Improved myelin water fraction mapping with deep neural networks using synthetically generated 3D data

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Biophysics

Background:

• Myelin water fraction (MWF) mapping is crucial for neuroimaging but faces challenges with acquisition time, signal-to-noise ratio (SNR), and repeatability.
• Current MWF estimation techniques require significant improvements for widespread clinical adoption.

Purpose of the Study:

• To introduce a novel generative model for synthesizing large-scale 3D datasets for quantitative MWF parameter mapping.
• To enable the training of deep learning models, specifically convolutional neural networks (CNNs), for accurate MWF estimation.

Main Methods:

• The model integrates a magnetic resonance (MR) physics signal decay model with a probabilistic multi-component parametric T2 model.
• Generative spatial models, conditioned on tissue segmentations, capture spatial variations in synthesized data.
• The system generates high-quality synthetic signals and parameters across a range of realistic values.

Main Results:

• The synthetically trained CNN demonstrated superior accuracy compared to existing methods, particularly under low SNR conditions.
• Achieved a normalized root mean squared error (nRMSE) of less than 7% on synthetic data.
• The proposed method showed at least a 4x improvement in the coefficient of variation (CoV) on a test-retest dataset.

Conclusions:

• The generative model effectively synthesizes realistic 3D datasets for MWF mapping.
• CNNs trained on these synthetic datasets significantly improve MWF estimation accuracy and repeatability.
• This approach holds promise for overcoming current limitations and facilitating clinical use of MWF mapping.