Computer-aided diagnosis of spinal deformities based on keypoints detection in human back depth images

Abstract

Physicians often need to manually measure physiological indicators through examinee’s X-ray for spinal deformities. However, this process is time-consuming and harmful to the human body because of radiation. Currently, computer-aided diagnosis technology has been gradually applied in the screening of spinal deformities. Due to the increased pressure on the spine and lower back during pregnancy, individuals may experience varying degrees of spinal deformities. We propose a novel method to help doctors non-invasively, accurately, and quickly diagnose spine deformity for postpartum women. First, this proposed method captures depth images of the human back by a Kinect DK camera. Second, it employs a region-growing method to extract the human back region and converts the back depth image into a point cloud image through mapping. After that, due to the issues of noise and voids in the depth image, joint bilateral filtering algorithm is used for repair and smoothing. Third, the OpenPose network is utilized to extract spinal keypoints (i.e., C7, T12, L5 and S5). Then, we use Delaunay triangulation and linear interpolation to transform the point cloud image into a three-dimensional rectangular mesh. Subsequently, based on the concavity and convexity of the back surface, interpolation fitting of spine point set is performed using horizontal adjustment and B-spline fitting methods to reconstruct the spinal curve. Last, the experimental results indicate that the prediction accuracies for the six physiological indicators are 0.83, 0.43, 0.81, 0.82, 0.80, and 0.76, respectively. Additionally, the two PCK indicators are 0.87 and 0.90 for spinal point detection.