Deep Learning Enhanced Microwave Imaging for Brain Diagnostics

Permittivity reconstruction of head tissues has an essential role in the application of microwave imaging for brain stroke diagnostics. In this paper, we propose a deep learning enhanced microwave imaging approach for estimating permittivities of tissues inside the head assuming that only the outer boundary of the head is known. The approach first retrieves the inner domain boundaries and then determines the permittivity of each domain. The first task is performed by a U-Net neural network trained to predict the inner boundaries based on the qualitative images obtained using the first order solution computed via truncated singular value decomposition. Then, the permittivities of the domains inside the head are iteratively estimated using the distorted Born iterative method. An assessment of the approach with a simplified but realistic head model consisting of two homogeneous tissues is provided.