Spinal Curve Guide Network (SCG-Net) for Accurate Automated Spinal Curvature Estimation

Abstract

Cobb angle plays an important role in the diagnosis of scoliosis, which can effectively quantify the degree of scoliosis. Manual measurement of Cobb angles is time-consuming, and the results are also heavily affected by the expert’s choice. In this paper, we propose a spine curve guide framework to directly regress the cobb angle from single AP view X-rays images. We firstly design a segmentation network to accurately segment two spine boundary, and then aggregate the obtained boundary scoremap with the original spinal X-rays images to input another angle estimation network to make high-precision regression prediction for cobb angle. We evaluate our method in the AASCE19 challenge, and our result achieves 22.1775 SMAPE that shows strong competitiveness compared to other excellent methods.