Lesion filling effect in regional brain volume estimations: a study in multiple sclerosis patients with low lesion load

Abstract

Introduction Regional brain volume estimation in multiple sclerosis (MS) patients is prone to error due to white matter lesions being erroneously segmented as grey matter. The Lesion Segmentation Toolbox (LST) is an automatic tool that estimates a lesion mask based on 3D T2-FLAIR images and then uses this mask to fill the structural MRI image. The goal of this study was (1) to test the LST for estimating white matter lesion volume in a cohort of MS patients using 2D T2-FLAIR images, and (2) to evaluate the performance of the optimized LST on image segmentation and the impact on the calculated grey matter fraction (GMF). Methods The study included 110 patients with a clinically isolated syndrome and 42 with a relapsing-remitting MS scanned on a 3.0-T MRI system. In a subset of consecutively selected patients, the lesion mask was semi-manually delineated over T2-FLAIR images. After establishing the optimized LST parameters, the corresponding regional fractions were calculated for the original, filled, and masked images. Results A high agreement (intraclass correlation coefficient (ICC) = 0.955) was found between the (optimized) LST and the semi-manual lesion volume estimations. The GMF was significantly smaller when lesions were masked (mean difference −0.603, p < 0.001) or when the LST filling technique was used (mean difference −0.598, p < 0.001), compared to the GMF obtained from the original image. Conclusion LST lesion volume calculation seems reliable. GMFs are significantly reduced when a method to correct the contribution of MS lesions is used, and it may have an impact in assessing GMF differences between clinical cohorts. Keywords

The Impact of MRI T1 Hypointense Brain Lesions on Cerebral Deep Gray Matter Volume Measures in Multiple Sclerosis

“…The second part of our study focused on differences between manually and automatically segmented/in‐painted images. Pareto and colleagues validated LST‐based lesion in‐painting using 2D T2 FLAIR and 3D MPRAGE images from patients with MS . Consistent with their results, we found high agreement between manually and automatically segmented/in‐painted images.…”

“…Since the number of patients with high lesion volume (typically, >15 mL) was not enough to sufficiently power a separate study of this subsample, our sample comprised a mix of patients with low, medium, and high lesion volume. Considering the results of previous studies suggesting that higher lesion volume may cause more bias in global volumetric measures, we investigated the relationship between lesion load and the resulting error in DGM structure volumes by running correlation analyses between lesion load and absolute differences of DGM volumes derived from non‐in‐painted and manually segmented/in‐painted images. P values of these correlations ranged between .09 (left amygdala) and .84 (right thalamus).…”

“…Therefore, understanding the impact of hypointense T1 lesions on brain tissue volume measurement in MS is essential. Several studies explored the effects of these lesions on global brain measures; however, literature on the impact of hypointense T1 lesions on DGM volumes is sparse. In two studies where the analyses were carried on images acquired from a 1.5 Tesla (T) scanner, researchers reported a direct effect of WM lesions on DGM volumetry measures .…”

Quantitative spinal cord MRI in radiologically isolated syndrome

“…Lesion maps were used to fill the segmented lesions in the T1-image with estimated healthy white matter (WM) tissue. 17 Gray matter (GM), WM, and CSF volumes were obtained using the Voxel-based morphometry 8 (VBM8) toolbox 18 for SPM8. Lesion, GM, WM, and CSF maps were reviewed for quality control by an experienced neuroradiologist (P.A.-L.).…”

ExploreASL: an image processing pipeline for multi-center ASL perfusion MRI studies

“…The presence of WMH can affect the GM/WM classification of T1w images in two ways: 1) WMH themselves can be incorrectly segmented as GM; 2) image intensities of WMH bias global modeling of GM and WM intensity distributions (Pareto et al 2016; Battaglini et al 2012). ExploreASL alleviates these complications by lesion-filling the T1w image before initiating the segmentation (Battaglini et al 2012): voxel intensities in the hypointense WMH regions on the T1w images are replaced by bias field-corrected values from the surrounding, normal-appearing WM (Chard et al 2010) (Figure 1).…”