Machine learning techniques for microwave brain stroke detection and classification

The innovative microwave brain imaging system (MWI) allows to carry out a pre-diagnosis of the stroke in an ambulance and a continuous monitoring of bedridden patients, thanks to a non-invasive, easy-to-use, portable and low-cost device. The operating principle of the system exploits the dielectric contrast between healthy and pathological tissues at microwaves frequencies. The combination of artificial intelligence techniques and the proposed imaging method can effectively assist the clinician in making decisions about the therapeutic treatment of potential stroke patients. In this regard, my thesis project consists in the development of algorithms capable of solving classification problems. The aim of the work is therefore to identify the presence and the location within the head of the cerebral stroke, distinguishing the cases of ischemia from those of hemorrhage. Classes are detected via supervised Machine Learning Algorithms (ML) as Support Vector Machine (SVM), Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (k-NN). The S parameters measured at the antennas ports of the MWI system represent the features that are given as input to the ML algorithms. They are provided in the form of amplitude and of real and imaginary part. Data collection and processing are two key aspects in the learning process: algorithms need thousands of known examples to identify patterns useful to build a model that is then able to correctly recognize the class of an unknown case. However, carrying out a sufficiently large number of measurements requires a great effort in terms of time. For this reason, the first step was to create a series of synthetic training data, using the Born approximation and performing a linearization of the scattering operator. This method allowed to generate 10000 examples in a very short time. The relative permittivity and conductivity values adopted for the creation of the synthetic training set refer to the dielectric characteristics of the brain tissues at the considered frequencies. Ad hoc mixtures that mimic the dielectric characteristics of both ischemic and hemorrhagic stroke and healthy brain tissue, intended as a homogeneous medium, were created. At this point the tuning of the hyper-parameters, the model construction and the training of ML algorithms were performed. The second part of the work involved the creation of a testing-set used to evaluate the performance of the previously trained algorithms. This dataset consists of examples much more similar to reality, obtained through full-wave Finite Element Method (FEM) simulations. It came out that all the classifiers can identify the presence or not of the stroke and among the algorithms used, the MLP proved to be the most performing. From the results achieved it is evident that the linearization of the scattering operator is a reasonable approximation. Future developments will consist in testing ML algorithms on a series of experimental measurements performed with the MWI system and the 3D human head phantom.