Deep learning based vertebra localization and Cobb angle estimation using x-ray images

Scoliosis (a spinal ailment) with improper lateral curvature and rotational abnormalities may have a significant impact on the general health and physical development of a person. For thediagnosis of this problem, Cobb angle measurement is required from Anterior-Posterior (AP) X-ray images. At present, a manual estimate is performed, which is time-consuming and prone to error.Usingthis dual techniques, the model is able to collect both structural and edge details information, and the IEM is able to combine these characteristics to improve vertebra localisation and tilt estimates.

The proposed approach defines a new Deep Learning (DL) architecture (i.e. Scolio-Net) that consists of a Dual-Input Convolutional Neural Network (DICNN) and an Information ExchangeModule (IEM) to analyzes the original X-ray images and its edge-detected counterparts.

The model is tested on a manually annotated dataset (where one node is creating heat maps to detect vertebral centres and another for calculating tilt) under the expert guidance of orthopaedic specialists of the Postgraduate Institute of Medical Education and Research (PGIMER), Chandigarh, India.

The dual-pathway approach addresses tilt calculation and vertebral center detection to ensure the precise analysis of spinal curvature. The proposed Scolio-Net model has demonstrated outstanding performance compared to existing state-of-the-art methods, achieving the lowest SMAPE value of 7.68.