RESUMO
Stroke is one of the leading causes of death and disability. To address this challenge, microwave imaging has been proposed as a portable medical imaging modality. However, accurate stroke classification using microwave signals is still an open challenge. In addition, identified features of microwave signals used for stroke classification need to be linked back to the original data. This work attempts to address these issues by proposing a wavelet convolutional neural network (CNN), which combines multiresolution analysis and CNN to learn distinctive patterns in the scalogram for accurate classification. A game theoretic approach is used to explain the model and indicate distinctive features for discriminating stroke types. The proposed algorithm is tested in simulation and experiments. Different types of noise and manufacturing tolerances are modeled using data collected from healthy human trials and added to the simulation data to bridge the gap between the simulation and real-life data. The achieved classification accuracy using the proposed method ranges from 81.7% for 3D simulations to 95.7% for lab experiments using simple head phantoms. Obtained explanations using the method indicate the relevance of wavelet coefficients on frequencies 0.95-1.45 GHz and the time slot of 1.3 to 1.7 ns for distinguishing ischemic from hemorrhagic strokes.