Recurrent inference machine for medical image registration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Medical image registration aligns voxels across multiple images for analysis.
• Deep learning methods offer speed but can sacrifice accuracy and require large datasets.
• Optimization-based methods are training-free but may be slower.

Purpose of the Study:

• To develop a novel, data-efficient deep learning method for medical image registration.
• To improve both registration accuracy and training data efficiency.
• To address limitations of existing deep learning and optimization-based registration techniques.

Main Methods:

• Proposed the Recurrent Inference Image Registration (RIIR) network, a meta-learning solver.
• Formulated registration as an iterative process learning optimization update rules.
• Integrated implicit regularization with explicit gradient input.

Main Results:

• RIIR demonstrated superior registration accuracy and high data efficiency across brain MRI, lung CT, and cardiac MRI datasets.
• Achieved state-of-the-art performance using only 5% of training data compared to other deep learning methods.
• Ablation studies confirmed the significant contribution of hidden states in the recurrent inference framework.

Conclusions:

• The RIIR network offers a highly data-efficient framework for deep learning-based medical image registration.
• This approach effectively balances accuracy and data efficiency, crucial for clinical applications.
• RIIR represents a significant advancement in medical image analysis and AI-driven diagnostics.