Improved deep learning-based IVIM parameter estimation via the use of more "realistic" simulated brain data

Area of Science:

• Medical Imaging
• Biophysics
• Machine Learning

Background:

• Precise estimation of intravoxel incoherent motion (IVIM) parameters is challenging due to low signal-to-noise ratio (SNR) and limited b-values.
• Brain imaging is particularly affected by subtle differences in diffusion (D) and pseudo-diffusion (D*) parameters, leading to inaccurate and noisy results.

Purpose of the Study:

• To develop a synthetic data-driven supervised learning method (SDD-IVIM) for enhanced precision and noise robustness in IVIM parameter estimation.
• To achieve this without requiring real-world data for neural network training.

Main Methods:

• A novel model-based approach generated synthetic human brain IVIM data by sampling parameters from complex distributions and modulating with brain texture.
• Synthetic multi-b-value diffusion-weighted images were created using the IVIM bi-exponential model.
• A U-Net model was trained using this synthetic data for IVIM parameter mapping.

Main Results:

• The SDD-IVIM method demonstrated superior performance in numerical phantom experiments, achieving lower error and higher structural similarity, especially at low SNR.
• In glioma patient studies, SDD-IVIM yielded lower coefficient of variation and improved contrast-to-noise ratios between tumor and healthy tissue.

Conclusions:

• The proposed SDD-IVIM method significantly enhances parametric map quality and parameter estimation precision.
• The technique shows strong noise resistance and improved lesion characterization capabilities in IVIM analysis.