Vertebral Labelling in Radiographs: Learning a Coordinate Corrector to Enforce Spinal Shape

Abstract

Localizing and labeling vertebrae in spinal radiographs has important applications in spinal shape analysis in scoliosis and degenerative disorders. However, due to tissue overlaying and size of spinal radiographs, vertebrae localization and labeling are challenging and complicated. To address this, we propose a robust approach for landmark detection in large and noisy images and apply it on spinal radiographs. In this approach, the model has a holistic view of the input image irrespective to its size. Our model predicts the labels and locations of vertebrae in two steps: Firstly, a fully convolutional network (FCN) is used to estimate the vertebrae location and label, by predicting 2D Gaussians. Then, we introduce the Residual Corrector (RC) component, that extracts the coordinates of each vertebral centroid from the 2D Gaussians, and correct the location and label estimations by taking into account the entire image. The functionality of the RC component is differentiable. Thus, it can be merged to the deep neural network, and trained end-to-end with other sub-networks. We achieve identification rates of 85.32% and 52.28% for sagittal and coronal views and localization distance of 4.57 mm and 5.33 mm in sagittal and coronal views radiographs, respectively.