Accuracy and reproducibility of automated white matter hyperintensities segmentation with lesion segmentation tool: A European multi-site 3T study

Brain vascular damage accumulate in aging and often manifest as white matter hyperintensities (WMHs) on MRI. Despite increased interest in automated methods to segment WMHs, a gold standard has not been achieved and their longitudinal reproducibility has been poorly investigated. The aim of present work is to evaluate accuracy and reproducibility of two freely available segmentation algorithms. A harmonized MRI protocol was implemented in 3T-scanners across 13 European sites, each scanning five volunteers twice (test-retest) using 2D-FLAIR. Automated segmentation was performed using Lesion segmentation tool algorithms (LST): the Lesion growth algorithm (LGA) in SPM8 and 12 and the Lesion prediction algorithm (LPA). To assess reproducibility, we applied the LST longitudinal pipeline to the LGA and LPA outputs for both the test and retest scans. We evaluated volumetric and spatial accuracy comparing LGA and LPA with manual tracing, and for reproducibility the test versus retest. Median volume difference between automated WMH and manual segmentations (mL) was -0.22[IQR = 0.50] for LGA-SPM8, -0.12[0.57] for LGA-SPM12, -0.09[0.53] for LPA, while the spatial accuracy (Dice Coefficient) was 0.29[0.31], 0.33[0.26] and 0.41[0.23], respectively. The reproducibility analysis showed a median reproducibility error of 20%[IQR = 41] for LGA-SPM8, 14% [31] for LGA-SPM12 and 10% [27] with the LPA cross-sectional pipeline. Applying the LST longitudinal pipeline, the reproducibility errors were considerably reduced (LGA: 0%[IQR = 0], p < 0.001; LPA: 0% [3], p < 0.001) compared to those derived using the cross-sectional algorithms. The DC using the longitudinal pipeline was excellent (median = 1) for LGA [IQR = 0] and LPA [0.02]. LST algorithms showed moderate accuracy and good reproducibility. Therefore, it can be used as a reliable cross-sectional and longitudinal tool in multi-site studies.

Training labels for hippocampal segmentation based on the EADC-ADNI harmonized hippocampal protocol

BACKGROUND

The European Alzheimer's Disease Consortium and Alzheimer's Disease Neuroimaging Initiative (ADNI) Harmonized Protocol (HarP) is a Delphi definition of manual hippocampal segmentation from magnetic resonance imaging (MRI) that can be used as the standard of truth to train new tracers, and to validate automated segmentation algorithms. Training requires large and representative data sets of segmented hippocampi. This work aims to produce a set of HarP labels for the proper training and certification of tracers and algorithms.

METHODS

Sixty-eight 1.5 T and 67 3 T volumetric structural ADNI scans from different subjects, balanced by age, medial temporal atrophy, and scanner manufacturer, were segmented by five qualified HarP tracers whose absolute interrater intraclass correlation coefficients were 0.953 and 0.975 (left and right). Labels were validated as HarP compliant through centralized quality check and correction.

RESULTS

Hippocampal volumes (mm(3)) were as follows: controls: left = 3060 (standard deviation [SD], 502), right = 3120 (SD, 897); mild cognitive impairment (MCI): left = 2596 (SD, 447), right = 2686 (SD, 473); and Alzheimer's disease (AD): left = 2301 (SD, 492), right = 2445 (SD, 525). Volumes significantly correlated with atrophy severity at Scheltens' scale (Spearman's ρ = <-0.468, P = <.0005). Cerebrospinal fluid spaces (mm(3)) were as follows: controls: left = 23 (32), right = 25 (25); MCI: left = 15 (13), right = 22 (16); and AD: left = 11 (13), right = 20 (25). Five subjects (3.7%) presented with unusual anatomy.

CONCLUSIONS

This work provides reference hippocampal labels for the training and certification of automated segmentation algorithms. The publicly released labels will allow the widespread implementation of the standard segmentation protocol.