Automating spinal alignment analysis with machine learning and PCdare software: focus on idiopathic scoliosis

Optical 3D scanning is an emerging technology for evaluating spinal deformities in patients with idiopathic scoliosis. Despite the potential of surface scanning to avoid radiation exposure, its reliability and accuracy remain moderate compared to radiography. PCdare is an open-source software that is used to analyze and compare the optical 3D surface scans of the patients’ back with the corresponding biplanar radiographic images of the patients’ spine to better understand the relationship between internal spinal alignment and 3D back shape. It includes two important components: the line through the centroids of the vertebral bodies, the internal spinal line (ISL), and the line connecting all spinous processes on the back, the spinous process line (SPL). However, this software requires users to manually draw the ISL, which can be time consuming and subjective. Here we present a fully automatic approach for ISL estimation to improve the workflow of PCdare. We implemented and trained a convolutional neural network-based U-Net on 330 anterior-posterior radiographs of the spine and subsequently detected the centroids of vertebral bodies to estimate the ISL. We evaluated the model’s performance on a test set of 63 images, obtaining a median dice similarity coefficient of 0.93. The median root mean square error between ISL estimations and ground truth was 1.16 mm. Our method demonstrated promising results and will improve the applicability of the PCdare software.