Deep-Learning-Enabled Microwave-Induced Thermoacoustic Tomography Based on Sparse Data for Breast Cancer Detection

Abstract

As a rapidly developing novel electromagnetic imaging technique, microwave-induced thermoacoustic tomography (MITAT) has found many applications and attracted tremendous research interest. Using sparse data to reconstruct images is very challenging for MITAT. This work proposes a novel deep-learning-enabled MITAT (DL-MITAT) modality to address the sparse data reconstruction problem and applies it in breast cancer detection. The applied network is a domain transform network called feature projection network (FPNet) + ResU-Net. Detailed structure and implementation method of the network is described. We conduct both simulation and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ex vivo</i> experiments with breast phantoms to test the validity of the DL-MITAT approach. The obtained images given by the trained network exhibit much better quality and have much less artifacts than those obtained by a traditional imaging algorithm. We show that only 15 measurements can still reliably recover an image of the breast tumor for both full-view and limited-view configurations in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ex vivo</i> experiments. We also provide detailed discussions on the capability and limitations of the proposed scheme. This work presents a new paradigm for MITAT based on sparse data and can be applied in all related applications of MITAT, including biomedical imaging, nondestructive testing, and therapy guidance.