Enhanced Magnetic Resonance Imaging-Based Knee Cartilage Segmentation Using a Swin-UNet Conditional Generative...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Orthopedic Surgery

Background:

• Accurate knee cartilage segmentation from MRI is vital for diagnosing knee osteoarthritis and planning surgeries.
• Manual segmentation is time-consuming, and CT-based systems lack cartilage visualization capabilities.

Purpose of the Study:

• To develop and evaluate a deep learning framework, Swin-UNet conditional generative adversarial network (cGAN), for automatic femoral and tibial cartilage segmentation in MRI.
• To compare its performance against conventional UNet, UNet cGAN, and Swin-UNet baseline models.

Main Methods:

• Utilized a dataset of 232 knee MRI scans for quantitative experiments.
• Compared Swin-UNet cGAN against UNet, UNet cGAN, and Swin-UNet using Dice similarity coefficient, mean intersection over union, and surface distance metrics.
• Evaluated the model's performance on an external validation dataset.

Main Results:

• Swin-UNet cGAN achieved superior Dice similarity coefficient and intersection over union scores for both femoral and tibial cartilage segmentation.
• The model demonstrated significantly improved performance in distance metrics for tibial cartilage and comparable results for femoral cartilage.
• Consistently high segmentation accuracy was observed on both internal test and external validation datasets.

Conclusions:

• The Swin-UNet cGAN offers more accurate knee cartilage segmentation, especially in boundary definition, compared to baseline models.
• The model exhibits promising generalizability across different cohorts, addressing limitations of CT-based systems by enabling cartilage visualization.
• This MRI-based deep learning approach has the potential to enhance surgical precision and patient outcomes in total knee arthroplasty.