AUTOMATED SCOLIOSIS ASSESSMENT FROM SPINE X-RAY USING GENERATIVE ADVERSARIAL NETWORKS (GANs)

Abstract

Scoliosis, a lateral or S-shaped curvature of the spine, needs to be diagnosed accurately and as early as possible for effective treatment. Conventional assessment relies on the manual measurement of cobb angle which is a labour-intensive process potentially subject to human error. We propose a novel two-phase Generative Adversarial Networks (GAN) based method for automatic scoliosis evaluation in this work. We propose the first stage of a GAN based model to specifically segment the spine, which is then followed by a second stage in which the main clinical features such as the cobb angles are extracted and analysed to evaluate how severe the curvature is. The model, which is trained on labelled spine X-ray datasets, shows high accuracy in segmentation and scoliosis. We present a model trained on labelled datasets of spine X-rays, that has high segmentation accuracy and scoliosis classification and reduces manual analysis. The experimental findings reaffirm the strength of this framework, and highlight its possibilities for integration into clinical routine in order to improve diagnostic accuracy and minimize the existence of unnecessary imaging