DINOMotion: Advanced Robust Tissue Motion Tracking With DINOv2 in 2D-Cine MRI-Guided Radiotherapy.

Area of Science:

• Medical imaging
• Radiotherapy
• Artificial intelligence

Background:

• Accurate 2D-Cine MRI tracking is vital for radiotherapy safety and efficacy.
• Current methods struggle with large misalignments and lack interpretability.
• Deep learning offers potential for improved motion tracking.

Purpose of the Study:

• Introduce DINOMotion, a novel deep learning framework for robust, efficient, and interpretable motion tracking.
• Enhance image registration accuracy and interpretability in 2D-Cine MRI-guided radiotherapy.
• Address limitations of existing motion tracking techniques.

Main Methods:

• Developed DINOMotion using DINOv2 with Low-Rank Adaptation (LoRA) layers.
• Implemented automatic landmark detection for optimal image registration.
• Utilized deep learning for direct computation of image registration, avoiding iterative optimization.

Main Results:

• Achieved high Dice scores: 92.07% (kidney), 90.90% (liver), 95.23% (lung).
• Obtained low Hausdorff distances: 5.47 mm (kidney), 8.31 mm (liver), 6.72 mm (lung).
• Demonstrated real-time processing (approx. 30ms per scan) and superior performance, especially with large misalignments.

Conclusions:

• DINOMotion provides a robust, efficient, and interpretable solution for real-time motion tracking.
• The framework significantly improves accuracy and handling of misalignments in 2D-Cine MRI-guided radiotherapy.
• DINOMotion shows potential to enhance treatment outcomes and safety.