Machine Learning-based Multi-resolution Algorithm for Inverse Electromagnetic solution towards Breast Cancer Detection

Studies on breast cancer statistics worldwide indicate the necessity of early detection for improved mortality rates. Radio-Frequency (RF) based 3D imaging provides a low-cost, non-invasive, non-ionizing alternative to present day early diagnostic procedures. In RF imaging, the measured field data from an antenna array is used to predict malignant tissues in the breast profile through reconstruction of the dielectric properties. The reconstruction process entails the solution of a non-linear, ill-posed inverse problem. In this work, a multi-resolution approach is proposed to address the computational challenges of the traditional uniform grid approach. First, the measured scattered field data is fed to a Machine Learning based classifier to localize the dense tissue with coarse accuracy. Using this information, adaptively sized voxels are generated leading to a drastic reduction in the number of voxels and hence the number of unknowns for reconstruction. The reduced problem is solved using a traditional optimization algorithm like the Levenberg-Marquardt (LM) method. Numerical experiments demonstrate that the proposed method has significant advantages both in convergence profile and reconstruction efficiency as compared to that on a uniform grid.