Performance Evaluation of Deep Learning Models in Lumbar Spine MRI Segmentation

Spinal disorders have become a prevalent health problem around the world today, with far-reaching impacts on patients’ quality of life and ability to work. Spine segmentation is an effective tool for accurately identifying spinal disorders and is important in medical image processing and computer-aided diagnosis. Efficient and automated spine segmentation is essential to help doctors better identify spinal lesions and perform preoperative simulation and preparation. In recent years, the development of deep learning techniques has greatly contributed to the advancement of this field. However, there is a lack of studies comparing and evaluating the latest medical image segmentation models. In this paper, we evaluate the performance of state-of-the-art models proposed in recent years for spine MRI image segmentation, including U-Net, U-Net++, Attention-U-Net, MiniSeg, SENet, TransUNet, SwinUNet, Deeplabv3-ResNet 50, Deeplabv3-ResNet 101, Deeplabv3-MobileNet, Deeplabv3-Xception, Deeplabv3+ResNet 50, Deeplabv3+−ResNet 101, Deeplabv3+−MobileNet and Deeplabv3+−Xception. In this paper, each model is trained on the spine MRI dataset and a PCA-Driven Dynamically Weighted Hybrid Subjective-Objective Evaluation Algorithm for Comprehensive Model Performance (PCA-DHSEA) is proposed. The table selection of each model on each index is evaluated with this algorithm, and the results show that the Deeplab class model outperforms the U-Net class with other models, with the Deeplabv3+ model embedded in ResNet101 performing best. This study systematically summarizes deep learning models based on spine image segmentation, and the insights gained from it can provide a reference for the selection and development of models for medical image processing applications, thus promoting the development of computer-aided diagnosis in the management of spinal diseases.