Matrix-free photoacoustic image reconstruction via sensor-token networks
TL;DR
A direct, matrix-free reconstruction network, the Sensor-Token Reconstruction Network (STRN), which represents each sensor's time series as a token in a Transformer encoder and maps raw measurements to the reconstructed image in a single forward pass, without invoking the system matrix at inference.
Abstract
Photoacoustic tomography (PAT) couples optical absorption contrast with ultrasound resolution, but reconstructing the initial pressure distribution from sparse-view measurements is an ill-posed inverse problem. Iterative compressive-sensing solvers and unrolled deep networks both retain a dependence on the system matrix at inference, which makes fast reconstruction computationally expensive. We present a direct, matrix-free reconstruction network, the Sensor-Token Reconstruction Network (STRN), which represents each sensor's time series as a token in a Transformer encoder and maps raw measurements to the reconstructed image in a single forward pass, without invoking the system matrix at inference. An analytical k-space H-matrix was constructed, validated against the k-Wave pseudo-spectral solver (mean per-sensor Pearson correlation r ̄ = 0 . 919 ± 0 . 049 ), and used to synthesize training data. STRN was benchmarked on a 46-sample held-out set against ISTA, split-Bregman total variation (SBTV), and learned ISTA (LISTA). STRN attained the best overall fidelity, with the highest SSIM (0.522 vs. 0.212 for ISTA), a PSNR gain of 7.8–12.7 dB, and the lowest NMSE (0.233), all significant at p < 10 −8 under paired t-tests and Wilcoxon signed-rank tests. On independently k-Wave-simulated measurements it showed a substantial but structure-preserving and rapidly closable domain gap (SSIM 0.52 → 0.19 zero-shot, Pearson 0.79 retained; ~160 adaptation samples recovered ~90% of in-distribution fidelity). A parameter-matched ablation showed the self-attention sub-layers are not the source of accuracy: a matched attention-free variant matched or exceeded the full model, with the learned spatial decoder as the dominant component. The analytical operator is an efficient training surrogate rather than a source of an inverse crime, and fidelity is driven by the learned decoder rather than self-attention. By removing the per-iteration Hx multiplication, STRN reduces reconstruction time by at least an order of magnitude on the same hardware, and we report the architectural finding openly to guide simpler direct-reconstruction designs.