14:30   Lower extremities & Motion II Chair: Dennis Janssen AI-accelerated prediction of optimal implant alignment in total knee arthroplasty Nico Verdonschot, Linda ten Klooster, Periklis Tzanetis, Jelmer Wolterink Abstract: Background: Total knee arthroplasty (TKA) is effective for treating end-stage knee osteoarthritis (OA). However, dissatisfaction for one in five patients persists. A novel musculoskeletal model-based approach has been previously developed to predict the optimal implant position toward recreating the knee's pre-diseased functional behaviour, ultimately improving patient satisfaction post-operatively1. However, its computational complexity poses challenges for its implementation in clinical practice. The study aims to accelerate optimal implant alignment estimation using an AI-based surrogate model for real-time predictions. Methods: A previously obtained dataset of 21 knee OA patients2 was employed to train a neural network. It consists of candidate implant positions and their deviations from the knee’s pre-diseased state. The neural network was trained to learn this relation for individual patients, acting as a surrogate model enabling rapid determination of the optimal position with minimal error. Results: For all patients, candidate implant positions were examined to assess the network’s ability to predict the optimal position. Strong correlation between predictions and computed values in the musculoskeletal model was shown in 16 patients, while for the 5 other patients this relation was weaker. Evaluation metrics include Spearman’s rank coefficient, Kendall Tau’s rank coefficient, and mean squared error. Factors that affect model performance include limited data and complexity in distinguishing subtle implant position variations. The computation time reduced drastically from approximately 32 hours to a maximum of 7 seconds for all patients. Conclusion: The neural network significantly accelerates the prediction of optimal implant positioning in TKA, potentially aiding orthopaedic surgeons to predict optimal implant alignment in a personalized matter. Ultimately this may contribute to improving patient satisfaction. References 1. Tzanetis P, Fluit R, Souza K de, Robertson S, Koopman B, Verdonschot N. 2023. Pre-planning the surgical target for optimal implant positioning in robotic-assisted total knee arthroplasty. Bioengineering 10: 543. 2. Tzanetis P, Souza K De, Robertson S, Fluit R, Koopman B, Verdonschot N. 2024. Numerical study of osteophyte effects on preoperative knee functionality in patients undergoing total knee arthroplasty. J Orthop Res 42: 1943–1954. 3D GRF estimation during daily activities using 3 IMUs and Pressure Insoles Alessandro Castellaz, Frank Wouda, Bert-Jan van Beijnum Abstract: Accurate estimation of three-dimensional ground reaction forces (3D GRF) is essential for understanding joint loading, particularly in conditions like knee osteoarthritis (OA). Combining 3D GRF with lower limb kinematics provides valuable insights into joint forces, which can inform treatment and rehabilitation strategies. Traditional methods, however, require advanced laboratory setups, limiting their applicability in real-world environments. This study aims to develop a portable system for estimating 3D GRF during daily activities, using three inertial measurement units (IMUs) and pressure insoles, validated against force shoes as a reference. Building on the methods of Refai et al. [1], which track the relative distance between the feet and the center of mass (CoM), and Castellaz et al.'s [2] Virtual Pivot Point (VPP) concept, we estimate 3D GRF by combining data from three IMUs (one on the pelvis and one on each foot) with pressure insoles. The total GRF is computed from the pelvis IMU and split into individual foot components using the VPP approach. In this preliminary study, data from one healthy participant were analyzed during various activities, including the Timed Up and Go (TUG) test, free walking around a living environment, stair ascent/descent, and treadmill walking at two different speeds (1.5 and 2.5 km/h). The results demonstrate strong accuracy in estimating shear forces (anterior-posterior and mediolateral), with both components showing an average correlation of 0.89 and 0.87 and relative root mean square errors (rRMSE) of 6.11% and 6.12% for AP and ML components respectively. The vertical ground reaction forces (GRF) estimation yielded even higher accuracy, with an average correlation of 0.97 and an rRMSE of 1.96%. All estimates were validated through comparison with force shoes, serving as the reference. In conclusion, this method accurately estimates 3D GRF using a portable system, making it a viable alternative to traditional laboratory-based setups. While this is a preliminary study based on data from a single participant, the approach holds the potential for developing an ambulatory system to monitor biomechanics and joint