Lumbar spinal stenosis (LSS) is a lumbar disease with a high incidence in recent years. Accurate segmentation of the vertebral body, lamina and dural sac is a key step in the diagnosis of LSS. This study presents an lumbar spine magnetic resonance imaging image segmentation method based on deep learning. In addition, we define the quantitative evaluation methods of two clinical indicators (that is the anteroposterior diameter of the spinal canal and the cross-sectional area of the dural sac) to assist LSS diagnosis. To improve the segmentation performance, a dual-branch multi-scale attention module is embedded into the network. It contains multi-scale feature extraction based on three 3 × 3 convolution operators and vital information selection based on attention mechanism. In the experiment, we used lumbar datasets from the spine surgery department of Shengjing Hospital of China Medical University to evaluate the effect of the method embedded the dual-branch multi-scale attention module. Compared with other state-of-the-art methods, the average dice similarity coefficient was improved from 0.9008 to 0.9252 and the average surface distance was decreased from 6.40 to 2.71 mm.
Keywords: Deep learning; Dual-branch multi-scale attention module; Feature extraction; Lumbar spinal stenosis; Magnetic resonance imaging image.
© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2021.