Automated deep-learning quantification of nine patellofemoral instability parameters on multislice CT images : development and validation of the GU2Net model.
Authors
Affiliations (5)
Affiliations (5)
- Center for Rehabilitation Medicine, Department of Radiology, Zhejiang Provincial People's Hospital, Affiliated People's Hospital, Hangzhou Medical College, Hangzhou, Zhejiang, China.
- Department of Radiology, The People's Hospital of Pingyang, Wenzhou, China.
- The Second School of Clinical Medicine, Hangzhou Normal University, Hangzhou, Zhejiang, China.
- Taizhou Integrated Traditional Chinese and Western Medicine Hospital, Taizhou, Zhejiang, China.
- Department of Radiology, Tongxiang First People's Hospital, Jiaxing, Zhejiang, China.
Abstract
Objective and precise measurement of patellar instability (PI) parameters on CT images is essential for accurate diagnosis and treatment planning. However, manual assessment is tedious, time-consuming, and prone to error. This study aimed to develop and validate a deep learning model that automatically quantifies PI parameters on axial knee CT images. CT scans of 1,125 knees were randomly divided into training, validation, internal test, and hold-out test sets. A deep learning-based model was trained to localize anatomical landmarks and calculate nine PI parameters: lateral patellar tilt (LPT), bisect offset ratio (BSO), congruence angle (CA), sulcus angle (SA), trochlear groove depth (TGD), lateral trochlear inclination (LTI), trochlear groove-transepicondylar axis (TG-TEA) distance, tibial tubercle-trochlear groove (TT-TG) distance, and tibial tubercle-Roman arch (TT-RA) distance. Model performance was compared with manual measurements using the successful detection rate, mean absolute error (MAE), intraclass correlation coefficient (ICC), and Pearson correlation coefficient. The model accurately predicted landmark locations (MAE 0.84 to 2.72 mm) and PI parameters (ICC 0.826 to 0.997, <i>r</i> 0.705 to -0.994, pā<ā0.001) except for SA (ICC 0.701 to 0.862, <i>r</i> 0.542 to 0.744, pā<ā0.001). On the hold-out test set, the model outperformed inexperienced radiologists for LPT, CA, SA, LTI, and TGD (model: ICC 0.701 to 0.996, r 0.542 to 0.992, pā<ā0.001; radiologists: ICC 0.413 to 0.959, r 0.281 to 0.923, p<i>ā</i><ā0.05). The proposed deep learning model reliably automates PI measurement, reducing the time and variability associated with manual assessment and mitigating dependence on examiner experience.