With the rising incidence of degenerative lumbar spine disorders, accurate segmentation of spinal structures based on magnetic resonance imaging (MRI) is crucial for intelligent clinical diagnosis and surgical planning, while automated measurement of the lumbar lordosis angle based on segmentation can further support quantitative assessment of the condition. To overcome the issues of limited receptive field and loss of detail in existing deep learning methods when processing low-resolution, edge-blurred images, this paper proposes a segmentation architecture called MDVM-UNet. This architecture integrates three complementary mechanisms: the VSS module constructs a computationally efficient, multi-scale collaborative receptive field through parallel multi-scale hole convolution; the dual-path fusion module performs feature alignment using global average pooling and channel-spatial attention; and the edge enhancement module sharpens blurred boundaries through reverse attention and Laplacian pyramid decomposition. Experiments conducted on both private and public lumbar MRI datasets demonstrate that this method achieves excellent segmentation results for vertebral bodies and intervertebral discs, with an average Dice coefficient of 0. 943, a 6. 4% improvement over the standard U-Net;The mean absolute error between the lumbar lordosis angle measured automatically based on segmentation results and the manually annotated measurements was 2. 23^, which is below the clinically acceptable threshold. Overall, the method described in this paper demonstrates excellent performance in both segmentation accuracy and clinical quantification, offering a viable approach for the intelligent assessment of degenerative lumbar spine diseases.
Ji et al. (2026) studied this question.