Fig. 1: Framework for extracting and validating core components for a knee joint digital twin.
This figure presents a scalable framework for the systematic extraction and analysis of biomarkers from knee joint imaging, laying the groundwork for future digital twin development and studies in osteoarthritis prognosis and intervention. a Imaging acquisition and segmentation: Illustrates the dataset derived from 4796 unique patients with 3D-DESS MRI scans. Automatic segmentation of bone and cartilage is performed using neural networks, comprising ensembles of 2D and 3D V-Nets previously trained on the OAI dataset. The segmentation captures detailed anatomical structures of the knee joint. The MRI icon in this panel was created by LAFS and obtained from Flaticon. b Biomarker extraction: Demonstrates the conversion of segmentation masks into 3D point clouds for bone and meniscus shape analysis. Surface registration is achieved using Iterative Closest Point matching. Cartilage thickness is quantified using Euclidean distance transforms, while T2 relaxation times are calculated from segmented cartilage in MSME-DESS (Multi-Slice Multi-Echo Double-Echo Steady State) images using a mono-exponential fitting model. Principal component analysis is applied to reduce the dimensionality of these biomarkers, retaining critical geometric variations. c PCA visualization and expert interpretation: Radiologists use a MATLAB application to visualize and interpret principal component modes, adjusting them within ±3 standard deviations (SDs) from the mean to assess their impact. This is complemented by Clinical Cohort Matching, employing t-SNE embedding and non-parametric testing for cohort analysis in OA Incidence and Knee Replacement studies. d Feature selection and regression analysis: Elastic net regularization within a generalized linear model framework is applied, iterating over 1000 bootstrap samples for feature stability. The final logistic regression model evaluates feature importance, calculating coefficients, standard errors, p-values, confidence intervals, and odds ratios to identify significant predictors of OA incidence and knee replacement.
