Simulating Federated Learning to Enable Multi-Hospital Collaboration for Lumbopelvic Alignment Estimation

Background: Accurate computation of radiological parameters related to spinal alignment is clinically crucial for diagnosing and managing conditions, such as adolescent idiopathic scoliosis and adult spinal deformities. Key parameters, including sacral slope, pelvic tilt, pelvic incidence, and lumbar lordosis, are required to assess lumbosacral alignment. Artificial Intelligence (AI) has demonstrated strong potential in automating these assessments, reducing clinician workload and improving consistency. However, AI models require large, diverse, high-quality datasets to perform reliably across different clinical settings. Privacy concerns and data ownership issues often hinder data sharing, limiting the creation of centralized datasets. Methods: In this study, we demonstrate that federated learning (FL) enables the training of deep learning models across four hospitals without compromising patient privacy. In particular, we compared FL against a centralized approach, where data from all the hospitals are pooled together and a model is trained on them, and a local approach consisting of training individual models exclusively on data from each respective hospital, resulting in distinct hospital-specific models. Results: FL achieved performance comparable to centralized training (errors ~5°), where data is pooled, and consistently outperformed models trained on data from individual hospitals, both in internal (~8°) and external (~10°) testing. Conclusion: This work highlights FL as a viable solution for collaborative AI development in spinal imaging, facilitating the use of diverse, multi-institutional data while circumventing privacy barriers and complex data-sharing agreements. Additionally, FL demonstrates particular benefits for smaller hospitals, enabling them to achieve superior model performance by effectively leveraging data from hospitals with larger datasets.