Depth-resolved phase velocity estimation in layered tissue based on an efficient additive attention network with surface acoustic wave – optical coherence elastography

This study developed a depth-resolved phase velocity estimation framework for surface acoustic wave optical coherence elastography in layered tissues. The authors combined numerical spectral analysis with a deep learning inversion model, PVNet, built on an efficient additive attention network to estimate depth-specific phase velocity directly from optical coherence elastography phase images. The framework was validated in homogeneous agar phantoms, layered agar phantoms, and in vivo human skin from the palm, forearm, back of the hand, and face. Numerical spectral analysis produced phase velocities that agreed closely with independent mechanical testing in agar phantoms, while layered phantoms showed clear detection of subsurface interfaces through abrupt phase velocity transitions. In human skin, the method identified depth-dependent phase velocity gradients and epidermal interface locations across anatomical regions with different layer thicknesses. PVNet achieved the best overall predictive performance among compared models, with mean absolute errors of 0.123 ± 0.024 m/s in agar models and 0.145 ± 0.114 m/s in human skin, supporting the use of additive-attention deep learning for clinically relevant, depth-resolved stiffness assessment in layered tissues.