-P086) on basel-mp2rage data

branch: nlm/add_sc_manual_correction
commit: 961a0c0afc753ba960155548e14c227d14703c1b

QC here