Phase-driven complex-valued network for quantitative microwave-induced thermoacoustic tomography

Deep learning has demonstrated significant strides in the field of microwave-induced thermoacoustic imaging, achieving superior imaging accuracy compared to conventional methods. However, the current end-to-end reconstruction networks predominantly process time-domain signals while overlooking the critical influence of sensor–tissue spatial relationships on feature extraction, which may degrade imaging fidelity. To address this, we propose a complex-valued neural network architecture for quantitative reconstruction of microwave thermoacoustic conductivity. The proposed framework incorporates two essential components: a 2D Fourier transform operator that converts sensor signals into spatial-frequency representations and complex-valued matrix operations that synergistically optimize amplitude-phase feature learning. Comprehensive numerical simulations and experiments validate that our method achieves higher indicators compared to conventional algorithms and real-valued networks, establishing a state-of-the-art framework in conductivity map reconstruction.