Kneeview: An open-source repository of patient-specific knee geometries and structured meshes

The exploitation of finite-element modeling to understand knee biomechanics is often constrained by data availability. While the osteoarthritis initiative (OAI) [1] provides invaluable imaging data, converting these images into finite-element meshes is time-consuming. Open models such as OpenKnee [2] are valuable but restricted by unstructured meshes and low subject count (n=8). To address this, we present a repository of 80 patient-specific bone geometries of the femur, tibia, patella and fibula that is being expanded with personalized structured finite-element models through mesh morphing. The dataset included T2, fat saturated, knee MRI scans from 93 patients (38 male, 55 female) acquired at Hospital Del Mar, Barcelona. An nnUNet neural network was trained on 35 expert annotations verified by a musculoskeletal-specialized radiologist. We created a webservice, Kneeview (https://knee.view.upf.edu), to host the repository. Subjects with missing demographics were excluded and 80 subject segmentations were reported. We customize the Bayesian Coherent Point Drift (BCPD) algorithm to treat tissues distinctively: bones undergo object-specific morphing, while soft tissues are automatically interpolated.  The nnUNet achieved high segmentation accuracy on a hold-out test dataset of 15 volumes with an average Dice coefficient of 0.986 ± 0.001 and IOU of 0.971 ± 0.002. The average Hausdorff Distance (HD95) was 14.464 ± 12.727. This variance was driven exclusively by distal tibial artefacts (Tibia HD95: 16.49 ± 14.22) while femur, patella and fibula gave an average HD95 of 1.46 ± 0.242. All annotated 3D models were successfully incorporated in Kneeview, in STL format. The finite-element mesh included the whole knee joint with ligaments, bones and menisci without distorted or zero-volume elements. A collection of 80 patient-specific models were successfully created and accessible through the new Kneeview service. The high Dice scores confirm that our automated pipeline produces geometries comparable to manual segmentation. As kneeview continues to grow, it shall address the shortage of open-source simulation-ready knee geometries. Future work will include the structured FE meshes for all subjects after these meshes have been successfully tested against finite element simulations of full gait cycles. Acknowledgements: This research study was co-funded by the European Union under the Horizon Europe MSCA Joint doctoral network inSilicoHealth with grant No.101169278, by the European Research Council (ERC-2021-CoG-O-Health-101044828), and by the Spanish Ministry of Science, Innovation, and Universities project STRATO - PID2021-126469OB-C21. Views and opinions expressed are those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the REA can be held responsible for them.