Sensor Attention Network Enables Matrix-Free Photoacoustic Image Reconstruction

A novel architecture based on Transformers, known as the Sensor Attention Network (SAN), has been introduced for the reconstruction of photoacoustic tomography (PAT). This technique merges optical absorption contrast with ultrasound resolution, yet the challenge lies in recovering initial pressure from sparse sensor data, which is an ill-posed inverse problem. Existing iterative compressive-sensing methods and unrolled deep networks require the system matrix during inference, leading to high computational costs for real-time clinical applications. In contrast, SAN views the complete time series from each sensor as a token, directly translating raw data into the reconstructed image without needing the system matrix. An analytical k-space H-matrix was developed for training and validation against the k-Wave pseudo-spectral solver, achieving a mean per-sensor Pearson correlation of 0.919 ± 0.049. This research is available on arXiv with ID 2607.25576.