Unrolled Microwave Imaging via Physics-assisted Deep Learning

Microwave imaging (MWI) is a useful tool to identify the characteristics of unknown objects in a non-invasive fashion. It offers advantages such as low cost and portability compared to other medical imaging methods. However, MWI faces challenges in solving the underlying inverse scattering problem (ISP) due to non-linearity and ill-posedness. Traditional optimization methods have limitations related to computational burden and local minima, while learning-based approaches can provide real-time imaging but suffer from a lack of physical understanding and poor generalizability. This paper introduces an innovative unrolled optimization approach, named CNNs-MWI, that combines physics-based calculations and convolutional neural networks to improve generalizability with respect to classical end-to-end learning methods and computational load with respect to standard optimization methods. The proposed method is assessed using simulated and experimental data, demonstrating its effectiveness in addressing MWI challenges.