Abstract
<jats:p>Quantitative optical coherence tomography (qOCT) enables depth-resolved mapping of tissue attenuation coefficients, yet its accuracy is fundamentally limited by speckle noise, multiple scattering, and the ill-posed nature of the inverse problem. Existing deep learning methods, whether purely data-driven or based on simplified scattering models, fail to capture full-wave electromagnetic propagation, leading to poor generalization and degraded reconstruction accuracy. We propose SS-PhyNet, a spatial-spectral dual-domain physics-informed network that enforces full-wave electromagnetic constraints directly from Maxwell's equations, combined with data-driven priors for high-reliability inversion. In the spatial domain, we construct a hybrid U-shaped backbone with ConvNeXt blocks, Haar wavelet downsampling, and Transformer attention for multi-scale feature extraction and edge preservation. In the spectral domain, a pretrained Fourier neural operator (FNO) serves as a differentiable surrogate for the forward model, imposing wave-equation constraints during training at substantially reduced computational cost. Simulations and phantom experiments demonstrate that SS-PhyNet reduces the physical reconstruction error to below 30% of the best-performing baseline on simulated data and by approximately 47% on phantom data. On in vivo retinal datasets without ground truth, SS-PhyNet achieves the highest contrast and contrast-to-noise ratio, with physical consistency error halved relative to competing methods. Noise robustness testing shows that the structural similarity index remains above 0.88 under interference from 5 dB to 25 dB. By combining full-wave physical constraints with data-driven learning, SS-PhyNet improves both the accuracy and physical credibility of attenuation coefficient inversion while maintaining low computational cost, offering a reliable solution for physics-informed inverse problems in optical imaging.</jats:p>