loading in daily life, which could support the management of conditions such as knee OA in real-world settings. Threshold fluctuations in single human motor axons: insights from low- and high- threshold motor units in patients with ALS and healthy controls Isabelle A.M. Busman, Diederik J.L. Stikvoort García, H. Stephan Goedee, Leonard H. van den Berg, Boudewijn T.H.M. Sleutjes Abstract: Background: Human motor axons show intrinsic threshold variability in response to nerve stimulation, reflecting axonal excitability characteristics. Research on this variability largely focuses on healthy, low-threshold motor units (MUs). In amyotrophic lateral sclerosis (ALS), altered motoneuron excitability and preferential low-threshold MU loss hint at intrinsic differences within the MU pool. Assessing threshold variability across the MU spectrum could deepen our understanding of motoneuron properties and enhance motor unit estimation techniques for sensitive neurodegeneration monitoring. Research question: How does threshold variability differ across the MU pool in ALS patients compared to healthy controls, and how does it relate to MU size and their axonal threshold levels? Objective: We aim to quantify the threshold variability of single MUs from compound muscle action potential (CMAP) scans in healthy controls and ALS patients. Methods: We retrospectively analysed thenar CMAP scans of ALS patients and healthy controls. Threshold variability was quantified by relative spread (RS), defined as the standard deviation divided by the mean threshold. We further assessed MU sizes and defined low-to- high threshold axons based on their mean threshold relative to the stimulus current range of the CMAP scan. Results: We have analysed 119 CMAP scans from ALS patients and 51 CMAP scans from healthy controls. In total, 296 single MUs were identified (ALS, n=204; healthy controls, n=92). In ALS patients, a median RS value of 1.25% was found (5%-95%: 0.32%-2.76%) compared to 1.33% (5%-95%: 0.47%-2.55%) in healthy controls. We found a significant negative correlation between RS values and low-to-high threshold axons in ALS patients (from S0–S100 ΔRS (%) = -0.005, CI: -0.006 to -0.001, p<0.05). Conclusion: Increased threshold variability of low-threshold compared to high-threshold MUs in ALS patients may indicate altered axonal excitability among the MU pool. This variation may reflect differences in susceptibility and the heterogeneous nature of underlying neurodegenerative disease processes. Automatic Robotic Ultrasound Scanning for Muscle Segmentation and Reconstruction Dezhi Sun, Alessandro Cappellari, Bangyu Lan, Kenan Niu Abstract: Ultrasound (US) imaging is widely utilized for medical diagnostics thanks to its non-invasiveness, real-time imaging capability [1], and cost-effectiveness, making it ideal for visualizing and localizing soft tissues. Recent advancements in US technology allow for real-time 3D model reconstruction, enabling clinicians to dynamically monitor complex anatomical structures, which is especially valuable for analyzing muscle behavior during movement [2,3]. This capability benefits athletes, trainers, and clinicians, and the development of automated scanning systems further supports personalized care through patient-specific model generation. In this study, an automatic robotic ultrasound scanning (ARUS) system was developed to perform real-time segmentation and reconstruction of muscle and bone tissues. The system integrates a robotic arm, a US imaging system, and a stereo camera to perform automatic ultrasound scanning and real-time muscle reconstruction. A custom calibration approach was employed to enhance 3D reconstruction accuracy and consistency [4]. The ultrasound images are transmitted in real-time via UDP protocol from the Windows workstation to the main processing unit, which operates on an Ubuntu system and is responsible for data processing and robotic control. A hybrid control strategy was employed to ensure stable contact force, while visual servoing allowed real-time adjustments to the probe’s position to accommodate dynamic muscle contours and improve reconstruction quality. The ARUS system was evaluated using a customized muscle phantom. The evaluation was conducted from two aspects, i.e., image segmentation and 3D reconstruction. The segmentation model used was an attention-enhanced U-Net, achieving a high Dice coefficient of 0.84 and an Intersection over Union (IoU) score of 0.85, indicating robust segmentation performance. For reconstruction, the ARUS system yielded an average spatial error of 0.94 mm, with a root mean square error (RMSE) of 1.22 mm, when comparing with MRI reference data. In this study, an automatic robotic ultrasound scanning (ARUS) system was developed for real-time segmentation and 3D reconstruction of muscle and bone tissues. Evaluation results demonstrated promising performance in both segmentation and reconstruction, indicating that