MSRUNet: Multiscale Residual U-Net Network for Spine MRI Image Segmentation

• Published in: 2024 IEEE 25th China Conference on System Simulation Technology and its Application (CCSSTA) pp. 285 - 290
• Main Authors: He, Junjie, Li, Kun, Qin, Juan, Ji, Yuehui, Xia, Dan, Fu, Xulong
• Format: Conference Proceeding
• Language: English
• Published: IEEE 21.07.2024
• Subjects: ASPP; Computational modeling; Convolutional neural network; Deep learning; Image segmentation; Magnetic resonance imaging; Measurement; Medical diagnostic imaging; Multiscale residual; Performance analysis; Resilience; Spine; Spine MRI segmentation; Systems simulation
• DOI: 10.1109/CCSSTA62096.2024.10691809

Spinal medical images are essential for clinicians to diagnose spinal conditions. Computer segmentation can often be utilized to assist physicians in making a diagnosis, but traditional segmentation methods are time-consuming. Recently, deep learning has become more prevalent in medical imaging. This paper introduces a high-performance automatic segmentation method for spinal MRI images using deep learning. The proposed method, called multiscale residual UNet (MSRUNet), enhances the classical U-Net architecture by incorporating residual double convolution and a customized ASPP block. These improvements allow the model to accurately focus on and segment various tissue and lesion regions in MRI images. This study assesses the model's performance on SpineSagT2Wdataset3 by measuring its recall, accuracy, Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics. Compared to other models, including U-Net, MSRUNet showed superior segmentation performance, demonstrating its effectiveness. This work advances automated spine MRI segmentation and promotes automated diagnosis and treatment of spinal disorders.