Improving the prediction of white matter hyperintensities evolution in brain MRI of patients with small vessel disease using stroke lesions information

Abstract

ABSTRACTPredicting the evolution of white matter hyperintensities (WMH) (i.e., whether WMH will grow, remain stable, or shrink with time) is important for personalised therapeutic interventions. However, this task is difficult mainly due to the myriad of vascular risk factors (VRF) and comorbidities that influence the evolution of WMH, and the low specificity and sensitivity of the intensities and textures alone for predicting WMH evolution. Given the predominantly vascular nature of WMH, in this study, we evaluate the impact of incorporating stroke information to a probabilistic deep learning model to predict the evolution of WMH 1-year after the baseline image acquisition using brain T2-FLAIR MRI. The Probabilistic U-Net was chosen for this study due to its capability of simulating and quantifying uncertainties involved in the prediction of WMH evolution. We propose to use an additional loss called volume loss to train our model, and incorporate an influential factor of WMH evolution, namely, stroke lesions information. Our experiments showed that jointly segmenting the disease evolution map (DEM) of WMH and stroke lesions, improved the accuracy of the DEM representing WMH evolution. The combination of introducing the volume loss and joint segmentation of DEM of WMH and stroke lesions outperformed other model configurations with mean volumetric absolute error of 0.0092ml(down from 1.7739ml) and 0.47% improvement on average in shrinking, growing and stable WMH using Dice similarity coefficient.