Oral Presentation 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024
An AI-based approach for automated paraspinal muscle analysis and disc identification in patients affected by low back pain (#MP-1f)
Introduction: Low back pain (LBP) is an endemic musculoskeletal condition and a global leading cause of disability1. In the pursuit of enhancing early diagnostic and treatment capabilities, artificial intelligence (AI) has emerged as a transformative tool2. In this study, we developed a fully automated algorithm for disc identification and lumbar paraspinal muscle segmentation from lumbar magnetic resonance imaging (MRI) scans.
Methods:
Automated identification and spatial analysis of intervertebral discs: After Gaussian filtering for noise reduction, a 2D isotropic filter is applied. Column-wise average intensity is calculated, allowing to determine a reference coordinate for the spinal canal by resolving argmax over the intensity columns. The spinal canal coordinates guide a composite function for curve identification. Utilizing intensity, a linear approximation is made for the upper 20% of image rows, followed by a 3rd-degree polynomial for the remaining curve. The resulting shape is used to create a mask, shifting it leftwards to select the region encompassing the vertebrae and discs based on intensity. Extraneous columns are removed. Similarly, the total intensity per pixel row is calculated, identifying intensity peaks (vertebrae) and valleys (discs). Local minimal intensity reveals the vertical coordinates of intervertebral discs. Knowing each disc center facilitates identifying the axial slice corresponding to each disc.
Paraspinal muscle segmentation: The focus then shifts to the strategically chosen L4-L5 level. A U-Net architecture is employed for the automated segmentation of lumbar paraspinal muscles. This sophisticated algorithm processes MRI data, distinguishing between muscle and non-muscle structures. The outcome is a precise delineation of muscle boundaries, forming the foundation for subsequent analyses.
Adipose Infiltration Assessment: The second segmentation task involves distinguishing between muscle and fat subregions within previously predicted regions of interest (ROIs). A visible contrast difference between muscle and fat tissues is leveraged, employing an intensity-based segmentation strategy followed by region-based spatial refinement. Otsu’s thresholding method is applied to the paraspinal ROI. The raw fat mask is then propagated using a recursive empirical region-growing algorithm, exploring boundary pixels for high-contrast regions in a local neighbourhood. This process applies iteratively until no substantial area increase is detected. Hyperparameters are optimized through grid search, enhancing accuracy metrics for the final setup. T2-weighted lumbar MRI scans of 91 subjects were considered for the preliminary investigation, and images of 100 patients from a public dataset were utilized for further validation.
Results: We achieved accurate automated disc identification (100%) and subsequent selection of axial slices. At the L4-L5 level, the U-Net-based muscle segmentation achieved an average DICE score >95% on the internal dataset, and >94% on the public dataset. Adipose infiltration assessments complemented these findings, contributing to a comprehensive understanding of the musculoskeletal dynamics.
Discussion: This detailed methodology integrates cutting-edge AI techniques, revolutionizing LBP management. The automated disc-to-axial slice association streamlines the process, allowing for targeted analysis at the L4-L5 level. The implications for personalized treatment plans and the broader landscape of AI-enabled healthcare interventions will shed new light on LBP treatment.
