Introduction: Automatization of 3D reconstruction of the spine from frontal and sagittal radiographs is extremely challenging. For example, the overlying features of soft tissues and air cavities may interfere with image processing algorithms.
Method: To overcome these problems, the proposed method efficiently combines the partial information contained in two images from a patient with a statistical 3D spine model generated from a database of scoliotic patients. The algorithm operates through two simultaneous iterating processes. The first process generates a personalized vertebra model using 2D/3D registration with bone boundaries extracted from radiographs, while the other process infers the position and the shape of less visible vertebrae from the estimation of the well registered vertebrae using a statistical 3D model.
Results: The method is applied to 8 patients of the Erasme Hospital (Belgium) to obtain some results on shape accuracy based on 20 lumbar vertebrae (L1 to L4). The in-vivo experiments, which consist in comparing the 3D reconstructions (only the regions of the vertebral body and pedicles) obtained from the low-dose radiographic system EOS (biospacemed) to 3D surface models of the vertebral shapes reconstructed from CT-scan, show an average and a standard deviation error of less than 1.0 mm for the 20 vertebral shapes reconstructed by two users.
Conclusion: Experimental evaluations confirm that the proposed method gives viable 3D reconstructions and is an accurate and reliable alternative to competitive state-of-the-art methods. The proposed method requires only 3 minutes to complete, allowing an acceptable and fast enough 3D reconstruction for a routine clinical use.