Robust and fast nonlinear optimization of diffusion MRI microstructure models.

Area of Science:

• Neuroimaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) microstructure models offer greater specificity than DTI.
• Existing models use diverse, often incompatible, optimization algorithms and initialization strategies.
• Computational expense limits the application of dMRI models in large studies.

Purpose of the Study:

• To investigate the performance of various optimization algorithms and initialization strategies for popular dMRI microstructure models.
• To determine if a single, efficient optimization approach can be standardized across models.
• To assess the impact of these choices on run time, data fit, accuracy, and precision.

Main Methods:

• Implemented multiple dMRI models (e.g., NODDI, CHARMED), optimization algorithms, and initialization strategies on GPUs.
• Evaluated performance using metrics like run time, fit quality, accuracy, and precision.
• Assessed results on datasets from two population studies with different acquisition protocols.

Main Results:

• GPU implementation enabled whole-brain fits in seconds to minutes.
• The gradient-free Powell conjugate-direction algorithm demonstrated superior run time, fit, accuracy, and precision.
• Multi-step fitting, especially using simpler models to initialize complex ones, significantly enhanced performance.

Conclusions:

• Optimization algorithms and multi-step approaches critically influence dMRI model performance and stability.
• The Powell algorithm and fitting cascades offer a standardized, efficient method for robust dMRI microstructure analysis.
• These findings facilitate the use of dMRI microstructure modeling in large-scale population studies and integrated tractography.