An Automatic Classification Method for Adolescent Idiopathic Scoliosis Based on U-net and Support Vector Machine

The traditional manual method for adolescent idiopathic scoliosis diagnosis suffers from observer variability. Doctors need an objective, accurate and fast detection method which would help to overcome the problem encountered by the traditional classification. This study introduces new techniques, including automatic radiograph segmentation, scoliosis measurement and classification, based on artificial intelligence. Firstly, the vertebral region in the radiograph was segmented by U-net and the scoliosis measurement was performed on the segmented image. Secondly, SVM classification was conducted by extracting the curve features in posteroanterior images and supplementary parameters in lateral and bending images. Finally, the results of automatic scoliosis measurement were compared with the one made by surgeons and the accuracy of the proposed automatic classification method was verified by a test set. The U-net segmentation model was successfully established to segment the vertebrae and the differences between the measurement results obtained by the automatic and manual measurement method were less than one degree and the accuracy of the automatic curve identification approach was found to be 100%.