A Deep-Learning–Based, Fully Automated Program to Segment and Quantify Major Spinal Components on Axial Lumbar Spine Magnetic Resonance Images

Abstract

Objective The paraspinal muscles have been extensively studied on axial lumbar magnetic resonance imaging (MRIs) for better understanding of back pain; however, the acquisition of measurements mainly relies on manual segmentation, which is time consuming. The study objective was to develop and validate a deep-learning–based program for automated acquisition of quantitative measurements for major lumbar spine components on axial lumbar MRIs, the paraspinal muscles in particular. Methods This study used a cross-sectional observational design. From the Hangzhou Lumbar Spine Study, T2-weighted axial MRIs at the L4–5 disk level of 120 participants (aged 54.8 y [SD = 15.0]) were selected to develop the deep-learning–based program Spine Explorer (Tulong). Another 30 axial lumbar MRIs were automatically measured by Spine Explorer and then manually measured using ImageJ to acquire quantitative size and compositional measurements for bilateral multifidus, erector spinae, and psoas muscles; the disk; and the spinal canal. Intersection-over-union and Dice score were used to evaluate the performance of automated segmentation. Intraclass coefficients (ICCs) and Bland–Altman plots were used to examine intersoftware agreements for various measurements. Results After training, Spine Explorer (Tulong) measures an axial lumbar MRI in 1 second. The intersections-over-union were 83.3% to 88.4% for the paraspinal muscles, 92.2% and 82.1% for the disk and spinal canal, respectively. For various size and compositional measurements of paraspinal muscles, Spine Explorer (Tulong) was in good agreement with ImageJ (ICC = 0.85 ~ 0.99). Impact Spine Explorer (Tulong) is automated, efficient, and reliable in acquiring quantitative measurements for the paraspinal muscles, the disk, and the canal, and various size and compositional measurements were simultaneously obtained for the lumbar paraspinal muscles. Such an automated program might encourage further epidemiological studies of the lumbar paraspinal muscle degeneration and enhance paraspinal muscle assessment in clinical practice.

An externally validated deep learning model for the accurate segmentation of the lumbar paravertebral muscles

“…Xia and colleagues [11] tested several network architectures, including the U-Net as well as novel solutions, to segment psoas major, erector spinae, and multifidus, obtaining excellent Dice similarity scores between 0.913 and 0.95 for the best-performing neural network. Other studies achieved relatively lower performances [21], or focused on different imaging modalities or fields of view [22,23]. In general, it can be concluded that our model achieved state-of-the-art performance and is the only one that has been externally validated on a purposely created dataset so far.…”

Artificial Intelligence and Computer Vision in Low Back Pain: A Systematic Review

“…This study demonstrated the possibility to extract accurate information about soft tissues from spine CT without the necessity to order an MRI, which is often expensive and time-consuming. Other studies have also shown the possibility to automatically calculate the spinal canal area [73] as well as segmenting and reconstructing multiple structures at the same time [47,66,72,75,79] with an elevate degree of accuracy.…”

“…Kim et al [71] exploited a BSU-net for IVDs segmentation on 20 MRI from the SpineWeb dataset, achieving a DICE of 89.4%. Shen et al [72] used a Feedforward NN on MRI of 120 subjects, achieving a Jaccard index for the segmentation of IVDs, spinal canal and muscles of 87, 82 and 85%, respectively. Gaonkar et al [73] applied a U-net to segment IVDs on 39295 MRI images, achieving an 88% DICE; they also combined an SVM with a Regression Tree to segment the spinal canal with a DICE of 87%.…”

Multi-muscle deep learning segmentation to automate the quantification of muscle fat infiltration in cervical spine conditions

“…The CNN markedly improved the efficiency of the segmentation, reducing the processing time from 20 min per image to only seconds. In addition to our work, CNN's have been used to automate the segmentation of the lumbar paraspinal and iliopsoas muscles from T 1 -weighted and Dixon MRI scans, respectively 25,26 . CNN's have also been applied to other spinal structures, allowing for the automatic quantification of vertebrae and intervertebral disc morphology from MRI 27 .…”