2D Residual U-Net for Accurate Lumbar Vertebrae Segmentation in Mri-Based Low Back Pain Diagnosis Using the Spider Dataset Publisher



Borgi AR ; Zohrabi A ; Kazemi A ; Abdolghaffar M ; Kordi R ; Farnia P ; Ahmadian A
Source: 32nd National and 10th International Iranian Conference on Biomedical Engineering, ICBME 2025 Published:2025

Abstract

Low back pain affects roughly 12% of people worldwide and continues to be a principal cause of disability. Precise visualization of lumbar vertebrae and intervertebral discs is critical for detecting pathological changes and guiding clinical interventions. Robust structure segmentation is essential to ensure trustworthy diagnosis and informed treatment plans. Compared to X-ray or CT scans, magnetic resonance imaging (MRI) provides superior soft-tissue contrast, making it the preferred modality for comprehensive spinal assessments. However, manual segmentation of vertebrae is labor-intensive and prone to inter-observer variability. At the same time, semiautomatic approaches are often time-consuming and lack robustness in accurately identifying vertebral anatomical structures, particularly when applied to low-quality or diverse clinical MRI data. In this study, we propose a 2D Residual UNet for vertebral segmentation on the SPIDER dataset. The pipeline includes reorientation, resolution standardization, and morphological mask refinement, along with a hybrid Dice-Binary Cross-Entropy loss to address class imbalance, particularly in non-vertebral structures. The proposed model achieved a Dice score of 0.946 and an IoU of 0.897, slightly surpassing a standard 2D U-Net with Dice =0.931 and IoU = 0.823. These results demonstrate that accurate 2D segmentation enables reliable 3D reconstruction, providing an efficient and clinically applicable solution for spinal analysis and LBP diagnosis. © 2025 IEEE.