A multiclass deep-learning model with realistic data augmentation for improved 7T gray matter/spinal cord segmentation on patients and controls
Nilser J. Laines Medina 1,2, Charley Gros3,4, Julien Cohen-Adad3,4, Arnaud Le Troter1,2, and Virginie Callot1,2,5
1CRMBM, Aix-Marseille Univ, CNRS, Marseille, France, 2CEMEREM, APHM, CHU Timone, Marseille, France, 3NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montréal, Montreal, QC, Canada, 4MILA, Québec AI Institute, Montreal, QC, Canada, 5iLab-Spine, International Associated Laboratory, Marseille-Montreal, France
Automated methods for WM/GM segmentation in the spinal cord are now largely available. However, these techniques were mostly developed for conventional systems (≤3T) and do not necessarily perform well on 7T MRI data that feature finer details, contrasts, but also different artifacts or signal dropout.
The primary goal of this study was thus to propose a new deep-learning model allowing robust SC/GM multi-class segmentation based on high-resolution 7T T2*-w MR images. The second objective was to highlight the relevance of implementing a realistic hybrid data augmentation strategy to provide better model generalization.
This abstract and the presentation materials are available to members only;
a login is required.