Prediction of knee ligament parameters using deep neural network models

Accurate characterization of ligament mechanics is essential for realistic knee joint simulations, but in vivo quantification of ligament properties remains infeasible. Existing approaches rely on generic cadaveric data or on time-consuming optimization procedures to obtain subject-specific parameters. This study investigates the potential of deep neural networks (DNNs) to estimate ligament parameters from knee laxity data. A tibiofemoral joint model was used to generate a synthetic dataset comprising 200,000 laxity simulations of 100 virtual knees with varied stiffness and reference strain values for the anterior and posterior cruciate, medial and lateral collateral ligaments. Twenty-three "optimal" laxity tests were identified to train eight DNN models, using increasing dataset sizes. Each DNN model was dedicated to one ligament parameter. Prediction accuracy improved with more samples However, 1,000-4,000 laxity test samples were required to achieve errors below 5% across all parameters. Biases smaller than 40N per strain units and 0.013 strain units and limits of agreement smaller than 380N per strain units and 0.15 strain units were observed for all ligament stiffnesses and reference strains, respectively. These results encourage further research into the use of DNN models for predicting ligament parameters as a fast alternative to traditional optimization methods. Future work should expand training data, include anatomical variability and validate the approach on experimental datasets to enable subject-specific ligament modeling.