SCPA-MEUM U-Net: A High-Precision Segmentation Method for Multifidus Muscle MRI

An improved U-Net model was proposed to solve the problems of low precision and blurred boundary of MRI images of multifidus muscle in patients with lumbar disc herniation (LDH). Based on encoder-decoder structure, the model enhances the representation of anatomical features and improves the accuracy of boundary reconstruction. In the coding stage, the Spinal Channel Prior Attention (SCPA) module is introduced to guide multi-scale spatial feature extraction with the channel prior information, thereby improving the recognition ability of the neural network for low-contrast muscle structures. In the decoding phase, a multi-scale edge-sensing upsampling module (MEUM) was designed, which incorporates a multi-level edge enhancement strategy to improve the reconstruction of muscle-disc anatomical boundaries. Experiments on lumbar MRI datasets show that the Dice similarity coefficient (DSC), Jaccard similarity coefficient (JAC) and Hausdorff distance (HD) scores of the model are 78.7%, 64.9% and 3.711 mm, respectively, which are significantly better than the traditional U-Net (DSC: 70.3%, JAC: 54.2%, HD: 4.437 mm), DSC and JAC increased by 8.4 and 10.7 percentage points, respectively, and HD decreased by 16.4%. The visualization results showed that the SCPA module significantly improved the segmentation accuracy around the intervertebral disc, while the MEUM module effectively suppressed artifacts and preserved the fine anatomy of the muscle. This study provides a high-precision automatic segmentation tool for lumbar MRI analysis to help clinicians make quantitative assessment and diagnostic decisions for LDH.