AI and the World's Hard Problems · Physics · arXiv · 2026-10-19
Untrained CNN and coefficient-space regularization for joint reconstruction in photoacoustic tomography
The paper addresses joint reconstruction of initial pressure and spatially varying sound speed in photoacoustic tomography from boundary measurements. An untrained CNN supplies fixed nonlinear features, from which the initial-pressure wavelet coefficients are estimated by Tikhonov-regularized least squares, while a learned coefficient-space prior is used for the sound speed. The method reconstructs accurately on held-out data and generalizes to unseen phantom geometries, and it yields substantially smaller errors than a supervised trained CNN while needing about one-sixth of the computation time.