A comparison of machine learning methods for recovering noisy and missing 4D flow MRI data

Area of Science:

• Cardiovascular Imaging and Hemodynamics
• Biomedical Engineering
• Machine Learning in Medical Imaging

Background:

• Accurate blood flow measurement is crucial for understanding cardiovascular diseases.
• Time-resolved 3D phase-contrast MRI (4D flow MRI) offers noninvasive velocity measurements but suffers from noise and artifacts.
• Current denoising methods lack comprehensive comparison with advanced machine learning techniques.

Purpose of the Study:

• To compare traditional filtering with machine learning (SVD) and deep learning (autoencoders) for denoising and data imputation in 4D flow MRI.
• To evaluate the performance of these methods on both simulated (CFD) and experimental (in vitro) cardiovascular flow data.
• To identify robust methods for enhancing corrupt cardiovascular flow data in diseased arteries.

Main Methods:

• Singular Value Decomposition (SVD)-based machine learning algorithms.
• Autoencoder-type deep learning models, including Denoising Autoencoders (DAE) and Noise2Noise (N2N).
• Application and comparison on artificially corrupted CFD data and in vitro 4D flow MRI data.

Main Results:

• SVD methods performed well on idealized data but struggled with in vitro experimental data.
• Autoencoders demonstrated versatility and applicability across both simulated and experimental datasets.
• Denoising autoencoders (DAE, N2N) significantly outperformed traditional filtering for denoising in vitro 4D flow MRI data.
• N2N achieved noise-free velocity fields even without clean training data.

Conclusions:

• Deep learning autoencoders are superior to traditional filtering and SVD for enhancing corrupt 4D flow MRI data.
• The Noise2Noise autoencoder shows particular promise for generating high-quality, noise-free cardiovascular flow data.
• This study provides a comprehensive comparison of classical and modern methods for improving quantitative hemodynamic analysis from 4D flow MRI.