SymReg-GAN: Symmetric Image Registration With Generative Adversarial Networks

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Symmetric image registration ensures inverse-consistent transformations, reducing bias in analyses like computational anatomy and shape analysis.
• Existing symmetric registration methods, particularly for multimodal images, suffer from slow speeds, difficulty with inter-modality relationships, and high labeling costs.

Purpose of the Study:

• To introduce SymReg-GAN, a novel generative adversarial network (GAN) approach for efficient and accurate symmetric image registration.
• To overcome the speed, inter-modality relation exploration, and data labeling limitations of current symmetric registration techniques.

Main Methods:

• Formulated symmetric registration as a conditional generative adversarial network (GAN) for unimodal and multimodal images.
• Employed a semi-supervised training strategy to leverage both labeled and unlabeled data.
• Introduced a cycle-consistency loss to enforce registration symmetry.

Main Results:

• SymReg-GAN demonstrated superior performance compared to state-of-the-art methods.
• Experiments were conducted on six public brain MRI datasets and one combined CT/MRI dataset.

Conclusions:

• SymReg-GAN effectively addresses the limitations of existing symmetric image registration methods.
• The proposed GAN-based approach offers a promising solution for precise image analysis in various interdisciplinary applications.