Adaptive deformation decomposition network for unsupervised medical image registration.

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Deep Learning for Medical Imaging

Background:

• Deformable registration is crucial in medical image analysis but often relies on pre-processed data, limiting its clinical applicability.
• Existing methods struggle with direct alignment of unprepared datasets, hindering scalability and integration into clinical workflows.
• The integration of affine and deformable registration for unaligned data remains a significant challenge.

Purpose of the Study:

• To investigate the limitations of current registration methods on unprepared medical datasets.
• To develop a novel registration method capable of handling complex and large deformations directly on unaligned images.
• To improve the applicability and scalability of deformable registration techniques for clinical use.

Main Methods:

• Introduction of the Adaptive Deformation Decomposition Network (ADDNet) for unsupervised non-rigid registration.
• Development of a Deformation Decomposition Module (DDM) to capture multi-scale dependencies and semantic information for large displacement.
• Design of an Adaptive Contextual Fusion (ACF) module to efficiently fuse global and local features based on deformation complexity.

Main Results:

• ADDNet demonstrates superior performance in handling both large and small deformations, especially on datasets without affine pre-alignment.
• Achieved state-of-the-art registration accuracy on LPBA, ABCT, IXI, and Mindboggle datasets, evidenced by best-in-class HD95 and MSE metrics.
• ADDNet generates more realistic deformation fields, confirming its effectiveness and efficiency in complex registration tasks.

Conclusions:

• ADDNet offers a robust solution for direct, unsupervised non-rigid registration of unprepared medical image datasets.
• The network's ability to decompose complex deformations and adaptively fuse features enhances accuracy and efficiency.
• This method significantly advances the potential for applying advanced registration techniques in real-world clinical settings.