MATGAN: a generative adversarial network with multi-scale attention and Huber loss for brain MRI registration

Area of Science:

• Medical Image Analysis
• Neuroimaging
• Artificial Intelligence

Background:

• Accurate brain magnetic resonance image (MRI) registration is crucial for medical image analysis.
• Fine-grained alignment is challenging in small or complex brain regions.
• Generative adversarial networks (GANs) offer potential for improved registration accuracy.

Purpose of the Study:

• To propose MATGAN, a novel GAN-based framework for robust and accurate brain MRI registration.
• To integrate a multi-scale attention mechanism and Huber loss into adversarial learning for enhanced alignment.
• To improve the precision of deformation estimation and structural alignment in brain images.

Main Methods:

• Developed MATGAN, a GAN-based registration framework incorporating a multi-scale attention mechanism in the discriminator.
• Embedded a multi-scale attention module to extract multi-resolution semantic representations.
• Incorporated Huber loss into the similarity constraint to mitigate outlier influence and enhance training stability.

Main Results:

• MATGAN demonstrated enhanced ability to distinguish real from generated deformations via multi-scale attention.
• The framework achieved improved anatomical consistency in generated deformation fields.
• Experimental results on three public brain MRI datasets confirmed MATGAN's effective registration performance.

Conclusions:

• MATGAN provides a robust and accurate solution for brain MRI registration.
• The integration of multi-scale attention and Huber loss significantly improves alignment precision.
• MATGAN advances the field of medical image analysis by enabling better structural alignment in complex brain regions.