A deep learning approach for robust and accurate deconvolution of DSC MRI perfusion calculation

Abstract

The conventional deconvolution method in DSC perfusion MRI suffers from sensitivity to noise and threshold level. Regularization methods to mitigate the noise issue also suffers from other issues. In this study, we present a deep learning approach to perform deconvolution more robustly and accurately. Our result showed multi layers perceptron (MLP) performed deconvolution more accurately in synthetic data compared to the traditional regularization method. We also showed that MLP performed more robustly in patient data with varying levels of noise. This study provides a strong argument for using MLP as a stable and accurate deconvolution method for DSC perfusion calculation.