A Deep Learning Approach for the Photoacoustic Tomography Recovery From Undersampled Measurements

Photoacoustic tomography (PAT) is a propitious imaging modality, which is helpful for biomedical study. However, fast PAT imaging and denoising is an exigent task in medical research. To address the problem, recently, methods based on compressed sensing (CS) have been proposed, which accede the low computational cost and high resolution for implementing PAT. Nevertheless, the imaging results of the sparsity-based methods strictly rely on sparsity and incoherence conditions. Furthermore, it is onerous to ensure that the experimentally acquired photoacoustic data meets CS's prerequisite conditions. In this work, a deep learning-based PAT (Deep-PAT)method is instigated to overcome these limitations. By using a neural network, Deep-PAT is not only able to reconstruct PAT from a fewer number of measurements without considering the prerequisite conditions of CS, but also can eliminate undersampled artifacts effectively. The experimental results demonstrate that Deep-PAT is proficient at recovering high-quality photoacoustic images using just 5% of the original measurement data. Besides this, compared with the sparsity-based method, it can be seen through statistical analysis that the quality is significantly improved by 30% (approximately), having average SSIM = 0.974 and PSNR = 29.88 dB with standard deviation ±0.007 and ±0.089, respectively, by the proposed Deep-PAT method. Also, a comparsion of multiple neural networks provides insights into choosing the best one for further study and practical implementation.