SpineHRformer: A Transformer-Based Deep Learning Model for Automatic Spine Deformity Assessment with Prospective Validation

Abstract

<p>The Cobb angle (CA) serves as the principal method for assessing spinal deformity, but manual measurements of the CA are time-consuming and susceptible to inter- and intra-observer variability. While learning-based methods, such as SpineHRNet+, have demonstrated potential in automating CA measurement, their accuracy can be influenced by the severity of spinal deformity, image quality, relative position of rib and vertebrae, etc. Our aim is to create a reliable learning-based approach that provides consistent and highly accurate measurements of the CA from posteroanterior (PA) X-rays, surpassing the state-of-the-art method. To accomplish this, we introduce SpineHRformer, which identifies anatomical landmarks, including the vertices of endplates from the 7th cervical vertebra (C7) to the 5th lumbar vertebra (L5) and the end vertebrae with different output heads, enabling the calculation of CAs. Within our SpineHRformer, a backbone HRNet first extracts multi-scale features from the input X-ray, while transformer blocks extract local and global features from the HRNet outputs. Subsequently, an output head to generate heatmaps of the endplate landmarks or end vertebra landmarks facilitates the computation of CAs. We used a dataset of 1934 PA X-rays with diverse degrees of spinal deformity and image quality, following an 8:2 ratio to train and test the model. The experimental results indicate that SpineHRformer outperforms SpineHRNet+ in landmark detection (Mean Euclidean Distance: 2.47 pixels vs. 2.74 pixels), CA prediction (Pearson correlation coefficient: 0.86 vs. 0.83), and severity grading (sensitivity: normal-mild; 0.93 vs. 0.74, moderate; 0.74 vs. 0.77, severe; 0.74 vs. 0.7). Our approach demonstrates greater robustness and accuracy compared to SpineHRNet+, offering substantial potential for improving the efficiency and reliability of CA measurements in clinical settings.</p>