An artificial-intelligence model generated complete three-dimensional hip-prosthesis geometry from preoperative CT scans and allowed the output to be guided by a chosen stem model and related size settings. When the researchers repeatedly sampled outputs from the same bone data, most generated regions overlapped the registered postoperative prosthesis. Differences tended to appear at the cup rim, the upper stem boundary and the stem tip, while the cup and stem stayed in the same position and stem alignment remained stable.
The work is an arXiv preprint based on routine-care CT examinations and operative records from a single centre, with identifiers removed before processing. No additional intervention was assigned, so the report is a technical assessment rather than a test of patient benefit.
A model built around the bone
Researchers screened 1,483 hips from primary total hip arthroplasty episodes and excluded 128. The final cohort contained 1,355 hips from 1,149 patients: 649 left hips and 706 right hips. The data were divided into 1,188 training hips, 84 validation hips and 83 held-out test hips.
Training took place in two stages. The first used latent-space pretraining, meaning the system first worked with a compact numerical representation of the geometry. The second used conditional rectified-flow matching. Separate AutoencoderKL models handled bone and prosthesis geometry.
The bone representation entered the network as a spatial condition from its first convolution. Six 256-dimensional tokens carried the stem model, that model's size, cup outer diameter, head diameter, head offset and liner offset. The setup allowed bone-only generation or generation conditioned on a recorded stem model.
Before training, the researchers used separate rigid registrations for the acetabular and femoral sides to accommodate relative hip motion. They represented the surfaces in a dual-channel truncated signed distance field, or TSDF, a grid that records distance from a surface, with values truncated at plus or minus 5 millimetres. The cup and femoral head occupied the acetabular channel, while the stem occupied the femoral channel.
The strongest results were geometric
On validation, the bone reconstruction branch reported an L1 score of 0.007946, a PSNR of 35.85 dB and an SSIM of 0.9802. For the prosthesis TSDF branch, the figures were 0.003313, 47.11 dB and 0.9964. These are reconstruction measures, rather than clinical outcome measures.
The seven most prevalent femoral stem models covered 1,265 hips, or 93.4% of the cohort. Cross-design comparison used one validation case from each model. That makes the comparison illustrative, and the supplied analysis does not provide a quantitative held-out-test performance estimate.
In consecutive-section review, the generated femoral geometry entered the effective canal proximally without crossing the calcar boundary, then narrowed into a continuous, regular distal taper. On the acetabular side, the cup was complete and its outer surface followed the acetabular wall. The interface findings were visual and sectional; no quantitative interface effect was reported.
Adding the recorded stem model changed selected features—stem length, proximal width and lateral profile—while both generation modes produced complete acetabular and femoral geometry and reported anatomical fit was preserved. In repeated bone-only samples, most generated regions still overlapped the registered actual prosthesis. The authors interpret the limited local differences as consistent with a patient-matched conditional geometry distribution rather than a demand that every sample reproduce one postoperative solution.
Still short of a surgical plan
The preprint's limitations are practical as well as scientific. Its single-centre retrospective cohort may not generalise across patient populations, CT protocols or prosthesis inventories. If a requested model or size conflicts with the anatomy, the catalogue geometry could be distorted.
The output is continuous geometry, not a directly actionable commercial model-and-size selection. A bone-only result would still need measurement or retrieval from a model library, followed by reduction and assembly, before it could complete a surgical plan.
Raw and processed data cannot be made public because of patient privacy, ethics and institutional data-governance restrictions. The source code is reported as available at the GitHub repository listed by the preprint. Funding was not reported in the supplied text.
The disclosure statement says Y.W. and J.S. are employees of Changzhou Jinse Medical Information Technology Co., Ltd., while J.L. and L.W. reported no competing interests.
For now, the result is best understood as a technical demonstration in a single retrospective dataset. It shows a route to generating patient-conditioned geometry, while leaving catalogue selection, measurement and assembly outside the model's current output.
Paper data and sources
Original title: THA-Flow Generative Model: Prosthesis Geometry Prediction from Preoperative CT
Authors: Yiping Wang, Jie Li, Jingyu Shen, Liao Wang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text