A Comprehensive Evaluation of Multiple Enlarged Perivascular Space Segmentation Tools

TL;DR

This paper compares several automated tools for identifying enlarged perivascular spaces in brain MRI scans and finds that a new deep-learning algorithm called DORES performs best.

Contribution

The paper introduces and validates DORES, a new deep-learning algorithm for segmenting enlarged perivascular spaces in MRI scans.

Findings

DORES outperformed other algorithms in detecting and segmenting enlarged perivascular spaces in both basal ganglia and white matter regions.

DORES achieved a Dice score of 0.73 in the basal ganglia and 0.63 in the white matter, showing strong agreement with manual tracings.

Compared to SegCSVD, DORES had lower absolute volume and element differences, indicating better accuracy.

Abstract

Enlarged perivascular spaces (ePVS) are a marker of cerebral small vessel disease (SVD) detectable on magnetic resonance imaging (MRI). Due to their small size, diffuse distribution, and similar appearance to other markers of SVD, ePVS are inherently difficult to quantify on MRI, resulting in a lack of reliable quantification methods. We comprehensively evaluated and tested several automated ePVS segmentation algorithms by comparing their detection and segmentation performances to gold standard manual tracings.

Vanderbilt Memory and Aging Project participants (n = 30, 72.2±8.8 years, 47% female) underwent T1‐weighted, T2‐weighted, and fluid‐attenuated inversion recovery 3T MRI, and were manually traced by a neuroradiologist to measure ePVS burden. Five algorithms were used to segment ePVS on all participants: BigrBrain, SHIVA‐T1, MAPS‐T1, SegCSVD, and a locally developed nnU‐Net, Detection and Output of enlaRged pErivascular Spaces (DORES). To assess algorithm detection and segmentation capabilities, the Dice score, F1 score, absolute volume difference (AVD), and absolute element difference (AED) were calculated and compared against the expert manual rater separately for the basal ganglia and white matter.

The algorithm with the highest overall performance was DORES, followed by SegCSVD. In the basal ganglia, the DORES achieved a Dice score of 0.73±0.09, F1 score of 0.72±0.10, AVD of 88±120, and AED of 29±33. SegCSVD, by contrast, achieved a Dice score of 0.56±0.09, F1 score of 0.52±0.08, AVD of 109±58, and AED of 52±53. In the white matter, DORES achieved a Dice score of 0.63±0.17, F1 score of 0.68±0.16, AVD of 327±629, and AED of 68±79. By comparison, SegCSVD achieved a Dice score of 0.39±0.18, an F1 score of 0.39±0.18, an AVD of 380±643, and an AED of 135±127 in the white matter.

We present a new deep‐learning algorithm, DORES, as an effective tool for segmenting ePVS in our cohort. Compared to other algorithms, DORES demonstrated higher accuracy, agreement, and validity across metrics of detection, segmentation, and correlations with manual tracings in both basal ganglia and white matter.