Analyzing Racial Differences in Imaging Joint Replacement Registries Using Generative Artificial Intelligence:...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiographic Disparities

Background:

• Medical datasets can contain biases, impacting deep learning models and leading to inequitable clinical applications.
• Understanding racial biases in medical data is essential for creating fair healthcare technologies.
• This study investigates radiographic differences between races in total hip arthroplasty patients using generative deep learning.

Purpose of the Study:

• To explore and understand racial variations in pelvic radiographs.
• To identify and visualize systematic differences in radiographic features between racial groups.
• To leverage generative AI for bias detection in medical imaging.

Main Methods:

• Retrospective analysis of pelvic radiographs from total hip arthroplasty patients.
• Utilized denoising diffusion probabilistic models to generate synthetic radiographs conditioned on demographics and imaging features.
• Generated transition videos comparing White and African American pelvises, analyzed by expert surgeons and radiologists.

Main Results:

• Analyzed 480,407 pelvic radiographs, noting a predominance of White patients.
• Generative model produced high-quality images with a Fréchet Inception Distance of 6.8.
• Identified six differentiating radiographic characteristics: interacetabular distance, osteoarthritis degree, obturator foramina shape, femoral neck-shaft angle, pelvic ring shape, and femoral cortical thickness.

Conclusions:

• Generative models show promise in uncovering disparities within medical imaging datasets.
• Visualizing race-based radiographic differences helps identify bias in AI models.
• This approach supports the development of more equitable healthcare AI.