Sparse-sampling photoacoustic computed tomography: Deep learning vs. compressed sensing

Abstract

Sparse-sampling photoacoustic computed tomography (SS-PACT) is an effective high-speed and low-cost modality in photoacoustic imaging. For recovering high-quality images from the sparsely sampled data in SS-PACT, two methods have been discussed in the literature in recent years: compressed sensing (CS) and deep learning (DL) methods. While both ways improve image quality, however, no work has been conducted to compare the reconstruction performance of the two methods comprehensively. Such comparison would facilitate the selection of appropriate techniques for improving the image quality for given imaging parameters. Here, we conduct in vivo imaging of SS-PACT on a human hand and a rat to compare an advanced CS reconstruction model with partially known support (CS-PKS) and a typical DL reconstruction network of U-Net. Experimental results demonstrate that both CS and DL can effectively suppress the reconstruction artifacts and dramatically improve the reconstructed image's signal-to-noise ratio. Compared to CS, DL has a higher imaging speed and can reconstruct images similar to CS when the sparse-sampling rate is not very high. However, when a lower sampling rate is used, some weak and tiny signals are lost in the recovered images from DL while CS recovers more accurate photoacoustic images, In summary, CS and DL methods are beneficial under different circumstances, and our research would potentially provide principles for optimal selection in the practical use of SS-PACT in the clinical setting.