Fourier-inspired Automatic Cobb Angles Measurement for Adolescent Idiopathic Scoliosis

Measuring Cobb angles in adolescent idiopathic scoliosis (AIS) patients using anteroposterior (AP) radiographs is time-consuming and prone to variability due to subjective vertebrae selection. To address this, we propose an automated model for Cobb angle measurement using a Fourier-inspired method. The model starts by estimating the spinal curvature spectrum and then predicts the Cobb angles. We use a VGG-16 transfer learning model to regress the concatenated real and imaginary components of the spine's spectral representation. This feature is used to reconstruct the curvature through an inverse Fourier transform for improved model explainability. Simultaneously, it serves as input to a projection head for Cobb angle estimation. Evaluating on the MICCAI 2019 SpineWeb dataset, our model shows a Pearson correlation of 0.9382 for curvature estimation and a symmetric mean absolute percentage error (SMAPE) of 12.25% for Cobb angle measurement, performing similarly to other top models. By treating spinal curvature prediction as a frequency domain analysis, our method offers more accurate curvature shapes than spatial domain approaches. Additionally, applying zero padding improves spectral resolution and reduces SMAPE in Cobb angle estimation. Spectral prediction consistently outperforms spatial prediction in both curvature and Cobb angle measurement.