Author(s): GUZZI, A., LEDERGERBER, R., FAUDE, O., KELLER, M., RITSCHE, P., Institution: UNIVERSITY OF BASEL, Country: SWITZERLAND, Abstract-ID: 1683

INTRODUCTION:
Quantification of the anatomical cross-sectional area (ACSA) of the patellar tendon is clinically relevant for monitoring tendon adaptations, injuries, and rehabilitation outcomes. Although several openly available and validated tools exist for automatic segmentation of muscle ACSA, no open-source and peer-reviewed solution is currently available for the patellar tendon. Therefore, the aim of this study was to develop and evaluate an automatic approach for patellar tendon ACSA segmentation from ultrasound images and to compare its performance with manual analysis.
METHODS:
Ultrasound images were obtained from 30 participants (age: 46.87 ± 6.03 years; BMI: 25.45 ± 4.14 kg/m²). Transverse images were acquired at 25%, 50%, and 75% of the patellar tendon length. To determine the consistency of the reference method, intra-rater and inter-session reliability were assessed using manual ACSA segmentation. A total of 497 labeled images were used to train and test three neural networks for automatic segmentation. Model performance was evaluated by comparing automatic and manual measurements using intraclass correlation coefficients (ICC), standard error of measurement (SEM), mean absolute error (MAE), and standardized mean difference (SMD). Additionally, model inter-session reliability was also examined.
RESULTS:
Manual intra-rater reliability was good (ICC = 0.804; 95% CI: 0.628–0.902), with SEM = 0.05 cm² (0.03–0.07) and MAE = 0.05 cm² (0.04–0.07). Inter-session reliability of manual segmentation was excellent (ICC = 0.980; 95% CI: 0.970–0.987), with SEM = 0.02 cm² (0.02–0.02) and MAE = 0.02 cm² (0.01–0.02). After exclusion of erroneous automatic predictions, agreement between automatic and manual segmentation was excellent (ICC = 0.918; 95% CI: 0.821-0.955), with SEM = 0.05 cm² (0.04–0.07) and MAE = 0.05 cm² (0.05–0.06). A small standardized mean difference was observed (SMD = 0.53; 95% CI: 0.33–0.75). Model inter-session reliability was moderate (ICC = 0.708; 95% CI: 0.471–0.849). Automatic analysis time per image ranged from 0.302 to 0.414 s.
CONCLUSION:
The proposed automatic segmentation approach enables rapid and less operator-dependent assessment of patellar tendon ACSA from ultrasound images. While some differences between manual and automatic measurements and moderate model inter-session reliability were observed, the method demonstrates acceptable agreement and substantial time efficiency. With cautious implementation and appropriate quality control, this open-source tool may provide valuable support for clinical assessment and research applications involving patellar tendon morphology.