Multi-scale generative adversarial network: three-dimensional reconstruction of the scoliotic spine from biplanar X-rays.

Background: Adolescent idiopathic scoliosis (AIS) causes lateral spinal curvature, affecting adolescent growth, nerve function, and cardiopulmonary health, with an incidence of 1% to 3% in adolescents. Accurate three-dimensional (3D) spinal assessment is critical for AIS management, yet clinical imaging faces dilemmas. Specifically, two-dimensional (2D) X-rays lack 3D information due to vertebral overlap, while computed tomography (CT) involves excessive ionizing radiation that increases tumor risks in adolescents. Existing biplanar X-ray-based 3D reconstruction methods suffer from feature fusion loss and poor adaptability to scoliotic deformities. This study aimed to develop a low-radiation and high-precision deep learning framework for 3D spinal reconstruction from orthogonal X-rays. Methods: We proposed a multi-scale generative adversarial network (MS-GAN) to reconstruct 3D spinal CT volumes from two orthogonal X-rays of both normal and scoliotic spines. The model incorporates three innovative designs: a residual-dense encoder for retaining fine vertebral details, an adaptive cross-view fusion module for integrating orthogonal projections, and a multi-scale fusion discriminator (MSFD) to ensure structural consistency. It was validated on a mixed dataset including 1,087 normal spines from CTSpine1K and clinical AIS data covering 138 cases with Cobb angles ranging from 15° to 85°. Reconstruction performance was evaluated using mean squared error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and Dice coefficient (Dice). Clinical applicability was further assessed using sagittal vertical axis (SVA), lumbar lordosis (LL), and thoracic kyphosis (TK) measurements on 38 real biplanar X-ray cases. We compared the performance gap between the model on real clinical data and simulated data, and verified the effectiveness of the model improvement through ablation experiments. Results: Compared to the baseline X2CT-GAN, MS-GAN achieved a 30% reduction in MSE; its PSNR reached 73.20 dB, an improvement of 12.2%; its SSIM reached 0.956, an improvement of 5.5%; and its Dice reached 0.92, an improvement of 8%. Analysis demonstrated strong agreement between reconstructed spines and CT gold standards, with mean differences of -0.198 mm for SVA, 0.087° for LL, and 0.095° for TK. When applied to real clinical X-rays, performance decreases remained within 5% for key metrics. And the results of the ablation experiment demonstrated the effectiveness of the model improvement. Conclusions: MS-GAN resolves feature fusion and dimension mismatch in biplanar X-ray 3D reconstruction, with robust performance on scoliotic spines. It avoids CT radiation and reduces equipment reliance, providing a low-cost, high-precision tool for AIS assessment and is valuable for primary hospitals.