Unsupervised Photoacoustic Tomography Image Reconstruction from Limited-View Unpaired Data using an Improved CycleGAN

Abstract

Photoacoustic tomography (PAT) is a hybrid imaging method with great applications in preclinical research and clinical applications. However, due to the limited-view issue, it is often hard to cover the desired tissue completely, thus resulting in severe artifacts in reconstructed images. Enhancing a reconstructed image to become artifact-free could be considered an image-to-image translation task which is addressed easily by the well-known Pix2Pix generative adversarial network (GAN). Training Pix2Pix usually requires a large paired dataset. Preparing such datasets can be difficult or even in some cases impossible. In this paper, we propose an improved unsupervised reconstruction method based on cycle-consistent adversarial networks (CycleGAN), to overcome the need for paired datasets. CycleGAN can learn image-to-image translation tasks from an unpaired dataset without the need for one-to-one matching between low-quality and high-quality images. Experimental results demonstrate that the proposed architecture outperforms the original CycleGAN in terms of image similarity metrics including PSNR and SSIM.