Vertebrae Segmentation with Generative Adversarial Networks for Automatic Cobb Angle Measurement.

Scoliosis is a lateral deformity of the spine, usually diagnosed in patients with a Cobb angle (CA) greater than 10°. Accurate measurements of the CA are necessary for timely intervention and subsequent effective treatment of scoliosis. However, the existing gold standard to calculate the CA requires individual vertebrae pairs to be manually identified, which is a laborious process. The segmentation of vertebrae on X-ray images can be ridden with various challenges including image artifacts and varying intensity and contrasts. Generative Adversarial Networks have been applied to a variety of segmentation tasks with high accuracy. In our study, we aim to investigate if GANs can produce more distinct and defined vertebrae compared to traditional segmentation models. Our results demonstrate that compared to traditional segmentation models, the Pix2Pix ensemble models were better than the other traditional models in distinguishing boundaries between different vertebrae. This would contribute to the development of a more robust CA estimation pipeline downstream that is capable of generalizing across datasets with greater variability in noise.