Abstract
Osteoarthritis affected 595 million people worldwide in 2020, with the knee most commonly affected and total knee arthroplasty (TKA) volumes continuing to rise. Robotic-assisted TKA improves alignment and component positioning relative to manual techniques, but most commercial image-based systems require a patient-specific 3D bone model from preoperative computed tomography (CT), imposing radiation, cost, and logistical burden that limit scalable access. Reconstructing 3D anatomy from standard radiographs offers an alternative, yet prior methods have largely been evaluated on synthetic radiographs or against global surface metrics that mask localized error in the alignment-defining and articular regions most relevant to surgical outcomes.
This dissertation introduces BoneVision, a prior-guided, region-aware deep learning framework that reconstructs patient-specific distal femur and proximal tibia surfaces from paired anteroposterior and lateral clinical radiographs. It combines a statistical shape model, a dual-view convolutional encoder, and a 3D Attention U-Net refinement network, linked by a differentiable signed distance field voxelization for end-to-end training. A hierarchical accuracy framework partitions each bone into clinically weighted tiers.
Validated under leave-one-out cross-validation on 100 real clinical cases (Kellgren-Lawrence grades II-IV) against CT ground truth, BoneVision achieved root mean square errors of 0.98 mm (femur) and 0.92 mm (tibia), Dice coefficients above 0.96, and angular errors within published CT inter-observer variability, with inference in about 2.5 seconds per case. These findings indicate radiograph-based reconstruction can approach CT-comparable fidelity for most elective TKA candidates.