The number of total knee arthroplasties performed world-wide are on a rise. Patient-specific planning and implants may improve surgical outcomes, but require 3D models of the bones involved. Ultrasound may become a cheap and non-harmful imaging modality, if the short comings of segmentation techniques in terms of automation, accuracy and robustness are overcome. Furthermore, any kind of ultrasound-based bone reconstruction must involve some kind of model completion, in order to handle occluded areas, e.g., the frontal femur. A fully-automatic and robust processing pipeline is proposed, generating full bone models from 3D freehand ultrasound scanning. A convolutional neural network is combined with a statistical shape model to segment and extrapolate the bone surface. We evaluate the method in-vivo on 10 subjects, comparing the ultrasound-based model to a magnetic resonance imaging (MRI) reference. The partial freehand 3D record of the femur and tibia bones deviate by 0.7 to 0.8mm from the MRI reference. After completion and on average, the full bone model shows sub-millimetric error in case of the femur and 1.24mm in case of the tibia. Processing of the images is performed in real-time, and the final model fitting step is computed in less than one minute. On average, it took 22 minutes for a full record per subject.