Zero-shot physics-supervised neural network for data-free photoacoustic tomography reconstruction

Abstract

To address the ill-posedness inherent in photoacoustic tomography reconstruction, we propose ZS-PSNN, a zero-shot physics-supervised neural network framework that eliminates the reliance on external paired training data. By explicitly constructing a discretized acoustic system matrix and embedding it as a differentiable forward operator within the loss function, the proposed method enforces acoustic measurement consistency to supervise the neural network optimization directly from a single scan. Results on simulated vascular data and experimental mouse data demonstrate that ZS-PSNN achieves competitive reconstruction performance compared with FBP, TR, model-based TV regularization, plug-and-play TV, and implicit neural representation methods. This work establishes a single-measurement physics-supervised neural optimization strategy for zero-shot photoacoustic reconstruction applicable to scenarios where high-quality paired training data are challenging to obtain.