Fig. 3

MRI processing steps. To obtain the WMH mask, we applied the algorithm developed by Andermatt et al., which employs Multidimensional Gated Recurrent Units with optimizations including data augmentation, selective sampling, residual learning, and DropConnect in the RNN state. This algorithm utilized T1 and FLAIR images to generate the WMH mask in the native T1 space. Additionally, the CAT 12 toolbox (Computational Anatomy Toolbox 12, version 12.8_1977), implemented in Statistical Parametric Mapping (SPM12, r7771, Wellcome Trust Centre for Neuroimaging, University College London), was used to generate GM/WM/CSF masks in native space for each participant. The derived WMH and WM masks facilitated the extraction of the average susceptibility value within WMH and nonWMH WM voxels. Abbreviations: WMH = white matter hyperintensity