Automate detection of spinal cord compression

[valosekj]

Background

Spinal cord compression is highly prevalent in the elderly, and its severity is considered in clinical decision-making. Currently, the evaluation of the compression is done manually by radiologists. Such manual evaluation is time-consuming and introduces inter-rater and inter-trials variability.

Recently, we showed that the logistic model combining morphometric metrics such as cross-sectional area (CSA), solidity, compressive ratio (CR), and torsion computed from T2*-w axial image could predict spinal cord compression automatically. For details, see the paper.

Methods

It would be great to automate the process of compression detection fully. Ideally to be run by a single command. This would include the following:

• obtaining SC segmentation and SC labeling, including quality control with potential manual corrections
• computation of the morphometric metrics (i.e., CSA, CR solidity, torsion) across individual slices by the sct_process_segmentation function.
• running the predictive model to detect the slice with the compression
• determine which intervertebral disc the slice corresponds to (based on the SC labeling)
• provide commonly used radiological metrics for the compression level (i.e., CSA, CR) as an output

I would be glad for any suggestions or ideas.

Related to:

• Train model to detect cord compression spinalcordtoolbox/spinalcordtoolbox#4333
• Develop a grading model for cervical cord compression ivadomed/ivadomed#762

[valosekj]

UPDATE: now implemented in sct_detect_compression

I implemented the first draft; see here

Steps to run:

1. Download the attached zip archive containing T2*-w image, SC seg, and manually created disc labels for a DCM patient with compression at C5/6 disc.
   example_data-compression_detection.zip

2. Run the following code:

cd ${SCT_DIR}
git pull
git checkout jv/add_compression_detection
source ${SCT_DIR}/python/etc/profile.d/conda.sh
conda activate venv_sct

cd <downloaded_data>
python ${SCT_DIR}/spinalcordtoolbox/scripts/sct_detect_compression.py -s t2s_seg.nii.gz -discfile labels.nii.gz
# -s - segmentation
# -discfile - disc labels

Output:

$ python ${SCT_DIR}/spinalcordtoolbox/scripts/sct_detect_compression.py -s t2s_seg.nii.gz -discfile labels.nii.gz

--
Spinal Cord Toolbox (git-jv/add_compression_detection-d89f8a3de5253cdca5c5041a467422aa8214fa24)

sct_detect_compression -s t2s_seg.nii.gz -discfile labels.nii.gz
--

Compute shape analysis: 100%|#################| 42/42 [00:00<00:00, 85.02iter/s]
Compression probability for disc 4 (corresponding to slice 32) is 0.226.
Compression probability for disc 5 (corresponding to slice 27) is 0.014.
Compression probability for disc 6 (corresponding to slice 22) is 0.984.
Compression probability for disc 7 (corresponding to slice 16) is 0.257.

Compression(s) was detected at:
	disc 6 (corresponding to slice 22). CSA = 54.91 mm2. CR = 43.22.

A few notes for a discussion:

• metrics normalization
  The predictive model is based on quantitative metrics, which have naturally different anatomy across vertebral levels (e.g., smaller CSA at C6/7 compared to C3/4). This means that the information about vertebral levels must be taken into account during the compression prediction. So far, I have implemented perlevel metrics normalization by normative values computed from healthy volunteers within Horakova et al., 2022. Potentially, some more advanced normalization approach also utilizing other variables (such as age, sex, sequence parameters) should be explored. See below.

• disc identification
  Since the axial ME-GRE image usually covers only a limited field of view, fully automatic labeling (=disc identification) might be challenging. In such cases, users can a) provide the initial C2/3 label (sct_label_vertebrae -initz), b) do the whole labeling manually, c) use labeling from some other image, e.g., from T1w image as proposed in spine-generic.

Idea:

• Would it make sense to replace the current predictive model based on quantitative metrics by a deep learning model? The DL model could be conditioned by metadata (compression level, age, sex, sequence parameters, ...). Could this conditioning serve instead of normalization?

[valosekj]

Few comments from Julien

• For the DL method, look at the recent paper from the Fehlings group ref, repo
• About the normalization: we've been discussing some strategies with @sandrinebedard link to GH issue

[valosekj]

In progress within the dcm-metric-normalization project. UPDATE: detect-compression project

Relevant discussions:

• Extend -create-viewer to allow users to create compression disc labels (all 1 values) spinalcordtoolbox/spinalcordtoolbox#3984
• Add normalized metrics for degenerative cervical myelopathy (DCM) spinalcordtoolbox/spinalcordtoolbox#3733
• Add new CLI script to compute normalized metric ratios (MSCC, etc.) for compressed levels spinalcordtoolbox/spinalcordtoolbox#4003
• Add script for compression detection spinalcordtoolbox/spinalcordtoolbox#3946