Unsupervised Image Registration towards Enhancing Performance and Explainability in Cardiac and Brain Image Analysis

Area of Science:

• Medical image analysis
• Deep learning in radiology
• Computational anatomy

Background:

• Magnetic Resonance Imaging (MRI) uses multiple sequences (modalities) providing diverse clinical information.
• Disparities across MRI modalities necessitate robust image registration for accurate biomarker analysis.
• Existing unsupervised models often fail to address both affine and non-rigid transformations simultaneously, and lack inverse-consistency.

Purpose of the Study:

• To introduce an unsupervised deep learning methodology for simultaneous affine and non-rigid medical image registration.
• To address the critical property of inverse-consistency in inter-modality registration.
• To develop a versatile and explainable registration model for clinical applications.

Main Methods:

• Proposed the "FIRE" (unsupervised deep learning registration) model capable of handling both affine and non-rigid transformations.
• Implemented bi-directional cross-modality image synthesis for modality-invariant latent representations.
• Incorporated factorised transformation networks and an inverse-consistency loss for topology-preserving transformations.

Main Results:

• FIRE demonstrated superior performance compared to the ANTs-based Symmetric Normalization baseline.
• The model achieved improved registration accuracy on multi-modality brain 2D/3D MRI and intra-modality cardiac 4D MRI datasets.
• FIRE operates in a memory-saving mode, learning topology-preserving registration during training.

Conclusions:

• The FIRE model offers an efficient and versatile solution for multi-modal medical image registration.
• The methodology enhances model explainability by focusing on model-data components.
• FIRE shows significant potential for clinical applications requiring accurate and robust image registration.