Simulation-based Machine Learning Training for Brain Anomalies Localization at Microwaves

Machine learning enters the world of medical application and, in this paper, it joins microwave imaging technique for brain stroke classification. One of the main challenges in this application is the need of a large amount of data for the machine learning algorithm training that can be performed via measurements or simulations. In this work, we propose to make the algorithm training via simulations based on a linear integral operator that reduces by three orders of magnitude the data generation time with respect to standard full-wave simulations. This method is used here to train the multilayer perceptron algorithm. The data-set is organized in nine classes, related to the presence, the type and the position of the stroke within the brain. We verified that the algorithm metrics (accuracy, recall and precision) reach values close to 1 for each class.