Compressive sensing combined deep learning for fast microwave-induced thermoacoustic tomography

Abstract

Breast cancer is the second leading cause of cancer death in women, and early detection of breast cancer is essential for more effective treatment. Recently, microwave-induced thermoacoustic tomography (MITAT) based on compressive sensing has been proven to have great potential as a new detection tool for early breast cancer within low sampling time. However, the traditional MITAT reconstruction method based on compressive sensing requires many computing resources. To find a balance between low computing resources and high-resolution images for the method based on compressive sensing, especially in the environment of a non-uniform tissue, we developed a MITAT based on deep learning (DL-MITAT) imaging scheme compressive sensing-super-resolution thermoacoustic imaging (CS-SRTAI) method which decomposed the single reconstruction step into the initial reconstruction part and the super-resolution part. The initial reconstruction part extracts the necessary physical information into the low-resolution image space. The super-resolution part maps the low-resolution image space to the high-resolution image space. Specifically, we proposed a neural network in the super-resolution part. Both numerical simulation and the experiment demonstrate the effectiveness of the proposed method. The proposed method achieved 88% structural similarity index measure within computing resources of 21 s and 1.0 GB for the numerical simulation. Moreover, for the real breast tumor and non-uniform tissue experiment, the CS-SRTAI performs well at recovering the location, size, and number of the tumor within computing resources of 65 s and 1.1 GB. It is worth noting that the proposed DL-MITAT imaging strategy reduces computing resources with great imaging quality. It is promising to use in the fields where the computing resources for imaging are restricted.