Introduction

Quantitative analysis of magnetic resonance imaging (MRI) data has become useful in both research and clinical studies, and is now crucial for the diagnosis and disease monitoring of patients with multiple sclerosis (MS). Indeed, MS is characterized by the presence of white matter (WM) lesions, detectable on MRI scans as hyper-intense areas on T2-weighed Fluid Attenuated Inversion Recovery (FLAIR) images. However, manual segmentation of such lesions is highly time-consuming and prone to significant intra- and inter-rater variability that makes wider application of MRI analysis and longitudinal follow-up still hardly feasible. Another challenge of lesion longitudinal follow-up is linked to the inconsistency of their patterns of evolution: their location and size can change significantly between two timepoints for the same subject. Consequently, the cross-segmentation of scans without reference to the preceding timepoints can lead to errors in MS lesion detection. Here, we first aimed at optimizing and validating a fully automated segmentation algorithm on large samples with a wide range of lesion loads. We then introduced a new tool for the longitudinal follow-up of MS lesions. Both algorithms are included in QyScore, a software that provides automated volumetric measurements of brain structures.

Materials and Methods

Longitudinal MRI brain scans from 53 patients (26 Clinically Isolated Syndrome (CIS) and 27 Relapsing-Remitting (RRMS) patients, average age 39.03 ± 11.35 y.o) were acquired using 3T scanners (48 on a Philips Achieva and five on a GE MR Discovery) with two timepoints one year apart. Manual segmentation was performed by two experts to create a Gold Standard (GS) with volumes ranging from 0.58 to 66.62 mL (mean = 8.31mL ± 11.38). Fully automated segmentation was obtained with an optimized version of White matter Hyperintensities Automatic Segmentation Algorithm (WHASA) [1], a method to automatically segment WM hyperintensities from T2-FLAIR and T1 images which relies on the coupling of non-linear diffusion and watershed segmentation, with intensity and location constraints. The longitudinal analysis performed afterwards required a fast and efficient T2-FLAIR intensity normalization method. To do so, we adapted the Standardization of Intensities (STI) [2] method to FLAIR data with WM lesions: STI uses joint intensity histograms in a common space to determine spatial intensity correspondences between two acquisitions. The difference between T2-FLAIR intensities from these scans was therefore computed, then masked by the lesion information obtained from WHASA (Figure 1). Longitudinal results can be displayed as labeled regions of increased or decreased lesion volumes, or as a color-scaled map showing intensity variations within lesions. As the longitudinal analysis is heavily related to the cross-sectional segmentation, performance was evaluated on WHASA results by computing indices relative to the GS, including Dice coefficient (DC) [3] and Absolute Volume Error difference in mL (AVE = |V_GS - V_WHASA|) for the whole dataset and then for patients categorized in different subsets according to their lesion loads (very low, low, medium and high, described in Figure 2)

Results

As presented in Figure 2, QyScore performed well on low lesion loads and remained accurate on very low lesion loads, while maintaining its accuracy on medium and higher lesion loads. Figure 3 shows a linear regression between manual and automated quantification in order to visualize the volumetric agreement: we found an excellent agreement (R² = 0.97) between manual and automated lesion volume for the whole dataset. Figure 4 shows the color-scaled maps resulting from the longitudinal analysis of WM lesions evolution. The color gradient illustrates the magnitude of lesion evolution, derived from intensity differences: increase (from red to yellow), decrease (from green to blue) and static (no color).

Discussion and conclusion

Compared to recent studies [4, 5], the proposed automated lesion segmentation method provides similar results for medium and high lesion loads and more accurate lesion segmentation for lower lesion loads. We have also shown that our longitudinal characterization enables visual assessment of lesion evolution. It could be, thus, a valuable tool for clinical routine and clinical trials, since it provides reliable, reproducible and robust quantification of lesions within minutes. With these algorithms, QyScore will deliver better comparisons and analysis of MS evolution, even in the early phases of the disease when the lesions load is low. A further validation of the lesions segmentation method will be to assess its performance on (a) two additional MS clinical phenotypes (Primary Progressive and Secondary Progressive) and (b) a larger variety of MR scanners (different manufacturers and models). The latest will also be useful for the longitudinal method validation, since the subject will be scanned at two different timepoints on several scanners sets.

Acknowledgements

The data have been provided by CHU-Bordeaux and we thank the team for their help.