Deep Learning Based Non-Iterative Solution to the Inverse Problem in Microwave Imaging

Abstract

A deep learning-based approach in conjugation with Fourier Diffraction Theorem (FDT) is proposed in this paper to solve the inverse scattering problem arising in microwave imaging. The proposed methodology is adept in generating a permittivity mapping of the object in less than a second and hence has the potential for real-time imaging. The reconstruction of the dielectric permittivity from the measured scattered field values is done in a single step as against that by a long iterative procedure employed by conventional numerical methods. The proposed technique proceeds in two stages; with the initial estimate of the contrast function being generated by the FDT in the first stage. This initial profile is fed to a trained U-net to reconstruct the final dielectric permittivities of the scatterer in the second stage. The capability of the proposed method is compared with other works in the recent literature using the Root Mean Square Error (RMSE). The proposed method generates an RMSE of 0.0672 in comparison to similar deep learning methods like Back Propagation-Direct Sampling Method (BP-DSM) and Subspace-Based Variational Born Iterative Method (SVBIM), which produce error values 0.1070 and 0.0813 in the case of simulation (using Austria Profile). The RMSE level while reconstructing the experimental data (FoamDielExt experimental database) is 0.0922 for the proposed method as against 0.1631 and 0.1037 for BP-DSM and SVBIM, respectively.

Reconstruction of Mixed Boundary Objects and Classification Using Deep Learning and Linear Sampling Method

“…Deep learning is emerging as a dominant tool for providing precise and reliable solutions in the field of microwave imaging, yet maintaining computational efficiency [12][13][14]. A deep learning-based approach in conjugation with the Fourier diffraction theorem is proposed in [15] to solve the inverse scattering problem encountered in quantitative microwave imaging (MWI). In [16], the orthogonality sampling method (OSM) is used along with deep learning architecture called the U-Net to reconstruct the permittivites of the dielectric objects.…”

Applications of Microwaves in Medicine Leveraging Artificial Intelligence: Future Perspectives

“…Typically, iterative inversion methods such as the Born iterative method (BIM), distorted Born iterative method (DBIM), contrast source inversion, etc., are used, but even with advances in numerical methods, solving the inverse problem is still challenging due to slow convergence, non-linearities, and ill-posedness leading to false solutions and unstable outcomes. This difficulty is further complicated by the 3D nature of the imaging domain, increasing the computational demand and processing times [14,15]. This is where deep learning (DL), a subset of artificial intelligence (AI), comes to the rescue, as it can quickly reconstruct the images within a few seconds or minutes, making the overall process suitable for real-time applications.…”