Comparative Analysis of Deep Learning Models for Lumbar Vertebrae Segmentation in MRI Images

Abstract

Lower back pain is a widespread health concern globally, leading to significant social and medical expenses. Magnetic Resonance Imaging (MRI) is widely used for evaluating lumbar intervertebral disc degeneration due to its non-invasive nature and superior tissue differentiation capabilities. However, traditional 2D image analysis methods are often hindered by noise and various external factors, complicating accurate diagnosis and surgical planning. To address these challenges, this study investigates the application of deep learning models for lumbar vertebrae segmentation in MRI images. We employ U-Net, HRNet, and EfficientNet architectures, to develop an accurate segmentation model. The U-Net model, characterized by its encoding and decoding phases, demonstrated superior performance with a Precision of 0.9809, Recall of 0.9715, F1-score of 0.9742, and mAP of 0.7084. Comparatively, HRNet and EfficientNet also showed promising results, with HRNet achieving a Precision of 0.6684, Recall of 0.9153, F1-score of 0.77208, and mAP of 0.6568, while EfficientNet achieved a Precision of 0.7995, Recall of 0.9247, F1-score of 0.8491, and mAP of 0.7666. Our findings indicate that deep learning models, particularly U-Net, can significantly enhance the accuracy and efficiency of lumbar vertebrae segmentation in MRI images. This advancement holds potential for improving clinical diagnostics and surgical planning. Future work will focus on refining these models with larger datasets and exploring additional architectures to further enhance segmentation performance and robustness.