Imposing implicit feasibility constraints on deformable image registration using a statistical generative model.

Area of Science:

• Medical image analysis
• Computational anatomy
• Machine learning for medical imaging

Background:

• Traditional deformable registration relies on regularization, often requiring case-specific tuning.
• A single regularization weight may not capture complex, spatially varying tissue properties.
• This limitation can hinder optimal registration performance.

Purpose of the Study:

• To incorporate a spatially variant deformation prior into image registration using a statistical generative model.
• To develop a method that eliminates the need for explicit regularization and manual tuning.
• To improve the accuracy and efficiency of deformable image registration.

Main Methods:

• A generator network trained in an unsupervised manner to maximize the likelihood of image pairs.
• Utilizing an alternating back-propagation approach for training.
• Performing optimization over a learned low-dimensional deformation parametrization.

Main Results:

• Significantly lower registration errors compared to SimpleElastix and DIRNet on synthetic and CT images.
• Demonstrated physical and physiological feasibility of deformations in cardiac MRI.
• Achieved superior Dice scores and comparable mean average surface distance on left ventricle contours.
• Reduced 3D registration time compared to SimpleElastix.

Conclusions:

• The learned implicit parametrization offers an effective alternative to traditional regularized models like B-splines.
• The proposed method is more flexible in handling spatial heterogeneity in deformations.
• This approach enhances accuracy, efficiency, and physical plausibility in deformable image registration.