CT Perfusion Imaging of the Brain with Machine Learning

Abstract

Computed tomography (CT) perfusion imaging is a routinely used technique in the field of neurovascular imaging. The progression of a bolus of contrast agent through the neurovasculature is imaged in a series of CT scans. Relevant perfusion parameters, such as cerebral blood volume (CBV), flow (CBF) and delay (Tmax), can be computed by deconvolution of the contrast-time curves with the bolus shape measured at one of the feeding arteries. These parameters are crucial in the medical management of acute stroke patients, where they are used to identify the extent of likely salvageable tissue and irreversibly damaged infarct core. Deconvolution is normally achieved using singular value decomposition (SVD). However, studies have shown that such a technique is noise sensitive and easily influenced by artifacts in the source image, and may introduce further distortions in the output parameters. In this study, we present a machine learning approach to the estimation of perfusion parameters from CT imaging. Standard types of regression-based machine learning models were trained on the raw/native CT perfusion imaging data to reproduce the output of an FDA-approved commercial implementation of the SVD deconvolution algorithm. As part of our experiments, Kernel ridge regression and random forest models performed best (SSIM: $~82\%, ~84\%$), trading off quick run time for better prediction accuracy while requiring a relatively low number of training examples.