Prediction of White Matter Hyperintensity Progression: A Systematic Comparison of Preprocessing Methods and Deep Learning Models

White matter hyperintensity (WMH) progression is a critical marker of cerebral small vessel disease; however, its prediction remains challenging due to inter-institutional variability and the inherently non-deterministic nature of lesion evolution. To our knowledge, this is the first study to systematically evaluate preprocessing strategies and prediction frameworks for WMH progression using a large-scale, multi-institutional dataset from the PICASSO trial (67 institutions). We compared three preprocessing pipelines, harmonization, normalization, and their combination, with three prediction frameworks: fluid-attenuated inversion recovery (FLAIR)-to-FLAIR, WMH-to-WMH, and disease evolution map (DEM). These were assessed in conjunction with multiple deep learning architectures. Our results demonstrate that harmonization substantially reduces inter-institutional variability and that the DEM framework consistently outperforms direct image-to-image prediction in volumetric and spatial agreement. Furthermore, we provide the first direct comparison of U-Net, GAN, and diffusion models under identical experimental conditions, showing that diffusion models deliver balanced performance in capturing non-deterministic lesion dynamics. At the same time, U-Net achieves the highest spatial agreement. Collectively, these findings establish the first comprehensive benchmark of WMH progression prediction across preprocessing strategies, prediction frameworks, and model architectures, and underscore the translational potential of DEM-based prediction with harmonization for individualized risk stratification and clinical decision-making.