Development of 3D spatial measurement method based on computed tomography images of adolescent idiopathic scoliosis

Abstract

This thesis presents a deep learning framework specifically designed to automatically measure the 3D spatial angles from computed tomography (CT) images for pre-surgical patients with adolescent idiopathic scoliosis (AIS). AIS is a common spinal disorder among adolescents, characterized as a complex three-dimensional deformity that includes lateral curvature and vertebral rotation. Severe cases require surgical treatment. Traditional scoliosis assessment is based on 2D radiographic Cobb angle measurements. However, due to the three-dimensional complexity of the spinal deformity, 2D assessment may not capture the true spinal deformity, leading to inaccurate surgical planning and prognosis. In addition, segmentation of vertebrae can be a challenging task due to the morphological variations of the deformed vertebrae and the proximity of adjacent anatomical structures, which complicates identification and characterization. The 3D spatial angle gives a more comprehensive view of spinal alignment by considering the curvature in three dimensions rather than just two. This method is useful for surgeons to make more accurate surgical planning and expected outcomes. This study firstly utilized a dataset of 116 scoliosis patients to perform spine segmentation using U-net and a new developed neural network nnformer++. In addition, a new spine curve fitting network called NURBS-Net was developed using non-uniform rational B-spline curves (NURBS). The 3D spatial angle was then calculated by recognizing the maximum angular deviation between vertebrae along the curve. U-net and nnformer++ both have better performance in severe scoliosis spine segmentation compared to recent studies. The application of NURBS curves in spine curve fitting significantly outperforms traditional methods by providing finer control with fewer parameters, thereby minimizing the risk of overfitting and improving the reliability of the measurements. The 3D spatial angle predicted by the deep learning model correlated strongly with the traditional 2D Cobb annotated by the surgeon, with a Pearson correlation coefficient as high as 0.983. In conclusion, this method not only validates the feasibility of accurate, automated 3D spatial angle measurements preoperatively in scoliosis patients, but also emphasizes its potential for medical imaging and surgical planning. By providing a detailed 3D view of the spinal deformity, this method is expected to significantly improve surgical accuracy and outcomes for patients with AIS.