Accurate scoliosis vertebral landmark localization on X-ray images via shape-constrained multi-stage cascaded CNNs

Abstract

Vertebral landmark localization is a crucial step in various spine-related clinical applications, which requires detecting the corner points of 17 vertebrae. However, the neighboring landmarks often disturb each other because of the homogeneous appearance of vertebrae, making vertebral landmark localization extremely difficult. In this paper, we propose a multi-stage cascaded convolutional neural network (CNN) to split a single task into two sequential steps: center point localization to roughly locate 17 center points of vertebrae, and corner point localization to determine four corner points for each vertebra without any disturbance. The landmarks in each step were located gradually from a set of initialized points by regressing offsets using cascaded CNNs. To resist the mutual attraction of the vertebrae, principal component analysis (PCA) was employed to preserve the shape constraint in offset regression. We evaluated our method on the AASCE dataset, comprising 609 tight spinal anteroposterior X-ray images, and each image contained 17 vertebrae composed of the thoracic and lumbar spine for spinal shape characterization. The experimental results demonstrated the superior performance of vertebral landmark localization over other state-of-the-art methods, with the relative error decreasing from 3.2e−3 to 7.2e−4.