Using Convolutional Neural Networks and Surface Topography for Detection of Adolescent Idiopathic Scoliosis

Introduction: An abnormal lateral curvature of the spine that can develop during the onset of puberty up until skeletal maturity is known as adolescent idiopathic scoliosis (AIS). AIS may cause back discomfort, breathing difficulties, and poor self-image. Additionally, worsening of the spinal curve is more prevalent in young females with AIS compared to males. The Cobb angle measured from radiographs is the primary outcome for diagnosing scoliosis. For mild curves, observation is usually sufficient, bracing is used for intermediate curves, and corrective surgery is recommended for severe curves. Children who are actively growing require frequent monitoring of the spinal curve. However, ionizing radiation exposure from x-ray assessments can increase a person’s risk of developing cancer. In addition, posterior-anterior radiographs have a substantial level of inter- and intra-observer variability, can only capture the 2D curvature of the spine, and are not able to evaluate aesthetics. An alternative tool that does not have adverse side effects is the 3D markerless surface topography technique, which quantifies the severity of trunk asymmetry. The ST technique can potentially be used as a scoliosis screening tool. However, detecting AIS from typically developing individuals is essential before clinical implementation. Therefore, this study aims to distinguish AIS patients from healthy adolescents using asymmetries present in the torsos. Methods: Surface torso scans were available from patients with AIS (n=241) and healthy subjects (n=85). All participants were 10- to 18-year-olds. Participants with AIS were included for all curve types with curves between 10° – 45°. The analysis of the scans involved reflecting the torso’s 3D geometry around the best plane of symmetry to highlight the external asymmetry in a deviation colour map image. Using the images as inputs, a convolution neural network (CNN) was developed to classify asymmetry patterns observed in healthy adolescents and those with AIS. The data was split into 60% training, 10% validation, and 30% test sets. The architecture of the CNN model consists of two convolutional layers with a rectified linear activation function. After each convolutional layer, a pooling layer is applied to decrease dimensionality and reduce the size of input parameters. Next, two fully connected linear layers were applied, followed by a 20% dropout layer to avoid overfitting. Finally, a sigmoid layer was applied. 400 epochs and a learning rate of 0.0001 were applied for training. Outputs are probability distribution with outcome bins 0 (i.e., Healthy) and 1 (i.e., AIS). Results: During the model training phase, the training and validation set obtained an accuracy of 89.8% and 87.1%, respectively. The accuracy is characterized as the number of subjects correctly classified to their target output (Healthy or AIS) by the CNN model. Additionally, the testing set’s accuracy, sensitivity, and specificity were all 88.9%. The positive predictive value of the testing set was 97.9%. Likewise, a negative predictive value of 57.1% was obtained. Conclusion: To accurately distinguish AIS from healthy people, a CNN prediction model was created. Further work involves Increasing the dataset size to Improving generalization.