Efficient development of subject-specific finite element knee models: Automated identification of soft-tissue attachments

Abstract

Musculoskeletal disorders impact quality of life and incur substantial socio-economic costs. While in vivo and in vitro studies provide valuable insights, they are often limited by invasiveness and logistical constraints. Finite element (FE) analysis offers a non-invasive, cost-effective alternative for studying joint mechanics. This study introduces a fully automated algorithm for identifying soft-tissue attachment sites to streamline the creation of subject-specific FE knee models from magnetic resonance images. Twelve knees were selected from the Osteoarthritis Initiative database and segmented to create 3D meshes of bone and cartilage. Attachment sites were identified in three conditions: manually by two evaluators and via our automated Python-based algorithm. All knees underwent FE simulations of a 90° flexion-extension cycle, and 68 kinematic, force, contact, stress and strain outputs were extracted. The automated process was compared against manual identification to assess intra-operator variability. The attachment site locations were consistent across all three conditions, with average distances of 3.0 ± 0.5 to 3.1 ± 0.6 mm and no significant differences between conditions (p = 0.90). FE outputs were analyzed using Pearson correlation coefficients, randomized mean square error, and pairwise dynamic time warping in conjunction with ANOVA and Kruskal-Wallis. There were no statistical differences in pairwise comparisons of 67 of 68 FE output variables, demonstrating the automated method's consistency with manual identification. Our automated approach significantly reduces processing time from hours to seconds, facilitating large-scale studies and enhancing reproducibility in biomechanical research. This advancement holds promise for broader clinical and research applications, supporting the efficient development of personalized musculoskeletal models.

Keywords: Automated; Finite element; Knee; Ligament attachments; Subject-specific modeling.

Figure 1.

Muscle and ligament structures included in the FE model.

Figure 2.

The attachment site identification process. (A) the template model is built based on experimental dissection and probing, and anatomical landmarks, (B) the subject-specific model is built based on knee MRI, (C) the subject-specific model is scaled in medial-lateral and anterior-posterior directions and superimposed on the template, (D) for each attachment site, the nearby template anchor points are found, (E) for template anchor point, the equivalent subject anchor point is determined, (F) subject model scaled back to original size, (G) the vectors going from the template anchor points to the attachment site are determined, (H) these vectors are projected to their equivalent subject anchor point after scaling, (I) the weighted average of vector endpoints is calculated based on distance from the attachment site, (J) the subject attachment site is determined, (K) this process repeated for all attachment sites makes the subject specific musculoskeletal model.

Figure 3.

The local anatomical landmarks and coordinate system automatically determined, (A) most distal points on medial and lateral femoral condyles, (B) dwell points on medial and lateral tibial condyles, (C) medial and lateral tibial intercondylar tubercles, (D) Grood and Suntay axes for the knee.

Figure 4.

Flowchart of the overall study design. RMSE: root mean squared error, DTW: dynamic time warping, MRI: magnetic resonance image, FE: finite element, CC: correlation coefficient

Figure 5.

(A) The FE model of a deep knee bend activity, (B) von Mises stress in tibial and patellar cartilages in full extension and (C) at maximum flexion.

Figure 6.

The between-condition comparison for the distances between attachment site nodal locations. E1: Evaluator 1, E2: Evaluator 2, Au: Auto

Figure 7.

Similarity between conditions across all output variables based on an RMSE-based similarity index (left) and Pearson Correlation Coefficient (right). Each bar (slice) represents the similarity between the two manual models for a given variable, with the variable indices shown around the plot. The length of each bar was computed by averaging the relative index across all subjects for that variable. Red and blue markers indicate the similarity between the automated model and the manual models from evaluator 1 (red) and evaluator 2 (blue). The colors of the bars are used solely for visual distinction between variables and do not represent any specific value. The RMSE-based similarity index was calculated for plotting purposes by first normalizing the RMSE values via dividing them by their maximum value, and then reversed so that higher scores denote greater similarity between the conditions. The complete list of variables and their raw and normalized RMSE values can be found in Appendix 1. * A significant difference between pairwise comparisons was found only for patellofemoral medial-lateral translation (variable 13) as denoted by an asterisk.

Figure 8.

Average kinematic outputs across all subjects (showing a representative sample of kinematic outputs), with shaded regions illustrating the 25^th and 75^th percentile for each condition. A-P: anterior-posterior, M-L: medial-lateral, I-E: internal-external, V-V: varus-valgus, S-I: superior-inferior.

Figure 9.

Mean force and contact outputs across all subjects (showing a representative sample of soft-tissue forces, contact areas, joint forces and center of pressure outputs), with shaded regions illustrating the 25^th and 75^th percentile for each condition. CoP: center of pressure, PAT: patellar cartilage, TIB_MED: tibial medial cartilage. A-P: anterior-posterior, M-L: medial-lateral. Note: The mean ACL total force is higher than the 75^th percentile for a portion of the cycle because the mid-cycle ACL force drops to zero in several knees, skewing the 75^th percentile lower than the mean.

Figure 10.

Average 90^th percentile first principal logarithmic strain, and 50^th and 95^th percentile von Mises stress across all subjects, with shaded regions illustrating the 25^th and 75^th percentile for each condition. FEM: femoral cartilage, PAT: patellar cartilage, TIB_MED: tibial medial cartilage.