A novel locally-constrained GAN-based ensemble to synthesize arterial spin labeling images

Abstract

Arterial spin labeling (ASL) images that are capable to quantitatively measure the cerebral blood flow receive increasing research attention in recent dementia diagnosis studies. However, this important imaging modality is unfortunately not commonly seen in many well-established image-based dementia datasets. Hence, synthesizing ASL images to supplement the important modality in these datasets for further improving the accuracy of dementia diseases diagnosis is quite important and valuable. In this study, a novel locally-constrained generative adversarial networks (GAN)-based ensemble is introduced to fulfill the ASL image synthesis task for improving the dementia diseases diagnosis performance. Technically, new attention-based feature pyramid-GAN models are designed as local models of the novel ensemble. Also, multi-Gaussian-distributed noise is generated from a new flow-based generative model and utilized in medical image synthesis, for the first time. Experiments have been conducted to reveal the effectiveness of the novel GAN ensemble. Comparisons between the novel GAN ensemble and many other state-of-the-art methods in medical image synthesis have been carried out. Statistical analyses have suggested that, accuracies of dementia diseases diagnosis can be significantly improved with the help of the novel GAN ensemble, which brings about 41.62% performance improvement based on a 355-demented-patient dataset and approximately 25% performance improvement from the well-known ADNI-1 dataset.