Accurate automated 3D lumbar spine reconstruction from biplanar X-rays using multi-task deep learning and anatomy-aware optimization.
Authors
Affiliations (7)
Affiliations (7)
- School of Biomedical Engineering and Med-X Research Institute, Shanghai Jiao Tong University, Shanghai, China.
- Engineering Research Center of Digital Medicine and Clinical Translation, Ministry of Education, Shanghai, China.
- TAOiMAGE Medical Technologies Corporation, Shanghai, China.
- School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai, China.
- Department of Orthopaedic Surgery, The 3rd Hospital of Hebei Medical University, Shijiazhuang, Hebei, China.
- Guangdong Provincial Hospital of Chinese Medicine, The Second Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, China.
- Department of Spine Surgery, Tongji Hospital, School of Medicine, Tongji University, Shanghai, China.
Abstract
Accurate 3D assessment of the weight-bearing lumbar spine is crucial for diagnosing various spinal pathologies. However, existing biplanar X-ray reconstruction methods struggle with complex pathologies and low-contrast structures. This study aims to propose a fully automated framework for high-accuracy 3D lumbar spine reconstruction from biplanar X-ray images. Statistical shape models (SSMs) of L1-L5 vertebrae were constructed from 270 lumbar CT scans. A multi-task deep learning network was designed to simultaneously isolate vertebral signals from biplanar X-rays and detect anatomical landmarks. Landmark predictions were used for pose initialization, followed by SSM-based 2D-3D registration using an anatomy-aware weighted optimization strategy that emphasized the transverse and spinous processes. The method was validated against patient-specific CT segmentations registered to the biplanar imaging geometry. The method was evaluated in subjects without significant osseous abnormalities and in pathological patients from an independent center. The proposed network achieved Dice coefficients of 0.991 and 0.989 for anteroposterior and lateral vertebral signal isolation, respectively. The overall 3D reconstruction accuracy was 0.85 ± 0.24 mm. Reconstruction accuracy was 0.80 ± 0.15 mm in Center 1 and 1.32 ± 0.46 mm in the pathological cohort from Center 2. The method maintained stable performance under additive Gaussian noise and across different lumbar postures. Ablation studies confirmed the contributions of signal isolation, landmark-based initialization, and anatomy-aware optimization. This robust, high-accuracy automated method precisely reconstructs complex lumbar structures from biplanar X-rays, even for severely pathological vertebrae. Its ability to maintain accuracy under extreme movement and improve reconstruction quality in complex regions highlights its significant potential for clinical diagnosis and surgical planning.