A deep learning-based approach for measuring patellar cartilage deformations from knee MR images

Abstract

While knee osteoarthritis (OA) is a leading cause of disability in the United States, OA within the patellofemoral joint is understudied compared to the tibiofemoral joint. Mechanical alterations to cartilage may be among the first changes indicative of early OA. MR-based protocols have probed patellar cartilage mechanical function by measuring deformations in response to exercise. These studies, however, often rely on manual segmentation, which is time-intensive and may introduce variability. Therefore, our goals were (1) to develop convolutional neural networks to segment the patella and patellar cartilage from knee MR scans and (2) to evaluate the ability of these networks to measure exercise-induced cartilage deformations. Using a dataset of 109 knee MR scans, 2D and 3D U-Nets were developed and compared using the mean dice similarity coefficient (mDSC). Reliability of the best-performing networks was examined and the ability of these networks to detect patellar cartilage deformations following a hopping activity was evaluated. The 2D U-Net outperformed the 3D U-Net for both the patella (mDSC, 2D: 0.967 vs 3D: 0.960) and patellar cartilage (mDSC, 2D: 0.896 vs. 3D: 0.895). The 2D U-Nets demonstrated excellent reliability (ICC = 0.99, mean difference < 0.03 mm) in reproducing the mean patellar cartilage thickness across different days. Lastly, significant mean (mean ± standard deviation, 1.5 ± 1.8 %, P = 0.014) and maximum (10.6 ± 3.2 %, P < 0.001) patellar cartilage strains were detected following hopping. The autosegmentation tools developed herein provide a powerful framework for probing patellar cartilage mechanics in vivo.