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

Area of Science:

• Biomedical Engineering
• Radiology
• Orthopedics

Background:

• Knee osteoarthritis (OA) is a major cause of disability, with patellofemoral joint OA being understudied.
• Early OA detection is crucial, and mechanical changes in patellar cartilage are key indicators.
• Current MR-based methods for assessing patellar cartilage mechanics rely on manual segmentation, which is time-consuming and prone to variability.

Purpose of the Study:

• To develop convolutional neural networks (CNNs) for automated segmentation of the patella and patellar cartilage in knee MR scans.
• To evaluate the performance and reliability of these CNNs in measuring exercise-induced cartilage deformations.

Main Methods:

• Developed and compared 2D and 3D U-Net convolutional neural networks for segmentation using a dataset of 109 knee MR scans.
• Evaluated segmentation accuracy using the mean Dice Similarity Coefficient (mDSC).
• Assessed the reliability of the best-performing networks and their ability to detect cartilage deformations after a hopping exercise.

Main Results:

• The 2D U-Net achieved higher segmentation accuracy for both patella (mDSC: 0.967) and patellar cartilage (mDSC: 0.896) compared to the 3D U-Net.
• The 2D U-Nets demonstrated excellent reliability in measuring patellar cartilage thickness (ICC = 0.99).
• Significant mean (1.5%) and maximum (10.6%) patellar cartilage strains were detected following hopping exercise.

Conclusions:

• Automated segmentation using 2D U-Nets provides a reliable and efficient method for analyzing patellar cartilage.
• These AI tools offer a powerful framework for in vivo assessment of patellar cartilage mechanics.
• This approach can advance the study of early patellofemoral joint osteoarthritis.