Medical image registration via neural fields.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Image registration is crucial for medical image analysis, but traditional optimization methods are slow, and learning-based methods struggle with domain shift.
• Existing methods face challenges in balancing speed, accuracy, and robustness in deformable image registration.

Purpose of the Study:

• To introduce Neural Image Registration (NIR), a novel framework combining optimization with deep neural networks for accurate and efficient medical image registration.
• To model complex deformations using neural fields and optimize registration via gradient descent.

Main Methods:

• NIR utilizes neural fields to represent continuous transformations, outputting displacement or velocity vector fields for registration.
• The framework employs techniques like coordinate encoding, sinusoidal activation, and specialized sampling for improved optimization.
• Registration is achieved by updating neural field parameters through stochastic mini-batch gradient descent.

Main Results:

• NIR demonstrates highly competitive performance on 3D MR brain scan datasets, achieving superior accuracy and regularity compared to traditional methods.
• The framework significantly reduces computation time while improving registration outcomes.
• NIR shows promising performance in cross-dataset registration tasks, outperforming pre-trained learning-based methods.

Conclusions:

• NIR offers a powerful and efficient approach to medical image registration, overcoming limitations of existing methods.
• The integration of neural networks with optimization provides a flexible and accurate framework for complex registration tasks.
• NIR represents a significant advancement in automated medical image analysis and computational anatomy.