Comparison of manually segmented lesion and predicted lesion segmentation

[plbenveniste]

In this issue, we look at the predicted lesion segmentations from sct_deepseg lesion_ms.
To perform this inference, we use: git-plb/4944_add_new_ms_model-4df0921dcb73b267f03e1cde23d24f36b2bc53c5. This is the current development branch where the newest model has been uploaded.

Inference is currently running on joplin using: sct_run_batch -script batch_processing.sh -path-data <PATH_TO_BIDS_ROOT> -path-output <PATH_RESULTS>. QC will be added here as soon as it is available.

[plbenveniste]

The QC output is stored here: ~/duke/temp/plben/ms-critical-lesion/20250707_Batch_processing

It seems that most lesions are detected. Lesion segmentation can however be improved.

[plbenveniste]

Some processing crashed with this error:

Spinal Cord Toolbox (git-plb/4944_add_new_ms_model-4df0921dcb73b267f03e1cde23d24f36b2bc53c5)

sct_deepseg lesion_ms -i sub-011_acq-ax_T2w.nii.gz -o sub-011_acq-ax_T2w_label-lesion_seg.nii.gz -qc /home/GRAMES.POLYMTL.CA/p119007/critical_lesions/output_run_batch/qc -qc-subject sub-011 -qc-seg sub-011_acq-ax_T2w_seg.nii.gz
--

Model 'model_seg_ms_lesion' is up to date (Source: https://github.com/ivadomed/ms-lesion-agnostic/releases/download/r20250626/20250626_model_fold0_ESMRMB_abstract.zip)
perform_everything_on_device=True is only supported for cuda devices! Setting this to False
Running inference on device: cpu
Model loaded successfully.
Creating temporary folder (/tmp/sct_2025-07-04_11-31-58_sct_deepseg_bewf7pxc)
Copied sub-011_acq-ax_T2w.nii.gz to /tmp/sct_2025-07-04_11-31-58_sct_deepseg_bewf7pxc/sub-011_acq-ax_T2w.nii.gz
Orientation (based on dataset.json): RPI
Changing orientation of the input to the model orientation (RPI)...
Starting inference...

  0%|          | 0/1 [00:00<?, ?it/s]
100%|██████████| 1/1 [00:17<00:00, 17.79s/it]
100%|██████████| 1/1 [00:17<00:00, 17.79s/it]
Inference done.
Total inference time: 0 minute(s) 24 seconds
Reorienting the prediction back to original orientation...
Reorientation to original orientation RAI done.
�[33mImage header specifies datatype 'uint16', but array is of type 'uint8'. Header metadata will be overwritten to use 'uint8'.�[0m
Saving results to: /tmp/sct_2025-07-04_11-31-58_sct_deepseg_bewf7pxc/nnUNet_prediction/sub-011_acq-ax_T2w_pred_lesion_seg.nii.gz

*** Generating Quality Control (QC) html report ***
Resample images to 0.600000x0.600000 vox
�[33mConverting image from type 'uint16' to type 'float64' for spline interpolation�[0m
�[33mConverting image from type 'uint8' to type 'float64' for linear interpolation�[0m
�[33mConverting image from type 'uint8' to type 'float64' for linear interpolation�[0m
Bounding box: x=[1, 40], y=[0, 233], z=[0, 233]
Cropping the image...
Bounding box: x=[1, 40], y=[0, 233], z=[0, 233]
Cropping the image...
�[33mImage header specifies datatype 'float64', but array is of type 'int64'. Header metadata will be overwritten to use 'int64'.�[0m
Bounding box: x=[1, 40], y=[0, 233], z=[0, 233]
Cropping the image...
Find the center of each slice
/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/python/envs/venv_sct/lib/python3.10/site-packages/scipy/ndimage/_measurements.py:1548: RuntimeWarning: invalid value encountered in scalar divide
  results = [sum_labels(input * grids[dir].astype(float), labels, index) / normalizer
Traceback (most recent call last):
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/scripts/sct_deepseg.py", line 537, in <module>
    main(sys.argv[1:])
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/scripts/sct_deepseg.py", line 512, in main
    qc2.sct_deepseg(
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/reports/qc2.py", line 277, in sct_deepseg
    sct_deepseg_axial(
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/reports/qc2.py", line 353, in sct_deepseg_axial
    img = equalize_histogram(mosaic(img_input, centers, radius))
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/reports/qc2.py", line 819, in mosaic
    cropped.append(np.pad(slice, [[r] for r in radius])[
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/python/envs/venv_sct/lib/python3.10/site-packages/numpy/lib/arraypad.py", line 745, in pad
    raise TypeError('`pad_width` must be of integral type.')
TypeError: `pad_width` must be of integral type.
Total runtime; 35.720 seconds.

Or this error:

Traceback (most recent call last):
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/scripts/sct_deepseg.py", line 537, in <module>
    main(sys.argv[1:])
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/scripts/sct_deepseg.py", line 512, in main
    qc2.sct_deepseg(
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/reports/qc2.py", line 277, in sct_deepseg
    sct_deepseg_axial(
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/reports/qc2.py", line 339, in sct_deepseg_axial
    crop_with_mask(img_to_crop, img_qc_seg, pad=3)
  File "/home/GRAMES.POLYMTL.CA/p119007/spinalcordtoolbox/spinalcordtoolbox/reports/qc2.py", line 1092, in crop_with_mask
    first_slice = min(np.where(img_ref.data)[axis])
ValueError: min() arg is an empty sequence
Total runtime; 33.221 seconds.

This should be investigated and mentioned in SCT issues:

[plbenveniste]

This should be investigated and mentioned in SCT issues:

Issue was opened: spinalcordtoolbox/spinalcordtoolbox#4951

[plbenveniste]

Investigation into the quality of some predictions:

In today's meeting, we (@jcohenadad and I) looked at sub-012. It seems that again there is a problem when stitching the upper and lower parts of the image, creating a disaligned spinal cord.

I am not excluding this image as it does not have an impact on the SC.

[plbenveniste]

As for sub-007, we saw that the segmentation was of pretty poor quality. While I cannot find an explanation, here is an improved result by simply applying a denoising of the image (sct_denoising_onlm -p Rician)

Here is the comparison of the results:

[jcohenadad]

@plbenveniste I see no difference between the denoised and the non denoised #10 (comment)

[plbenveniste]

@plbenveniste I see no difference between the denoised and the non denoised #10 (comment)

You have to look at the outer parts of the images on the first GIF.

[jcohenadad]

You have to look at the outer parts of the images on the first GIF.

ok, now i see it but it is veeeeeeeery subtle.

[plbenveniste]

ok, now i see it but it is veeeeeeeery subtle.

Yes! That's one more reason why I am pretty surprised about the large difference in result.

[jcohenadad]

you might want to play around with denoising parameters then, be a bit more aggressive