Acute Infarct Volume Prediction Based on CT Perfusion Metrics Derived from an Automated Software Package using Machine Learning Models

The correct characterization of acute ischemic stroke is critical in the clinical setting, requiring a fast and accurate estimation of the main biomarkers that characterize it. In this context diffusion weighted imaging (DWI) MRI has proven to be the gold standard to measure the infarct core in the brain but in the clinical setting it is common practice to use computed tomography perfusion (CTP) protocols, which are faster and more available in stroke centers and hospitals.Here we analyze the CTP parameters generated by the automated software package Minerva using the Ischemic Stroke Lesion Segmentation (ISLES) 2018 challenge dataset which provides the raw CTP time series, derived perfusion maps and the manual segmentation from DWI MRI images as gold standard for the infarct core. We compare the automated outputs estimated by Minerva against these segmentations and perform models in predicting infarct volume based on linear regression and multilayer perceptron neural networks.The mean absolute error obtained when comparing the infarct automatically estimated by Minerva was 19.207 ml with a correlation coefficient of 0.839. The error was improved after optimizing and training two different models employing multiple perfusion metrics with and error of 16.453 ml for a linear regression fitting and 15.341 ml with a multilayer perceptron after training and validating both models with 5 fold cross-validation.