Application of Deep Learning Algorithms for Automated MRI Knee Joint Image Segmentation

Manual segmentation of knee joint structures in 3D Magnetic Resonance Imaging (MRI) scans is a labor-intensive process, often requiring several hours per scan for detailed segmentations. This paper presents a deep learning approach utilizing a modified 3D U-Net architecture for automated segmentation of knee MRI images from the Osteoarthritis Initiative (OAI) dataset. Trained on 507 subjects, the model achieves minimum Dice similarity coefficients of 98.5% for femoral and tibial bone, 85.0% for femoral cartilage, and 70.8% for tibial cartilage. The average inference time per one subject was under 30 seconds, demonstrating a reduction in processing time compared to manual annotation. These results suggest that CNN-based segmentation offers a reliable and efficient alternative to manual segmentation, enhancing diagnostic workflows in osteoarthritis monitoring.