Poster Presentation Sydney Spinal Symposium 2023

SpineQ 3D: The fully automated 3D quantitative assessment of lumbar spine (#20)

Xihe Kuang 1 , Jason Cheung 1 , Tao Huang 1 , Teng Zhang 1
  1. the University of Hong Kong, Hong Kong, HONG KONG, China

Introduction: 

Low back pain (LBP) is a common health problem, with a lifetime incidence of 80%. It is believed that LBP results from pathological changes that occur with lumbar degenerative diseases (LDD). Currently, the clinical diagnosis and treatment planning of LDD is usually done manually which is inefficient and inconsistent. Besides, based on the 2D slice of MRI or CT, the complex 3D parameters are difficult to measured accurately, which may need to across different slices. Therefore, the automated 3D quantitative analysis can have great significance to improve the efficiency, consistency, and accuracy of diagnosis and treatment planning.

Methods:

A dataset was esteblished based on an LDD cohort from the southern Chinese population1, which contained sagittal and axial MRI scans from 2473 subjects (mean age 45.2; 39.5% male). Our deep learning pipeline adopted the Spine-GFlow2, a robust unsupervised multi-tissue segmentation framework, that could accurately identify different anatomical structures from lumbar MRI without relying on any manual annotation. Further, based on the segmentation result, multiple parameters, including anteroposterior (AP) vertebral body (VB) diameter, midline VB width, mid-AP canal diameter, canal width, mid-AP dural sac (DS) diameter, pedicle width, lamina angle, and facet joint angle, were measured using the 3D symmetrical boundary searching and knowledge-based distance retrieve algorithm. The automated measurement accuracy was validated by comparing it with the manual measurement annotated by a spine specialist with over 20 years of clinical experience.

Results:

Preliminary validation showed that the deep learning pipeline achieved satisfactory performances on the measurement. For the distance parameters, the average absolute error was 3.721mm/4.538pix, and for the angle parameters, the average absolute error was 4.891 degree.  

Conclusions:

A deep learning pipeline for fully automated 3D quantitative assessment of lumbar is developed and tested. The fast and consistent 3D parameter measurement can assist clinicians in efficient and consistent diagnosis and treatment planning. The preliminary validation shows that our method can achieve good performance on the measurement of multiple parameters without relying on any human intervention. A prospective clinical study needs to be performed for further validation.

 

  1. [1] Samartzis D, Karppinen J, Chan D, Luk KD, Cheung KM (2012). The association of lumbar intervertebral disc degeneration on magnetic resonance imaging with body mass index in overweight and obese adults: a population-based study. Arthritis Rheum 64:1488-1496. doi: 10.1002/art.33462
  2. [2] Kuang, X., Cheung, J. P. Y., Wong, K. Y. K., Lam, W. Y., Lam, C. H., Choy, R. W., ... & Zhang, T. (2022). Spine-GFlow: A hybrid learning framework for robust multi-tissue segmentation in lumbar MRI without manual annotation. Computerized Medical Imaging and Graphics, 99, 102091.