Deep residual-SVD network for brain image registration.

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Medical image registration aligns spatial positions of two images.
• Current U-Net architectures for registration have limited learning capacity and are sensitive to image noise.
• Addressing these limitations is crucial for accurate medical image analysis.

Purpose of the Study:

• To develop a novel unsupervised 3D brain image registration framework.
• To enhance the learning ability of registration networks and mitigate noise effects.
• To improve the accuracy of medical image registration.

Main Methods:

• Proposed a U-Net architecture incorporating residual units and a singular value decomposition (SVD) denoising layer.
• Utilized Akaike information criterion for model order estimation in SVD denoising.
• Employed Exponential Linear Unit (ELU) as an activation function for enhanced noise robustness.

Main Results:

• The Deep Residual-SVD Network demonstrated superior performance compared to state-of-the-art methods on the Mindboggle101 and LPBA40 brain MRI datasets, particularly in Dice Score metrics.
• The method achieved comparable results in terms of mean folding voxels and registration time.
• Experimental results validated the effectiveness of residual units, SVD denoising, and ELU activation in improving registration accuracy and noise handling.

Conclusions:

• The Deep Residual-SVD Network significantly improves 3D brain image registration accuracy.
• Residual units enhance network learning capacity, while SVD denoising effectively removes noise.
• The proposed framework offers a robust solution for accurate and reliable medical image registration.