An Automatic Scoliosis Diagnosis and Measurement System Based on Deep Learning

Adolescent idiopathic scoliosis (AIS) is a three-dimensional structural deformity of the spine which affects 1-4% of adolescents and causes not only deformed appearance but also compromised mental status, pulmonary function, motor function and life quality. Currently, the diagnosis of AIS depends on the measurement of Cobb angle on spine radiographs, which is performed manually by doctors. Intra-observer and inter-observer variation exist in such method and causes errors in diagnosis. The purpose of this study is to design an automatic scoliosis diagnosis and measurement system based on deep learning to improve measurement accuracy and assist doctors in diagnosis. U-net segmentation network was used to segment the spine radiographs. Based on the segmented images, specific positions of upper and lower end vertebrae (UEV and LEV) and slopes of their endplates were identified by minimum outer envelope rectangle and least square method. Subsequently, Cobb angles were measured as angles between superior endplates of UEVs and inferior endplates of LEVs. After comparing manually measured Cobb angles with computer- measured Cobb angles, the result showed that the computer- measured Cobb angles were similar to the manually measured ones. To sum up, this system for automatic measurement of Cobb angle can assist doctors in clinical diagnosis.