Machine learning approaches for automated stroke detection, segmentation, and classification in microwave brain imaging systems

Abstract

In this paper, an intracranial hemorrhage stroke detection and classification method using microwave imaging system (MIS) based on machine learning approaches is presented. To create a circular array-based MIS, sixteen elements of modified bow-tie antennas around a multilayer head phantom with a spherical target with radius of 1 cm as an intracranial hemorrhage target are simulated in CST simulator. To obtain satisfied radiation characteristics in the desired frequency band of 0.5– 5 GHz a suitable matching medium is designed. Initially, in the processing section, a confocal image-reconstructing method based on delay-multiply-and-sum (DMAS) beam-forming algorithms is used. Then, reconstructed images are generated, which shows the applicability of the confocal method in detecting a spherical target in the range of 1 cm. Separating and categorizing targets is a challenging task due to the ambiguity in the extracted target from MIS. Thus, to distinguish between healthy and unhealthy brain tissues, a new compound machine learning technique, including filtering, edge-detection based segmentation, and applying K-Means and fuzzy clustering techniques, which reveal intracranial hemorrhage area from reconstructed images is adopted. Simulated results are presented to validate the proposed method effectiveness for precisely localizing and classifying bleeding targets.