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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Spinal Health

Background:

• Accurate radiological parameters of spinal alignment are vital for diagnosing and managing spinal deformities.
• Key parameters include sacral slope, pelvic tilt, pelvic incidence, and lumbar lordosis for lumbosacral assessment.
• Artificial Intelligence (AI) shows promise in automating these assessments, but requires large, diverse datasets, which are limited by privacy and data-sharing concerns.

Purpose of the Study:

• To demonstrate federated learning (FL) as a method for training deep learning models across multiple hospitals without compromising patient privacy.
• To compare the performance of FL models against centralized and local training approaches for spinal imaging analysis.

Main Methods:

• Federated learning (FL) was employed to train deep learning models across data from four participating hospitals.
• Model performance was evaluated against a centralized approach (pooled data) and local approaches (hospital-specific models).

Main Results:

• Federated learning (FL) achieved performance comparable to centralized training, with errors around 5 degrees.
• FL models consistently outperformed models trained on individual hospital data, showing lower errors in both internal (~8°) and external (~10°) testing.

Conclusions:

• Federated learning (FL) is a viable solution for collaborative AI development in spinal imaging, overcoming privacy barriers and data-sharing limitations.
• FL facilitates the use of diverse, multi-institutional data, enhancing model reliability across different clinical settings.
• FL particularly benefits smaller hospitals by enabling them to leverage larger datasets for improved AI model performance.