Gaussianmorph: deformable medical image registration with Gaussian noise constraints

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning excels in medical image registration, offering efficiency and improved outcomes through automatic feature extraction.
• Cascaded networks are increasingly used for coarse-to-fine registration, despite higher computational costs during training and inference.

Purpose of the Study:

• To leverage the high registration performance of cascaded networks for enhanced accuracy.
• To introduce a novel deep learning architecture, GaussianMorph, for improved medical image registration.

Main Methods:

• Cascaded two VoxelMorph convolutional neural networks for dense deformation field generation.
• Introduced a second network to output a noisy deformation field, minimizing error against Gaussian noise to boost performance.
• Integrated an Enhancement Features-encoder (EF-encoder) block with an attention mechanism in the encoder-decoder structure.

Main Results:

• GaussianMorph demonstrated superior performance compared to VoxelMorph (VM), VM × 2, and TST-Net on LPBA40 and HBN datasets.
• Achieved better Dice Similarity Coefficient, Average Symmetric Surface Distance, Structural Similarity, and Pearson Correlation Coefficient.
• Experimental results confirm improved registration accuracy with the proposed GaussianMorph method.

Conclusions:

• The proposed GaussianMorph method effectively enhances medical image registration accuracy.
• The combination of cascaded networks, noise-based deformation field refinement, and attention-based feature enhancement yields significant improvements.
• GaussianMorph represents a promising advancement in deep learning-based medical image registration techniques.