Scoliosis is a condition that affects 2-4% of adolescents. Orthopedic spine surgeons use the Lenke classification to categorize the various types of adolescent idiopathic scoliosis. This triad classification system utilizes upright coronal, upright sagittal and left/right bending radiographs and is composed of 3 different parts: a curve type (1-6), a lumbar spine modifier (A, B, C), and a sagittal thoracic modifier (-, N, +). This classification system has been adopted worldwide due to its reliability as a tool to assist surgeons in determining the best method of treatment based on curve patterns. While it allows for a clear 3-dimensional classification of scoliosis, it is complex and suffers from intra- and inter- observer reliability issues. Deep neural networks have been increasingly applied to medical imaging to solve complex image recognition problems at accuracies comparable to human physicians. Previous studies have applied various computer-assisted methodologies to help classify the various types of scoliosis. However, no study has used a deep neural network machine learning approach to directly classify scoliosis radiographs into the Lenke classification.
After acquiring the necessary radiographs for classification of juvenile indiopathic scoliosis, the radiographs for each patient will be labeled using the Lenke classification scheme. The full set of radiographs will be split into a training set and a test set. An 8-layer neural network comprised of 7 serial convolutional blocks and a single fully-connected layer will be trained utilizing the radiographs and labels. After training is completed, the neural network will be applied to the test set for cross-validation.
The aim of this study is to apply a deep neural network approach to classify adolescent idiopathic scoliosis radiographs into the Lenke system. Developing such a system will decrease the amount of time necessary for classification. Furthermore, it would increase the accuracy of classification and reduce inter-observer effects.