Self-labelled encoder-decoder (SLED) for multi-echo gradient echo-based myelin water imaging.

Area of Science:

• Medical Imaging
• Machine Learning
• Neuroscience

Background:

• Reconstructing high-quality myelin water imaging (MWI) maps from multi-echo gradient echo (mGRE) sequences is challenging.
• Traditional non-linear least squares fitting (NLLS) methods for MWI can yield maps with limited detail and suboptimal signal-to-noise ratio (SNR).

Purpose of the Study:

• To develop a novel, unsupervised learning method, self-labelled encoder-decoder (SLED), to enhance gradient echo-based MWI data fitting.
• To improve the accuracy, stability, and noise performance of myelin water fraction (MWF) estimation.

Main Methods:

• Collected ultra-high resolution MWI data from five mouse brains using a 7T preclinical MRI system and an mGRE sequence.
• Implemented a self-labelled encoder-decoder (SLED) network in TensorFlow for MWF calculation based on mGRE signal decay.
• Utilized a simulated MWI phantom to rigorously evaluate MWF estimation performance.

Main Results:

• SLED demonstrated superior MWF estimation accuracy and stability compared to NLLS in phantom tests.
• SLED produced less noisy MWF maps from high-resolution mouse brain MR microscopy images, with lower noise amplification across genotypes.
• Mean MWF values in white matter regions of interest (ROIs) derived from SLED strongly correlated with NLLS results, and SLED showed higher tolerance to low SNR data.

Conclusions:

• The unsupervised and self-labeling SLED method provides a robust alternative for analyzing gradient echo-based MWI data.
• SLED enables accurate and stable myelin water fraction (MWF) estimations, outperforming conventional techniques.