This model automatically detects and segments the Multifidus, Psoas, Quadatus Lumborum, and Erector Spinae muscles in T1 axial MRI slices of the lumbar spine. Automatic extraction of these features provides an opportunity to accelerate workflow in studying chronic lower back pain. The 2D VNet Convolutional Neural Network that drives this model was trained on data developed by a skilled technician who was trained by a board-certified radiologist.
Chronic lower back pain is the leading cause of disability worldwide, and manual extraction of paraspinous features for study is expensive and time consuming. This model provides a tool that can greatly accelerate segmentation and biomarker calculation in research settings, saving time and money for researchers.
This model is intended to be used to segment axial slices of lumbar spine from T1 acquisitions so accelerate any workflow that would involve manual segmentation of the paraspinal muscle. It should be used in conjunction with close inspection and post-processing techniques to ensure high quality results. This tool is not intended for diagnostic or prognostic use.
This tool was built and tested on clinical data from GE scanners in research settings. Performance metrics for further generalization have not been confirmed.
Model performance metrics
Dice Similarity Coefficient on Validation Data:
0.86, 0.91, 0.74, 0.90 (Mul, Ps, QL, ES)
Dice Similarity Coefficient on Test Data:
0.81, 0.95, 0.84, 0.92 (Mul, Ps, QL, ES)
Majumdar Lab, Center for Intelligent Imaging, UCSF
The UCSF RBI Center for Intelligent Imaging combines AI research, computational infrastructure and access to the UCSF clinical data record to facilitate quantitative imaging research and to develop cutting edge applications that can be delivered into the clinical workflow to improve patient care, decision support, radiology workflow efficiency and data utilization.