PAFormer: Photoacoustic reconstruction via transformer with mask mechanism

As an emerging biomedical imaging technology, Photoacoustic tomography (PAT) can provide high contrast and resolution images of optical absorption at depths, which is greater than traditional optical imaging techniques. However, the imaging quality of PAT usually suffers from limited view and sparse channels signal. Conventional reconstruction algorithms (e.g., time reversal) always lead to a degenerated solution due to its inherently ill-posed feature. Although deep learning has shown its potential to reconstruct from sparse channels signal in a data-driven way, existing methods can not handle well the domain gap between signal and image. In this work, we propose a new architecture, PA-Former, to recover the full channels signal from sparse input. To excavate the inter-channel relationship between each channel of the PA signal, we design a channel-wise attention mechanism, which can robustly improve the reconstructed full-sampled signals. We train the model on a public clinical dataset and in-vivo fish dataset. PA-Former shows great performance on both two datasets of sparse data.