Tensor-product kernel-based representation encoding joint MRI view similarity.

Area of Science:

• Medical Imaging Analysis
• Neuroimaging
• Machine Learning for Healthcare

Background:

• 3D magnetic resonance image (MRI) analysis requires understanding inter-slice relationships across axial, coronal, and sagittal views.
• Inferring mutual relationships from marginal image similarity (MIS) matrices is challenging due to nonlinearities between axes.

Purpose of the Study:

• To develop a novel framework, Tensor-Product Kernel-based Representation (TKR), for joint image similarity analysis of 3D MRI data.
• To encode patient-specific brain structure patterns by integrating MIS matrices from different views.

Main Methods:

• Estimated marginal image similarity (MIS) matrices for axial, coronal, and sagittal views.
• Developed a Tensor-Product Kernel-based Representation (TKR) to unify MIS matrices into a single framework.
• Trained the TKR model in a low-dimensional projected space to mitigate voxel-derived noise.

Main Results:

• The proposed TKR approach demonstrated superior performance in classifying patient categories (gender and age) compared to conventional voxel-wise sum of squared differences.
• Successfully encoded complex brain structure patterns, accounting for patient differences.

Conclusions:

• The TKR approach offers a robust method for MRI similarity inference and clustering.
• This technique can significantly support template-based image segmentation and atlas construction in neuroimaging.