-
Notifications
You must be signed in to change notification settings - Fork 2
Open
Description
I tested the Dataset322_LumbarRootlets
model (more details on this model here) on T2w axial images from the nrc-lumbar-balgrist
dataset.
Commands
NOTE: we have to crop the image first, as the Dataset322_LumbarRootlets
model was trained on cropped images.
cd nrc-lumbar-balgrist/sub-02/anat
# Segment spinal canal
python packaging_lumbar_rootlets/run_inference_single_subject.py -i sub-02_acq-tseTra_chunk-2_T2w.nii.gz -path-model $SCT_DIR/data/deepseg_models/model_seg_canal_t2w/nnUNetTrainer__nnUNetPlans__3d_fullres -fold 0 -o sub-02_acq-tseTra_chunk-2_T2w_canal_seg.nii.gz
# Crop
sct_crop_image -i sub-02_acq-tseTra_chunk-2_T2w.nii.gz -m sub-02_acq-tseTra_chunk-2_T2w_canal_seg.nii.gz -dilate 32x32x32 -o sub-02_acq-tseTra_chunk-2_T2w_crop.nii.gz
# Segment rootlets
python packaging_lumbar_rootlets/run_inference_single_subject.py -i sub-02_acq-tseTra_chunk-2_T2w_crop.nii.gz -path-model $HOME/models/rootlets/Dataset322_LumbarRootlets/nnUNetTrainerDA5__nnUNetPlans__3d_fullres/ -fold 0 -o ub-02_acq-tseTra_chunk-2_T2w_crop_rootlets_322_DA5.nii.gz
GIF - unfortunately, the performance is pretty low
Some rootlets are segmented, but there are also some false positives outside the spinal canal.
I also tried the Dataset302_LumbarRootlets
model (trained on uncropped data), but there were a looot of false positives outside of the spinal canal. Running the models on other subjects (sub-01
and sub-03
) got similar results (not shown).
A few important considerations:
- current
Dataset30X
lumbar models are very much alpha versions, trained on 6 images from a single dataset (lumbar-marseille
) - images from
lumbar-marseille
used to train the models were single-chunk 3T T2* ZOOMit with resolution0.292969, 0.292969, 0.5
covering T10-S2 region - while
nrc-lumbar-balgrist
images are 1.5T T2w TSE with resolutions0.260417, 0.260417, 3.3
to0.446429, 0.446429, 4.6
(i.e., much higher slice thickness) covering L1-L4 acquired in chunks, each chunk covering max 2 discs; example below
Metadata
Metadata
Assignees
Labels
No labels