DRIPS: Domain Randomisation for Image-based Perivascular spaces Segmentation

DRIPS, a synthetic-data-trained deep-learning framework, significantly outperformed prior methods for perivascular space segmentation across eight MRI cohorts, with mean AUPRC gains of at least 0.1.

• Why it matters: Robust PVS segmentation is hindered by the need for cohort-specific tuning in conventional methods and by limited generalisation and heavy annotation demands in supervised deep learning.
• What they did: DRIPS was trained exclusively on synthetic data and evaluated on 185 fully manually annotated MRI scans from eight independent cohorts spanning 3T and 7T, T1- and T2-weighted sequences, and isotropic and anisotropic resolutions.
• The result: Across cohorts, DRIPS beat the second-best method for AUPRC and clDice, including 0.146 (BG AUPRC), 0.125 (CSO AUPRC), 0.112 (BG clDice), and 0.186 (CSO clDice), and its performance was not associated with white matter hyperintensity burden.