Accurate representation of subject-specific bone anatomy in lower-limb musculoskeletal models is important for human movement analyses and simulations. Mathematical methods can reconstruct geometric bone models using incomplete imaging of bone by morphing bone model templates, but the validity of these methods has not been fully explored. The purpose of this study was to determine the minimal imaging requirements for accurate reconstruction of geometric bone models. Complete geometric pelvis and femur models of 14 healthy adults were reconstructed from magnetic resonance imaging through segmentation. From each complete bone segmentation, three sets of incomplete segmentations (set 1 being the most incomplete) were created to test the effect of imaging incompleteness on reconstruction accuracy. Geometric bone models were reconstructed from complete sets, three incomplete sets, and two motion capture-based methods. Reconstructions from (in)complete sets were generated using statistical shape modelling, followed by host-mesh and local-mesh fitting through the Musculoskeletal Atlas Project Client. Reconstructions from motion capture-based methods used positional data from skin surface markers placed atop anatomic landmarks and estimated joint centre locations as target points for statistical shape modelling and linear scaling. Accuracy was evaluated with distance error (mm) and overlapping volume similarity (%) between complete bone segmentation and reconstructed bone models, and statistically compared using a repeated measure analysis of variance (p<0.05). Motion capture-based methods produced significantly higher distance error than reconstructions from (in)complete sets. Pelvis volume similarity reduced significantly with the level of incompleteness: complete set (92.70±1.92%), set 3 (85.41±1.99%), set 2 (81.22±3.03%), set 1 (62.30±6.17%), motion capture-based statistical shape modelling (41.18±9.54%), and motion capture-based linear scaling (26.80±7.19%). A similar trend was observed for femur volume similarity. Results indicate that imaging two relevant bone regions produces overlapping volume similarity >80% compared to complete segmented bone models, and improve analyses and simulation over current standard practice of linear scaling musculoskeletal models.