Predicting the Evolution of White Matter Hyperintensities in Brain MRI using Generative Adversarial Networks and Irregularity Map

We propose a Generative Adversarial Network (GAN) model named disease evolution predictor GAN (DEP-GAN) to predict the evolution (i.e., progression and regression) of white matter hyperintensities (WMH) in small vessel disease. In this study, the evolution of WMH is represented by the “disease evolution map” (DEM) produced by subtracting irregularity map (IM) images from two time points: baseline and follow up. DEP-GAN uses two discriminators (critics) to enforce anatomically realistic follow up image and DEM. To simulate the nondeterministic and unknown parameters involved in WMH evolution, we propose modulating an array of random noises to the DEP-GAN’s generator which forces the model to imitate a wider spectrum of alternatives in the results. Our study shows that the use of two critics and random noises modulation in the proposed DEP-GAN improves its performance predicting the evolution of WMH in small vessel disease. DEP-GAN is able to estimate WMH volume in the follow up year with mean (std) estimation error of -1.91 (12.12) ml and predict WMH evolution with mean rate of 72.01% accuracy (i.e., 88.69% and 23.92% better than Wasserstein GAN).