Machine Learning Architectures for Microwave Imaging with Enhanced Reconstruction

In the present work, the inverse scattering problem arising in medical imaging is considered, with a particular focus on the accuracy enhancement in the reconstruction of brain and breast models. In particular, the quadratic Born Iterative Method is adopted as a basic framework for iteratively found the solution, which however is affected by a slow convergence and high sensitivity to noise, thus resulting into sub-optimal reconstructions. To address these issues, a U-Net convolutional neural network architecture is incorporated to significantly improve the estimation of the permittivity distribution. In addition, a tumor segmentation adopting U-Net and ResNet-18 is further applied to the reconstructed models, thus resulting into a precise definition of tumor boundaries, as well as into a better differentiation between tumor regions from healthy tissue.