Deep Learning Based 3D Reconstruction of the Spine from Low Dose Biplanar Radiographs

Personalized 3D reconstruction of the spine from biplanar radiographs is of great interest for characterizing scoliosis and has potential for early diagnosis of disease progression. However, due to the tedious manual intervention, this technique is not widely used in clinical routine. In this work, we proposed to partially automate this method thanks to the use of deep learning to leverage the burden of manual landmarking of the radiographs. Manual intervention remained necessary but limited to quality control and manual rigid adjustment of the 3D models. The proposed method was trained and validated on 135 subjects and tested on 34 subjects, including 10 scoliotic subjects. The reconstructed 3D models were compared with reference reconstructions obtained with a validated method. We found no major biases in the position and orientation of the vertebral bodies. The uncertainty in the position of the vertebral body centers was limited to 2 mm. Further efforts are needed to obtain reliable estimates of axial rotation in the severe scoliotic population.