Quantitative Reconstruction of Dielectric Properties Based on Deep-Learning-Enabled Microwave-Induced Thermoacoustic Tomography

Abstract

Quantitative reconstruction of dielectric properties has enabled a wealth of biomedical applications. Although traditional microwave imaging and microwave-induced thermoacoustic tomography (MITAT) techniques have been widely explored for quantitative reconstruction, it is still highly challenging for them to deal with biological samples with high permittivity and conductivity. This work leverages deep-learning-enabled MITAT (DL-MITAT) approach to quantitatively reconstruct dielectric properties of biological samples with high quality. We construct a new network structure to separately reconstruct the permittivity and conductivity. By simulation and experimental testing, we demonstrate that the DL-MITAT technique is able to reliably reconstruct inhomogeneous biological samples with tumor, muscle, and fat. The experimental reconstruction error is only 5%. The network exhibits excellent generalization capability in terms of sample's geometry. This work provides a useful paradigm and alternative way for quantitative reconstruction of dielectric properties and paves the way toward practical applications.