GradICON: Approximate Diffeomorphisms via Gradient Inverse Consistency.

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Medical image registration is crucial for analyzing and comparing medical scans.
• Existing methods often struggle with transformation regularity and convergence.
• Learning-based approaches require careful regularization to ensure accurate spatial transformations.

Purpose of the Study:

• To develop a novel regularizer for learning spatial transformations in medical image registration.
• To improve the convergence and performance of registration models.
• To achieve state-of-the-art registration without dataset-specific tuning.

Main Methods:

• A neural network predicts forward and backward transformation maps between image pairs.
• The composition of these maps is regularized by penalizing deviations of its Jacobian from the identity matrix.
• The proposed regularizer is named GradICON.

Main Results:

• GradICON significantly improves convergence during registration model training.
• It outperforms existing methods in promoting transformation regularity.
• State-of-the-art registration performance is achieved on various real-world medical image datasets.

Conclusions:

• GradICON offers an effective approach to learning regular spatial transformations for medical image registration.
• The method demonstrates robustness and generalizability across different datasets.
• This technique enhances the reliability and accuracy of medical image analysis.