Abstract
Abstract Recently, many machine learning techniques have been presented to detect brain lesions or determine brain lesion types using microwave data. However, there are limited studies analyzing the location of the lesions through machine learning approaches, despite the critical importance of lesion localization for treatment. This paper contributes to the literature on microwave brain lesion localization by comparing several machine learning algorithms rather than proposing a single algorithm. In the proposed approach, lesions of varying sizes and locations are randomly embedded into realistic 2D brain models. The brain models are divided into four regions, each representing the possible location of the lesion. Subsequently, training and test datasets are generated by computing the electromagnetic fields scattered from healthy brains and brains with lesions via the Method of Moments. Each sample in the datasets is labeled as healthy, region 1, region 2, region 3, or region 4 according to the presence and location of the lesion. Finally, we developed models using various classifiers and compared their effectiveness. Among these, the models trained with stochastic gradient descent classifier, ridge classifier, and logistic regression demonstrated notable accuracy (91%, 82%, and 77%, respectively) for both detection and localization of lesions.