Fully-Constrained Variable Projection for Water-Fat Models

Area of Science:

• Medical Imaging
• Biophysics
• Computational Science

Background:

• Variable Projection (VARPRO) is a dimensionality reduction technique for least squares optimization.
• Existing constrained VARPRO variants for water-fat models prevent overfitting but allow potentially vulnerable anti-parallel orientations.
• These limitations can affect the accuracy of parameter estimation in water-fat MRI.

Purpose of the Study:

• To derive and present a novel, fully-constrained VARPRO formulation for water-fat modeling.
• To address the limitations of existing constrained VARPRO methods by preventing anti-parallel orientations.
• To maintain the dimensionality reduction and numerical efficiency of the original VARPRO technique.

Main Methods:

• Redefined linear coefficients based on physical constraints, incorporating fat fraction (f) as a nonlinear parameter.
• Demonstrated that the maximum likelihood (ML) estimate of f can be expressed as a closed-form function of off-resonance frequency (ω) and relaxation rate (R2*).
• This reformulation allows the removal of the fat fraction parameter, restoring the original VARPRO dimensionality.

Main Results:

• The proposed fully-constrained VARPRO method demonstrated more stable estimation of ω, R2*, and f for challenging datasets.
• This enhanced stability was observed even in complex or noisy data scenarios.
• The method successfully resolved the limitations of previous constrained VARPRO approaches.

Conclusions:

• The fully-constrained VARPRO formulation enables more reliable local parameter estimation in water-fat modeling.
• This advanced method preserves the computational efficiency characteristic of existing VARPRO approaches.
• It offers a more robust solution for quantitative water-fat MRI analysis.