Evaluation of Fully Convolutional Networks for Dielectric Profile Reconstruction in Medical Microwave Imaging

The capabilities of deep learning networks in reconstructing the dielectric profile in medical microwave imaging are evaluated. Thus, different fully convolutional networks are trained on unhealthy brain (with lesions) data and then tested to generate dielectric profiles of brains using wideband (0.5-2 GHz) signals generated from a brain imaging domain. The quantitative evaluation results of UNet3+ show the best reconstruction performance not only in unhealthy brain but also in healthy brain (without lesion) tests. This work offers guidance on capabilities and limitations of fully convolutional deep learning networks for dielectric profile reconstruction in medical microwave imaging.