Personalized Knee Geometry Modeling Based on Multi-Atlas Segmentation and Mesh Refinement

Abstract

The development of personalized finite element models of the knee anatomy is critically important in the simulation of knee joint mechanics, prediction of optimal treatments in cases of pathological conditions and prevention of injuries. Subject-specific models can be obtained from diagnostic images with multi-atlas segmentation being a pertinent choice when prior anatomical information of the structures of interest is available. Although multi-atlas segmentation has been prevalent in some parts of the body, its exploitation for the segmentation of the knee complex has not been illustrated yet. This work utilizes a multi-atlas segmentation method based on deformable registration and joint label fusion in conjunction with anatomically-adopted mesh refinement in order to generate subject-specific models of the knee. The success of finite element simulations strongly depends on the properties of the 3D surface and the quality of the volumetric meshes. Therefore, emphasis was given to create structured meshes with well-shaped hexahedra for the knee cartilages and menisci. The segmentation performance is assessed using cross-validation on 7 subjects from the Open Knee project and 78 subjects from the Osteoarthritis Initiative. Our results indicate that our developed state-of-the-art processing scheme can achieve competitive performance, opening the path for better diagnostics and patient-specific interventions. The developed tools are freely available to download from SimTK at https://simtk.org/projects/knee-segment.

Real-Time Prediction of Joint Forces by Motion Capture and Machine Learning

“…No subject-specific information on knee geometry was available, hence overestimation of actual knee forces might have infiltrated our results, as shown in [58], where estimated medial knee forces were compared between a generic (uninformed) knee model and a (fully informed) model with known contact point distance and tibiofemoral alignment, extrapolated from medical images. Our future plans include the exploitation of previous work of our group on the automatic generation of subject-specific and anatomically adopted knee geometry meshes [78] that allow one to personalize the KCF computations, if imaging information is available. Additionally, the reported forces in the anteroposterial and mediolateral directions must be cautiously considered, since no mathematical formulation is incorporated in the knee model that allows the sharing of the prescribed loads between the two knee contact points in the respective planes of motion.…”

Total Knee Replacement: Subject-Specific Modeling, Finite Element Analysis, and Evaluation of Dynamic Activities

“…The FE construction procedure includes reconstructing the knee bone surfaces through image segmentation techniques on patient-specific pre-operational Computed Tomography (CT) data (Benos et al, 2020;Stanev et al, 2020). Many studies, like Woiczinski et al (2016), Li et al (2017), Moewis et al (2018), Nikolopoulos et al (2020), Su et al (2020), and Xu et al (2020) that aim at developing patient-specific FE models, adopt this method. These 3D bone models were later aligned with the knee joint's implant components by using segmentation software on post-operative CT images.…”

“…As suggested in a lot of studies (Kiapour et al, 2014;Marra et al, 2015;Woiczinski et al, 2016;Li et al, 2017;Moewis et al, 2018;Benos et al, 2020;Nikolopoulos et al, 2020;Stanev et al, 2020;Su et al, 2020;Xu et al, 2020), the best way to obtain a patient-specific 3D representation of the knee joint bones is to segment them from MRI or CT image data obtained from the subject under examination. For doing so, a segmentation method was performed with 3D Slicer (Kikinis et al, 2014) to extract the geometries of the knee bones before TKA, using subject-specific pre-operational CT scans obtained from the 6th GCC data set.…”

Machine learning methods to support personalized neuromusculoskeletal modelling

“…Regarding musculoskeletal modelling, machine learning has been used to rapidly and automatically process medical imaging to isolate structures of interest. In particular, artificial neural networks have proven particularly effective in medical image processing and have been used in a wide range of applications from automatically determine body composition (Hemke et al 2020), cartilage pathologies (Liu et al 2018) and geometries (Nikolopoulos et al 2020), bone geometries (Ambellan et al 2019), and muscle volumes (Yeung et al 2019) and geometries (Ni et al 2019). Using a dataset of reconstructed anatomical structures or organs exists, statistical shape models have been used to extract features from anatomical data, using principal component analysis (Rodriguez-Florez et al 2017;Varzi et al 2015;Williams et al 2010) to create representations of anatomical tissue with associated principal components for bone (Grant et al 2020;Suwarganda et al 2019;Zhang and Besier 2017;Zhang et al 2014), cartilage (albeit indirectly) (Van Dijck et al 2018), meniscus (Dube et al 2018;Vrancken et al 2014), and other connective tissues (Neubert et al 2015).…”