Cross-Modality Image Adaptation Based on Volumetric Intensity Gaussian Process Models (VIGPM).

Area of Science:

• Medical Imaging
• Computational Anatomy
• Machine Learning

Background:

• Multi-modal imaging is crucial for diagnosis but often limited by cost, invasiveness, and scanner availability.
• Three-dimensional (3D) morphable models are used for feature-based analysis in medical imaging.
• Existing methods struggle to generate complementary image modalities when secondary scans are unavailable.

Purpose of the Study:

• To develop a method for synthesizing 3D volumetric medical images in one modality from another.
• To enable cross-modality image translation using Gaussian Processes (GPs) and latent space representations.
• To provide registered complementary imaging information from a single source modality.

Main Methods:

• A Gaussian Process (GP) was constructed over transformations to establish anatomical correspondence within training image modalities.
• The GP's eigenspace (latent space) was utilized to create a parametric representation for each modality.
• An operator for cross-domain translation between different modalities was developed, forming the VIGPM framework.

Main Results:

• The latent space effectively generates samples representative of the encoded modality.
• 3D volumetric images can be efficiently encoded in the latent space and translated to synthesize corresponding images in other modalities.
• The VIGPM framework allows for cross-modality synthesis by learning observations in a given modality.

Conclusions:

• The VIGPM framework successfully synthesizes medical images across modalities, addressing limitations of multi-modal data acquisition.
• This method provides access to registered complementary imaging information from a single acquired modality.
• The approach enhances diagnostic capabilities by enabling the generation of otherwise unavailable image modalities.