Deep learning-based femoral reconstruction from intraoperative point clouds for enhanced knee arthroplasty...

Computer- and robotic-assisted technologies improve total knee arthroplasty (TKA) through intraoperative bone registration. However, limited bone exposure restricts point collection to the distal femur, omitting key geometric features and reducing registration accuracy and the surgical outcomes.

We introduce a deep learning-based method to improve bone registration by reconstructing the full femur from intraoperative point clouds. To reconstruct the femur, a point completion network (PCN) is fine-tuned in two stages: first, using points covering the entire femur and then only using the distal parts, which are usually resected. The reconstructed bone enhances iterative closest points registration by adding geometric details from otherwise inaccessible regions, significantly improving accuracy due to their distance from distal points. We evaluate on 35 VSD and 5 real femurs by measuring reconstruction accuracy using mean Chamfer and Euclidean distances, and by assessing registration error in rotation and translation with points sampled from the proximal epiphysis-metaphysis, femoral head, and shaft.

Point reconstruction errors averaged $2.36mm$ for VSD and $2.0mm$ for real femurs. The best registration performance is achieved by randomly sampling points across the entire reconstructed bone, offering superior geometric coverage. This approach outperformed the baseline ICP and chosen deep learning registration methods, reducing varus-valgus angle errors by $50\%$ , flexion-extension and internal-external errors by slightly less, and translation errors by over $1mm$ .

Our approach uses PCN to reconstruct complete femurs from limited surgical data, leveraging inaccessible regions to improve registration accuracy for computer-assisted TKA.