Imaging-Process-Informed Generative Strategy for Microwave Medical Image Processing

Abstract

Microwave medical imaging (MMI) has significant potential for stroke imaging due to its unique portability and timeliness, and is therefore highly valued by researchers. This paper proposes an imaging-process-informed generative strategy (IPIGS) that addresses the challenges posed by nonlinearity in the imaging process, thereby generating more intuitive and diagnosis-friendly microwave medical images. Aiming at the limitations of traditional imaging methods (e.g., distorted Born iterative method, DBIM), namely the blurry imaging background and the confusion of stroke areas, we first attribute this issue to the nonlinearity in the imaging process. That is, under certain circumstances, the dielectric constants of different brain tissues are misinterpreted by the DBIM imaging algorithm, resulting in incorrect imaging of stroke areas. Then, we establish a generalized model for this nonlinear problem, and introduce the corresponding a prior knowledge (APK), such as brain tissue information that is sensitive to changes in dielectric constants, into the generative adversarial network Pix2pix to restore the unclear imaging caused by this problem. The final results show that our method has a significant improvement over traditional methods in terms of the structural similarity and threshold segmentation effect, with an increase in structural similarity of 8% compared to traditional methods and an average increase in intersection-over-union of approximately 47%. This provides new possibilities for the restoration of microwave images.