A physics-assisted deep learning method for high-fidelity breast microwave image reconstruction in electromagnetic inverse scattering

Abstract

Microwave imaging is a promising next-generation breast imaging technology. At present, cutting-edge microwave image reconstruction methods based on deep learning (such as various variants of convolutional neural networks) mostly adopt the general convolution modules and architectures from visual image processing and do not fully integrate physical priors. For breast tissues with strong heterogeneity and complex structures, the existing methods have prominent drawbacks, such as limited reconstruction accuracy and excessive model parameters. To address these issues, this paper proposes a lightweight deep learning framework based on physical assistance for high-fidelity microwave image reconstruction. First, a physically reliable initial reconstruction is generated using the Born iterative method. Subsequently, the Microwave Image Enhancement Network (MIENet) is proposed to reconstruct high-fidelity microwave images with rich details. Within the MIENet architecture, a group progressive residual module is proposed that achieves both superior detail reconstruction capability and computational efficiency, and the convolutional block attention module is integrated into the U-shaped architecture to enable the selective learning of key features. In addition, a hybrid loss function integrating multiple evaluation criteria is proposed, which effectively guides the network to generate high-fidelity microwave images. Experiments conducted on the breast dataset show that MIENet outperforms classical image reconstruction architectures in terms of image quality while being more lightweight: it achieves a peak signal-to-noise ratio of 39.14 dB and a structural similarity index of 0.9904, with a total parameter count and floating-point operations accounting for only 67.16% and 1.3% of those of state-of-the-art networks, respectively.