A Novel Framework to Synthesize Arterial Spin Labeling Images using Difference Images

Abstract

Arterial spin labeling (ASL) images that are capable to quantitatively measure the cerebral blood flow receive increasing research attention in recent dementia diseases diagnosis studies. However, this important and relatively new imaging modality is unfortunately not commonly seen in many well-established image-based dementia datasets, including the ADNI-1/2/3/Go datasets. Hence, synthesizing ASL images to supplement this important modality is valuable. In this study, a new framework based on a cascade of generative adversarial networks (GANs) and difference images generated from a Laplacian pyramid is proposed. This framework is novel as it is the first attempt to incorporate difference images for synthesizing medical images. Experimental results based on a 355-demented patient dataset and ADNI-1 dataset suggest that, this new framework outperforms all state-of-the-arts in ASL image synthesis. Also, synthesized ASL images obtained by this new framework are capable to significantly improve the accuracy of dementia diseases diagnosis performance.