Introduction: Three-dimensional spine reconstruction methods currently used to study scoliosis require manual identification of several anatomical features, which is time-consuming, costly and error-prone.
Objectives: To provide fast 3D reconstructions of the spine that may be accomplished by non-expert users. More specifically, recovering the 3D position of 6 anatomical landmarks per vertebra based on a pair of radiographs with minimal user interaction.
Materials and Methods: Splines approximating the spine midline were identified by end-users using a small number of control points (typically 4 to 7 points on each radiograph). Then, vertebrae location, rotation and 3D anatomical landmarks were recovered by deforming a statistical articulated model that captures intervertebrae variability (compiled using 291 prior reconstructions). The deformation optimizes both the fit with the user's splines and the prior probability given by the statistical model.
A set of 14 in vivo exams of scoliotic patients were used for validating the method. 3D reconstructions obtained using a previously validated method were compared with reconstructions using the proposed method performed by 2 volunteers with limited knowledge on spine radiology. Volunteers only had 20 minutes of training with the software tool (figure 1).
Results: The mean reconstruction errors were 3.4mm for the endplates and 4.8mm for the pedicles. Furthermore, the average reconstruction time was 1min28s.
Conclusion and Significance: Results show that rough reconstructions of the spine may be rapidly achieved by non-expert users with very little training. This makes the method attractive when fast reconstructions are required, or when dedicated personnel would be too expensive.