Machine Learning Approach to Enhanced Resolution of Inverse Scattering for Cancer Detection

Microwave inverse scattering is an imaging technique potentially useful in medical diagnostics to non-invasively determine the tissue properties of the body, including the detection of tumors and other lesions. However, its implementation has inherent challenges, since Inverse Scattering Problems (ISPs) are typically nonlinear and ill-posed. This aspect implies that conventional techniques usually provide inaccurate reconstructions, with limited resolution and high computational costs, when strong scatterers or inhomogeneities are present in the investigation domain. To overcome the above disadvantages, a framework based on Convolutional Neural Net-works (CNN), combined with the quadratic Born Iterative Method (quadratic BIM), is presented to accurately reconstruct and identify brain tumors. Specifically, two stages can be identified. First, the quadratic BIM approach is applied to recover an initial estimate of the complex permittivity of brain phantoms. Then, a U-Net architecture is adopted to achieve high-quality final reconstructions. Numerical results are presented to demonstrate more accurate reconstructions as compared to traditional BIM, with optimal computational costs and excellent resolution.