Speaker
Description
In camera-based full-field detection photoacoustic tomography, snapshots of the acoustic pressure field are recorded with an optical camera, yielding two-dimensional projections of the wave pattern at a fixed time after excitation. For a constant speed of sound, the initial pressure projection can be recovered exactly via a Radon-domain backpropagation method: the Radon transform decomposes the wave pattern into two counter-propagating bands, which are shifted back to time zero using d’Alembert’s formula. A key property of this decomposition is that both bands encode equivalent source information, providing inherent redundancy. In practice, parts of the wave pattern may be obscured by the sample, its holder, or the limited field of view, resulting in spatially masked regions with missing data. We propose a self-supervised neural network approach to inpaint the masked wave pattern directly in the image domain. The network takes the incomplete wave pattern as input and predicts the missing regions. Crucially, no ground-truth training data are required. Instead, training is driven by two complementary loss terms: (i) a data-fidelity loss enforcing consistency with the measured pixels and (ii) a Radon-domain self-consistency loss that exploits the redundancy of the two counter-propagating bands. If the data completion is correct, both bands must yield identical reconstructions of the initial pressure after shifting by ±cT. We evaluate the method on simulated data with various mask geometries and compare the reconstruction quality against iterative approaches and supervised baselines.