Cascade neural approximating for few-shot super-resolution photoacoustic angiography

Abstract

High-resolution photoacoustic angiography images are reconstructed from undersampled images with the help of a super-resolution deep neural network, enhancing the ability of the photoacoustic angiography systems to image dynamic processes in living tissues. However, image degradations are difficult to estimate due to a lack of knowledge of the point spread function and noise sources, resulting in poor generalization capability of the trained super-resolution model. In this work, a high-order residual cascade neural network was developed to reconstruct high-resolution vascular images, which is a neural approximating approach used to remove image degradations of photoacoustic angiography. To handle overfitting in training super-resolution model with a limited dataset, we proposed a BicycleGAN based image synthesis method in data preparation, achieving a strong regularization by forging realistic photoacoustic vascular images that act to essentially increase the training dataset. The quantitative analysis of the reconstructed results shows that the high-order residual cascade neural network surpassed the other residual super-resolution neural networks. Most importantly, we demonstrated that the generalized model could be achieved despite the limited training dataset, promising to be a methodology for few-shot super-resolution photoacoustic angiography.