Groupwise Non-Rigid Registration with Deep Learning: An Affordable Solution Applied to 2D Cardiac Cine MRI...

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Anatomy

Background:

• Groupwise (GW) image registration is crucial for medical image analysis but is computationally intensive.
• Existing methods require significant processing power and time due to repeated calculations of transformations and gradients.

Purpose of the Study:

• To develop a computationally efficient deep learning (DL) architecture for groupwise elastic registration of 2D dynamic cardiac MRI sequences.
• To reduce the runtime of GW registration while maintaining or improving accuracy compared to existing methods.

Main Methods:

• A simplified U-net based deep learning architecture, termed dGW, was proposed for 2D dynamic sequence registration.
• The template image was iteratively obtained alongside registered images within the dGW framework.
• Network hyperparameters, controlling transformation smoothness, were optimized using a forward selection procedure.
• The model was trained and validated on two cardiac MRI datasets, with a separate dataset used for testing.

Main Results:

• The dGW model achieved a 9-fold reduction in runtime compared to traditional optimization-based implementations.
• Significant differences in registration accuracy, measured by the structural similarity (SSIM) index, were observed compared to an alternative DL solution (Voxelmorph).

Conclusions:

• The proposed dGW deep learning architecture offers a computationally efficient solution for groupwise elastic image registration.
• dGW demonstrates potential for accelerating medical image processing tasks involving dynamic sequences, such as cardiac MRI analysis.