End-to-End Fully Automatic Segmentation of Vertebrae in Spinal X-Ray Images

Accurate vertebral identification and labeling is essential in image-guided spinal disease diagnosis and treatment planning. Unfortunately, spinal assessments traditionally rely on tedious and time-consuming manual measurement subject to observer variability. In particular, the measurement of scoliosis requires the segmentation of individual vertebrae, but an automatic method that can accurately identify and segment vertebrae is unavailable in the literature. We introduce an end-to-end fully automatic segmentation method that leverages a carefully-adjusted U-Net model with progressive side outputs in order to provide reliable segmentations of the vertebrae associated with scoliosis measurement. Our experimental results with X-ray images of scoliosis patients indicate that our method, which achieves an average Dice score of 0.993, promises to be an effective tool in the identification and labeling of vertebrae for the reliable estimation of scoliosis.