Multi-model sequential analysis of MRI data for microstructure prediction in heterogeneous tissue

Area of Science:

• Biomedical Imaging
• Quantitative MRI
• Tissue Microstructure Analysis

Background:

• Analyzing heterogeneous tissues with diffusion-relaxation MRI often requires selecting a single model, which can lead to inaccuracies.
• The optimal model for tissue microstructure prediction can vary significantly within different tissue environments.

Purpose of the Study:

• To propose a general, multi-model approach for combining predictions of tissue microstructure from diffusion-relaxation MRI data.
• To overcome limitations of single-model approaches in heterogeneous tissues.
• To enable flexible biological interpretation by separating signal characterization from biological assignment.

Main Methods:

• A three-stage sequence for signal interpretation: 1. Applying multiple semi-phenomenological models to predict physical properties of water pools. 2. Using tissue microstructure models to predict relative volumes of tissue structures within each water pool. 3. Aggregating microstructure predictions weighted by model likelihood and water pool fractional volumes.
• Exemplar application using in vivo diffusion-relaxation MRI data from human prostate.
• Demonstration of combining multiple models to predict tissue microstructure.

Main Results:

• The multi-model approach reduces prediction variance in regions where complex models are overparameterized and bias where models are underparameterized.
• Separation of signal characterization and biological assignment allows for diverse interpretations of physical properties.
• Successful application demonstrated on human prostate diffusion-relaxation MRI data.

Conclusions:

• The proposed general method effectively combines multiple models for robust tissue microstructure prediction in heterogeneous tissues.
• This flexible framework enhances the analysis of diffusion-relaxation MRI data, offering broader applicability.
• The method has potential applications beyond prostate imaging, applicable to various domains where single models are insufficient.