An Efficient Spine Segmentation Method

Abstract

The spine is the most complex load-bearing structure in the human body, and herniated discs, spinal stenosis, and degenerative discs are common spinal disorders. MRI is an effective imaging method in medicine, but the identification and quantitative analysis of lesions require physician judgment, which is not only a huge workload but also carries the subjective judgment of physicians, and such drawbacks can be solved by using image segmentation technology. In this paper, we propose an efficient spine segmentation method consisting of selective preprocessing and post-processing and an improved UNET network structure. In the selective pre-post processing, meaningful parts of the MRI are selected for random input, and the selected parts are effectively restored back to the original size of the segmented image. In the improved UNET network, differing from the traditional UNET structure, the perceptual field of the image input is increased by using inflated convolution, and the attention mechanism is added in the up-sampling and down-sampling end parts for better filtering of features. The experimental results show that our method outperforms the traditional method by substantially reducing the training elapsed time and performing well in terms of the accuracy of the model.