Deep learning based high frame rate photoacoustic tomography

Abstract

Generating an image of acceptable quality will take several minutes in circular scanning geometry-based Photoacoustic tomographic (PAT) imaging systems. Although, the imaging speed can be improved by employing multiple single-element ultrasound transducers (UST) and faster scanning. The low signal-to-noise ratio at higher and the artifacts arising from sparse signal acquisition hamper the imaging speed. Thus, there exists a need to improve the speed of the PAT imaging system without compromising the image quality. To improve the frame rate of the PAT system, we propose a convolutional neural network (CNN) based deep learning architecture for reconstructing the artifact-free PAT images from the fast-scanning data. The proposed model is trained with the simulated dataset and its performance was evaluated using experimental phantom and in-vivo imaging. The efficiency to improve the frame rate was evaluated on both the single-UST and multi-UST PAT systems. Our results suggest that the proposed deep learning architecture improves the frame rate by six-fold in a single UST PAT system and by two-fold in a multi-UST PAT system. The fastest frame rate of ~ 3Hz was achieved without compromising the quality of the PAT image.