the ARUS system holds significant potential for dynamic muscle monitoring and personalized diagnostic and therapeutic applications. Wide-band Dielectric Tissue Property Characterization by Quantitative Multinuclear MRI based Tissue Composition Assessment Laura JC Barendsz, Kemal Sumser, Desmond HY Tse, Jacobus FA Jansen, Rob MC Mestrom, Maarten M Paulides Abstract: Medical device design and device safety assessment increasingly rely on computational modeling-based methods. Accurate knowledge of the dielectric properties of human tissue therefore becomes increasingly crucial. In this study, we investigated the feasibility for determining dielectric tissue properties in the frequency range of 50 to 600 MHz using mixture models combined with quantitative multinuclear MRI-based tissue composition electrical property tomography (EPT): “1H23NaTiCEPT”. The composition of muscles can be approximated by protein, fat, sodium, and water. Four phantoms with different compositions were prepared (healthy, obese, obese+, and dehydrated). Mixture models were used to relate the dielectric properties (DP) of the muscle to its constituents. The Looyenga mixture model was used for the relative permittivity, and the Maxwell-Garnett equation for the effective conductivity. The constituents of muscle are saline as background solution, protein and fat as additive, and the water fraction is used to find the volume fraction of the additive. All MRI measurements were performed using a Siemens 7T whole-body MRI system at Scannexus (Ultra-High-Field MRI center, Maastricht). The water fraction was determined based on the measured T1 times. Calibration phantoms, consisting of water, sodium chloride and sugar, were used to correlate the T1 times with the water fraction. The sodium concentration measurements of the phantoms were performed with a 23Na knee coil using an ultrashort TE sodium sequence. Four calibration samples with water, agar, and varying salt concentrations were used for a linear fit to find the sodium content in the muscle phantoms. A rational fit was used for the water fraction and a linear one for the sodium concentration. The measured water fraction and sodium concentration were inserted into the Looyenga and Maxwell-Garnett equation. The mean error is 8.2% for the relative permittivity and 15.5% for the effective conductivity. This study demonstrates the potential of reconstructing the DP’s of human tissue by quantitative multinuclear MRI supplemented by the Looyenga and Maxwell-Garnett mixture model. Pelvic organ mobility after pelvic mesh surgery assessed with upright MRI Mart Kortman, Bruno De Santi, Frank Simonis, Anique Grob Abstract: Introduction & aim Internal rectal prolapse (IRP), is an example of pelvic organ prolapse (POP) and comes with symptoms such as obstructed defecation and/or fecal incontinence. IRP can surgically be treated by implanting a mesh in the pelvis to fixate the organs, so called sacrocolporectopexy (SCRP). After SCRP, anatomical prolapse is reduced in >90% of the patients, however, in 30-40% of the patients, symptoms persist(1). To study the effect of the mesh on pelvic organ mobility (POM), dynamic imaging studies, such as magnetic resonance defecography (MRD), were used(2). Full assessment of POM is restricted by the unphysiological supine defecatory position(3). Upright MRI can assess POM more effectively without these limitations(4). The aim of our study therefore is to use upright MRD to assess the effect of SCRP on POM and to determine the difference with previous supine measurements. Method Twenty-one female patients with POP and proven IRP underwent SCRP. Upright MRD was performed before and 6 weeks after SCRP using a 0.25T tiltable MR-system (G-scan Brio, Esaote). Analysis of 2D dynamic data included measurements of the lowest point of the bladder, vaginal vault (VV) and anorectal junction (ARJ) with respect to the reference line at rest and during maximum straining(5). POM was defined as the difference between the organ position at rest and during maximum straining, with a smaller POM indicating more organ fixation. Significance was assessed using the Wilcoxon signed-rank test (α=0.05) and results were considered clinically relevant when ΔPOM ≥10 mm. Results & conclusion A statistically significantly higher fixation was found post-operatively for all analyzed organs. Clinical relevance was found for the bladder (ΔPOM=11 mm) and VV (ΔPOM=25 mm). In comparison to previous, supine MR based studies, our POM outcomes indicate more fixation post-operatively, illustrating a body’s position dependency in the assessment of POM(2). Future recommendations Current upright MRD analysis on POM is not based on movement and reshaping of the entire organs but rather on distance measurements of anatomical landmarks. We recommend to include 3D mesh trajectory and automatic registration methods to further increase the understanding on the effect of SCRP.

end %-->