Examining perivascular spaces (PVS) in cerebral small vessel disease (CSVD) using a novel T1- based automated PVS segmentation tool

Introduction: White matter hyperintensities (WMH) and MRI-visible perivascular spaces (PVS) are neuroimaging features of cerebral small vessel disease (CSVD). PVS are believed to play a role in cerebral metabolic waste clearance, particularly during sleep, and emerging evidence suggests that sleep disturbances are among the earliest symptoms of dementia. However, PVS quantification remains challenging and not accessible, particularly in complex patients with neurovascular and neurodegenerative disease. We developed and validated an automated deep learning-based PVS quantification method using multi-site patient MRI with varying degrees of CSVD burden. Additionally, we examined the correlation between PVS volumes and age in patients undergoing treatment for sleep apnea. Methods: MRI (training/validation data=141; testing=15) were obtained from various multi-site studies (ONDRI, CAHHM, Leducq SVD-PVS, CAIN). We developed a deep learning convolutional neural net (CNN) model with a U-net architecture for PVS segmentation using T1-weighted images. Ground truth data used for model training were generated by first applying the RORPO filter to extract small tubular structures, then refinement of the RORPO output around WMH, followed by removal of probable non-PVS objects from the RORPO-based output using Freesurfer regions of interest, and finally manually correcting the remaining errors. After ground truth generation and model training, the final CNN output was used to examine the association between PVS volumes and age in patients with sleep apnea (n=42). Results: After CNN training, our PVS tool completes full segmentation in under 3 minutes (NVIDIA RTX3090). It achieved excellent performance on the test data, with a mean Dice score of 0.95, indicating outstanding agreement with ground truth. When applied to the patients undergoing interventional sleep apnea treatment, a strong positive relationship was found between total PVS volume and age (r=0.57 [0.38 0.72]). Discussion: Our findings suggest that our automated PVS segmentation method can quickly and accurately quantify PVS in neurodegenerative and neurovascular patients with varying degrees of CSVD burden. The strong association between age and PVS in patients with sleep apnea further demonstrates the utility of our method and underscores the importance of further investigating the role of PVS in the context of sleep, aging, and neurodegenerative processes.