Poor segmentation quality of TotalSpineSeg in the cervical/thoracic spine #116
Description
Description
For more information about the training, see https://www.researchgate.net/publication/389881289_TotalSpineSeg_Robust_Spine_Segmentation_with_Landmark-Based_Labeling_in_MRI
Due to the time-consuming nature of manually segmenting the anatomical structures identified by TotalSpineSeg namely the spinal cord, spinal canal, vertebrae, and intervertebral discs, the PAM50 atlas was registered to the native space of each scan that included the cervical and/or thoracic regions to serve as ground truth segmentation. However, because of inter-subject anatomical variability, the atlas did not perfectly align with every subject anatomy. These subtle misalignments introduced inconsistencies that propagated into the model’s predictions, often resulting in over- or under-segmentation of certain structures. This issue affects models trained up to and including release r20250224.
Ongoing plan
To solve this issue and improve the quality of the predictions, active learning will be used to retrain TotalSpineSeg.