VLDNet:Spinal Vertebrae Landmark Detection with Adjacent Point Iterative Correction

Scoliosis is a prevalent structural deformity of the spine, and accurately locating vertebral keypoints from X-ray images is a fundamental step in clinical spinal diagnosis and treatment. In recent years, deep learning-based methods for automatic detection of vertebral keypoints have made significant progress. However, effectively utilizing spinal structure information from X -ray images and addressing errors in landmark detection and whether keypoints pairing is performed correctly remain challenging. This paper proposes a novel deep learning architecture called Spinal Vertebrae Landmark Detection with Adjacent Point Iterative Correction (VLD-Net), along with a patching mechanism based on adjacent point offsets and employs a method that integrates offset maps and heatmaps to enhance landmark localization accuracy. Additionally, the spine symmetry offset distance loss designed in this paper aims to further improve the accuracy and robustness of keypoint detection. Experimental results demonstrate that this method exhibits superior performance on the AASCE challenge dataset compared to other methods, providing strong support for medical image analysis.