WHITE-Net : White matter HyperIntensities Tissue Extraction using deep learning Network

White matter hyperintensities (WMH) are a hallmark of cerebral small vessel disease, a highly prevalent condition in aging, making their accurate detection in large-scale magnetic resonance imaging (MRI) studies critically important. Currently available semi-automated and automated segmentation methods are frequently limited by sensitivity to variations in imaging conditions and computational constraints, restricting their applicability across diverse acquisition settings. We present WHITE-Net, a deep learning-based framework for automated WMH segmentation built on a 3D ResUNet architecture and trained on multisite MRI data encompassing substantial variability in scanner vendors, acquisition protocols, spatial resolution, and image characteristics. Evaluated across three independent datasets - BrainLaus, ADNI, and WMH Challenge, WHITE-Net achieves high segmentation accuracy and ranks among the top-performing methods across datasets, enabling a comprehensive assessment of generalisability in varied imaging environments. WHITE-Net maintains stable performance in a wide spectrum of lesion loads and demonstrates a favorable balance between precision and recall, effectively limiting false positive detections - a common limitation of existing approaches. Beyond accuracy, WHITE-Net requires no parameter tuning and offers fast inference times, making it well-suited for deployment in large-scale neuroimaging studies. The obtained results highlight the value of training on diverse multisite data for improving generalisability and position WHITE-Net as a reliable, scalable tool for automated WMH segmentation in computational anatomy research.