A Machine Learning Approach to Perfusion Imaging With Dynamic Susceptibility Contrast MR

Area of Science:

• Medical Imaging
• Neuroscience
• Machine Learning

Background:

• Dynamic susceptibility contrast (DSC) MR perfusion imaging is crucial for neurovascular assessment.
• It calculates blood flow and Tmax by deconvolving contrast agent curves, aiding in acute stroke diagnosis.
• Traditional deconvolution methods like singular value decomposition (SVD) are sensitive to image noise and artifacts.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) approach for estimating DSC-MRI perfusion parameters.
• To assess the impact of training data size, patch size, and ML model choice on parameter estimation accuracy.
• To compare the ML approach's robustness against noise and artifacts with a commercial SVD-based method.

Main Methods:

• Trained various ML models using raw DSC-MRI source data to replicate SVD deconvolution outputs.
• Investigated the influence of training set size (up to 5,000 voxels) and patch size (e.g., 5x5 voxels).
• Compared performance of different ML models, including Random Forest and neural networks, using root mean squared error and phantom studies.

Main Results:

• ML model performance improved with training set size, plateauing after 5,000 samples.
• A 5x5 voxel patch size yielded optimal results, outperforming single-voxel analysis.
• Neural networks demonstrated greater robustness to noise than Random Forest models, despite higher initial error.

Conclusions:

• The ML-based approach provides DSC-MRI perfusion parameter estimates that are invariant to common MR acquisition noise and artifacts.
• This method enhances robustness, outperforming FDA-approved SVD software in evaluations with acute stroke patients and phantom data.