Performance Comparison between U-Net Variant Models in Spine Segmentation

Abstract

Spine Magnetic resonance imaging (MRI) is a crucial diagnostic technique for illnesses of the spinal cord. The UNET network, the most prominent neural network model for segmenting medical images has opened up new opportunities for spin MRI segmentation as a result of the rapid development of deep-learning algorithms. In this study, we compared the difference between UNet and five other variants (Unet++, Unet+++, Attention-UNet, Dense-UNet, and R2UNet) in performance and efficiency by training and testing them on the same Spine MRI image dataset that contained 200 patients. The results showed that Attention-UNet performed best on the Miou (83.33 percent) and Average dice(89.15 percent) metrics; R2UNet performed best on the Accuracy (97.12 percent) metric. Attention-UNet has the slightest difference between the basic segmentation and the baseline value in terms of segmentation performance. This study could provide a better understanding of different networks on the Spine MRI Segmentation task.