Prediction Model of Scoliosis Progression Bases on Deep Learning

Abstract

By deep learning technique, we present a new approach to model idiopathic single curve scoliosis. We leverage the advanced version of the recurrent neural network, that is, the long short-term memory network, to achieve the goal. We frame scoliosis as a classification problem and a regression problem. A network for classification is designed first. We perform the training and testing with real clinic records that are imputed by various tricks. Using this model, one can classify the current level of scoliosis into three predefined groups via a few publicly measurable indictors, such as body height or arm span. We also design a regression network that can predict the future progression of spine curvature. This model can infer the development in spine curvature at a certain time span according to the changes of other indictors. Both of these models are evaluated by various metrics. The experiment shows that the quantitative picture of the scoliosis can be captured by our models giving a significant performance boost. Hence, the resulting decision-support system can help to decide the necessity of a further intervene both for physicians and patients.