Combining Sparse and Dense Features to Improve Multi-Modal Registration for Brain DTI Images

Area of Science:

• Medical Imaging
• Neuroimaging
• Image Processing

Background:

• Multimodal medical image registration is crucial for analyzing subject-specific anatomical and functional data.
• Diffusion MRI, particularly with higher b-values, is highly susceptible to motion artifacts.
• Existing registration methods often struggle with inter-subject variability and rely on specific anatomical features.

Purpose of the Study:

• To propose a novel multimodal intra-subject image registration solution using histogram of oriented gradients (HOG) and mutual information (MI).
• To enhance the accuracy and robustness of aligning T2-weighted (T2w) and diffusion tensor imaging (DTI) MRI scans.
• To develop a registration method that addresses whole-brain variability and motion compensation.

Main Methods:

• A rigid, multimodal image registration algorithm employing linear transformation and oriented gradients.
• Utilizing mutual information (MI) of image histogram-oriented gradients (HOG) as a matching criterion.
• Comparing results against MI-based intensity registration using fiducial registration error (FRE) for transformation assessment.

Main Results:

• The proposed HOG-based MI registration method demonstrated improved accuracy and robustness for whole-brain alignment.
• Computational cost was comparable to traditional MI-based intensity methods, despite the additional HOG computation.
• The method effectively compensates for motion artifacts in diffusion MRI acquisitions.

Conclusions:

• The HOG-based MI approach offers a superior solution for multimodal intra-subject medical image registration compared to feature- or region-based methods.
• This technique provides a robust and accurate method for aligning T2w and DTI MRI data.
• The developed algorithm is efficient and suitable for clinical applications requiring precise image alignment.