A New Feature Extraction Method for Stroke Classification with Microwave Signals Based on Early Stroke Volume Change

Abstract

This study proposes a new feature extraction method for microwave-based brain stroke classification using differential scattering parameters. In addition to the differences in dielectric properties, a further clinical distinction between hemorrhagic and ischemic strokes after onset is that hemorrhagic stroke volume tends to evolve during the first few hours, whereas ischemic stroke volume remains relatively stable. This early-stage behavioral disparity leads to measurable variations in scattering parameters across successive measurements. Therefore, according to the proposed method, scattering parameters are acquired from the patient’s head at two distinct times, and the subtraction of these measurements is used as a feature for classification. A synthetic database of training and test data is generated using three-dimensional magnetic resonance imaging (MRI)-based numerical head phantoms enclosed by a 21-element modified printed monopole antenna array. The physical dimensions, dielectric properties, stroke sizes, and lesion locations are systematically varied to represent diverse clinical scenarios, and realistic noise is incorporated into the simulated data. A standard support vector machine (SVM) classifier is employed to evaluate the proposed feature extraction method on the generated dataset, yielding average accuracies of 97.11% and 78.55% at noise levels of −115 dB and −95 dB, respectively; these results indicate improvements of 34.15 and 16.68 percentage points compared to the use of raw scattering parameters.