Synthetic X-Q space learning for diffusion MRI parameter estimation: a pilot study in breast DKI.

Area of Science:

• Medical Imaging
• Machine Learning
• Biophysics

Background:

• Machine learning, particularly synthetic Q-space learning (synQSL), shows promise for diffusion MRI (dMRI) parameter estimation using synthetic data.
• Existing methods face challenges with robustness and noise in parameter estimation.

Purpose of the Study:

• To develop and investigate a novel machine learning method, synthetic X-Q space learning (synXQSL), for improved dMRI parameter estimation.
• To enhance the robustness of dMRI parameter estimation in the presence of noise.

Main Methods:

• Synthesized local parameter patterns (3x3 voxels) using linear combinations of bases (flat, linear, quadratic).
• Computed diffusion-weighted image signal values using the signal model equation for diffusional kurtosis imaging with Rician noise simulation.
• Employed a multi-layer perceptron for parameter estimation, trained across various noise levels defined by a noise ratio.

Main Results:

• synXQSL demonstrated superior performance over synQSL in estimating parameters from noisy synthetic datasets.
• Digital phantom experiments suggested a quadratic pattern as a potentially optimal choice for synXQSL bases.
• Clinical breast datasets showed that synXQSL effectively suppresses noise in parameter maps, leading to improved contrast.

Conclusions:

• The study characterized the basic properties of synXQSL across diverse datasets.
• synXQSL, with appropriate base selection during training data synthesis, holds potential for enhancing dMRI parameter accuracy in noisy conditions.