THA-Flow Generative Model: Prosthesis Geometry Prediction from Preoperative CT

Total hip arthroplasty (THA) planning is traditionally treated as a deterministic task, where a single prosthesis configuration is selected for a patient's anatomy. However, clinical practice often allows for multiple valid solutions. This study introduces THA-Flow, a generative AI model designed to treat THA planning as a one-to-many problem, generating a conditional probability distribution of 3D prosthesis geometries that fit a patient's specific bone structure.

Researchers developed a two-stage generative pipeline using a retrospective cohort of 1,355 hips. First, separate AutoencoderKL models were trained to compress preoperative bone anatomy and prosthesis geometry into latent spaces. Second, a 3D UNet was trained using rectified flow matching to generate prosthesis geometry from Gaussian noise, conditioned on the preoperative bone latent and optional structured prosthesis parameters (such as stem model and size). The model represents prostheses as dual-channel truncated signed distance fields (TSDFs), allowing for independent registration of the acetabular and femoral components to the patient's preoperative coordinate space.

THA-Flow successfully generated complete acetabular and femoral geometries across seven major stem models, which accounted for over 93% of the study cohort. The model demonstrated high fidelity in reconstruction and spatial consistency, with generated components adapting to individual anatomical features like the medullary canal and the femoral calcar. Repeated sampling showed that the model maintains stable component position and alignment while allowing for local geometric variations, effectively capturing a distribution of plausible surgical plans rather than a single fixed template. The system supports interactive candidate generation, with inference times under two seconds per case.