Osteoarthritis (OA) is a common chronic joint disease characterized by articular cartilage damage, joint pain, and limited function. There are around 654.1 million individuals with knee OA among people over 40 years old in 2020 worldwide (Cui et al., 2020). Joint disease such as OA is closely related to the change in joint mechanics. Musculoskeletal simulation provides useful information, such as muscle force and joint contact force, which are significant for understanding the initiation, progression, and treatment of OA.

Generally, musculoskeletal models are scaled based on the distance between skin-surface markers to accommodate variations in subject size (Rahman et al., 2022, Wu et al., 2022). However, in obese populations and patients with bone deformity, it’s challenging to represent their skeletal anatomy with marker-based linear scaling approach (Horsak et al., 2017, Kainz et al., 2017). The marker-based linear scaling method has low bone model accuracy, with one recent study reporting root mean square error (RMSE) of 13.95 mm and 9.85 mm for pelvis and femur, respectively. (Suwarganda et al., 2019). More advanced methods, such as statistical shape modeling (SSM), can used to reconstruct geometric bone models from surface markers or medical images (Bahl et al., 2019, Shi et al., 2022, Suwarganda et al., 2019). Image-based SSM method greatly enhanced the accuracy of bone models, with one study reporting a reduction of 89.9 % and 85.7 % in the RMSE for the pelvis and femur, respectively, compared to marker-based linear scaling method (Suwarganda et al., 2019). However, image-based SSM method is difficult to implement in clinical practices, as this method is cumbersome and costly, and 3D medical images are not a standard daily examination for the diagnosis and treatment of OA patients (Suwarganda et al., 2019). Marker-based SSM method exhibited moderate bone model accuracy, with one study reporting a reduction of 54.7 % and 56.9 % in the RMSE for the pelvis and femur, respectively, compared to marker-based linear scaling method (Suwarganda et al., 2019). Despite its independence from 3D medical imaging, the manual selection of key anatomical landmarks remains a cumbersome process.

Among OA patients, the full-length radiograph which is a standard daily examination for the diagnosis and treatment contains many anatomical parameters of lower limbs, such as femoral length, tibial length, width and height of the pelvis, hip-knee-ankle (HKA) angle and femoral neck-shaft angle (NSA) (Sabharwal and Zhao, 2009). These anatomical parameters can be incorporated into the musculoskeletal model to enhance model accuracy. Previous study reported that a 1 % scaling error of the thigh could resulting in a 72 % deviation in the peak hip joint contact force (Koller et al., 2021). These scaling errors can be avoided by scaling segments to their actual length measured from the full-length radiograph. A previous study included HKA angle measured from full-length radiograph in the scaled generic musculoskeletal model to obtain a more accurate estimation of knee contact forces in knee OA patients with non-neutral alignment (Lerner et al., 2015). Furthermore, changing the NSA can substantially affect the estimation of hip contact forces (Bosmans et al., 2014, Lenaerts et al., 2009). Although these anatomical parameters can enhance the precision of musculoskeletal models, incorporating anatomical parameters measured from full-length radiograph into musculoskeletal models is cumbersome and requires specialized skills. Therefore, when addressing larger-scale knee OA populations, efforts should be paid directed towards developing a fully automatic algorithm capable of extracting anatomical parameters from full-length radiograph to generate a musculoskeletal model.

Deep learning has been used to segment medical images rapidly and automatically, such as full-length radiograph (Saxby et al., 2020). Subsequently, parameters of interest can be extracted from the segmentation bone masks. Schock et al. segmented the full-length radiograph using a U-Net convolutional neural network and extracted the HKA angle and femoral anatomic-mechanical angle from the segmentation (Schock et al., 2021). In addition, leg length discrepancy in children was automatically measured from the full-length radiograph segmentation segmented using a convolutional neural network model with mixed residual connections (Zheng et al., 2020). Although these two studies employed different convolutional neural network models, both achieved comparable segmentation performance, with mean Sørensen-Dice coefficients of approximately 0.97. Therefore, the goal of this study was to use deep learning method for bone segmentation in the full-length radiograph and develop an algorithm that could automatically generate a musculoskeletal model by extracting anatomical parameters from the segmented bone masks. This full-automatic musculoskeletal modeling method could be applied in personalized musculoskeletal analysis for larger-scale OA populations.