Densely connected heatmap regression network for X-ray image landmark localization

The localization of landmarks in X-ray images of spinal columns is of significant value in the fields of disease diagnosis and surgical robot path planning. However, existing methods for landmark localization in X-ray images suffer from insufficient accuracy and poor robustness against severe occlusions and large-area uneven lighting. We propose a deep learning model, namely the Densely Connected Heatmap Regression Network (DHR-Net) for landmark localization in X-ray images. DHR-Net can effectively fuse feature information of different depths and scales, thus achieving higher accuracy and robustness in the task of spinal anatomical landmark localization on X-ray images compared with existing methods. The average localization errors of DHR-Net in the landmark localization tasks for anteroposterior and lateral X-ray images are 1.41 mm (mm) and 1.57 mm respectively. When using 2 mm as the judgment threshold, the success detection rates are 68.69% and 64.14% respectively; when using 3 mm as the judgment threshold, the success detection rates are 88.89% and 79.80% respectively. Therefore, DHR-Net exhibits promising clinical application potential in the spinal X-ray localization task.