DVVDNet: A dual-view vertebra detection network with anatomical constraints for automatic scoliosis assessment

• Published in: Neurocomputing (Amsterdam) Vol. 650; p. 130913
• Main Authors: Hu, Rui, Shu, Xinwu, Hu, Guoxiong, Zhang, Xiao
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 14.10.2025
• Subjects: Anatomical constraint; Deep learning; Scoliosis; Vertebra detection; Deep learning; Vertebra detection; Scoliosis; Anatomical constraint
• ISSN: 0925-2312
• DOI: 10.1016/j.neucom.2025.130913

Scoliosis is a common spinal deformity, with adolescent idiopathic scoliosis being the most prevalent form. It significantly affects adolescents’ health and quality of life. Accurate assessment of spinal anatomy is essential for diagnosis and personalized treatment planning. Recent advances in deep learning have improved automated scoliosis evaluation, offering promising tools for clinical decision support. However, current methods face challenges such as vertebral feature loss in single-view X-rays due to occlusion, and reduced accuracy caused by image noise and low contrast. To address these issues, a deep learning-based dual-view vertebra detection network is proposed, which effectively integrates anterior-posterior and lateral spine radiographs by leveraging their complementary anatomical information. The network incorporates a gated multi-scale attentive fusion module to selectively enhance informative features while suppressing redundancy, thereby improving the representation capability of dual-view fusion. Additionally, a vertebral anatomy constraint loss is introduced, which incorporates both geometric morphology and structural characteristics of vertebrae to regularize model predictions, enhancing output stability and clinical applicability. The method is evaluated on the public AASCE challenge dataset and a private hospital dataset. Experimental results show high agreement with radiologists’ assessments, demonstrating the model’s accuracy and robustness. This study presents an automated, high-precision tool for scoliosis assessment, supporting more efficient diagnosis and individualized treatment planning.