A memory based model for cartilage and meniscus segmentation in 3D knee MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Accurate knee osteoarthritis monitoring requires precise cartilage morphometrics from MRI.
• Current cartilage segmentation methods are challenging, labor-intensive, and prone to variability.
• Visual Foundational Models (VFMs) offer potential for improved segmentation generalizability.

Purpose of the Study:

• Introduce SAMRI-2, a novel transformer-based deep learning method for 3D knee MRI segmentation.
• Enhance cartilage and meniscus segmentation using interactive, memory-based VFMs.
• Improve annotation efficiency and segmentation accuracy in musculoskeletal imaging.

Main Methods:

• Developed SAMRI-2, a transformer-based deep learning model for 3D knee MRI segmentation.
• Incorporated a Hybrid Shuffling Strategy (HSS) for improved spatial awareness and convergence.
• Utilized segmentation mask propagation to enhance annotation efficiency.
• Compared SAMRI-2 against 3D-VNet, 3D nnU-Net, SAMRI2D, and SAMRI3D on diverse datasets.

Main Results:

• SAMRI-2 achieved superior segmentation performance, outperforming all other models.
• Demonstrated an average Dice Similarity Coefficient (DSC) improvement of 0.05, with a maximum gain of 0.12 for tibial cartilage.
• Reduced cartilage thickness errors by up to threefold compared to other methods.
• Maintained high accuracy with minimal user interaction (as few as three clicks per volume).

Conclusions:

• SAMRI-2 offers a reliable and efficient AI-assisted approach for knee MRI segmentation.
• The memory-based VFM with spatial awareness advances deep learning in musculoskeletal imaging.
• This method holds promise for improving osteoarthritis diagnosis and management.