Performance of five automated white matter hyperintensity segmentation methods in a multicenter dataset

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• White matter hyperintensities (WMH) are key markers of cerebral small vessel disease.
• Large, pooled multicenter datasets are crucial for studying WMH but challenge automated segmentation.
• Current evaluations of automated WMH segmentation methods lack multicenter validation.

Purpose of the Study:

• To evaluate the performance of five freely available, automated WMH segmentation methods in a multicenter setting.
• To compare the consistency of these methods across different MRI scanners and vendors.
• To provide guidance for selecting appropriate automated WMH segmentation tools for multicenter research.

Main Methods:

• WMH segmentation was performed on 60 patients across six different MRI scanners using five automated methods: Cascade, kNN-TTP, Lesion-TOADS, LST-LGA, and LST-LPA.
• Manual WMH segmentations served as the reference standard.
• Method performance was assessed using Dice's similarity coefficient (DSC) and intra-class correlation coefficient (ICC) for spatial and volumetric accuracy, respectively, both within and across scanners.

Main Results:

• The kNN-TTP method showed the best performance, exhibiting superior spatial and volumetric correspondence with manual segmentations across scanners.
• LST-LPA and LST-LGA also demonstrated good performance, outperforming Lesion-TOADS and Cascade.
• Performance variations were observed across different methods and scanner configurations, highlighting the impact of multicenter data.

Conclusions:

• kNN-TTP is a reliable method for automated WMH segmentation in multicenter studies.
• The study underscores the need for robust evaluation of automated segmentation consistency across diverse MRI environments.
• Findings can inform the selection of WMH segmentation tools, improving the reliability of multicenter neuroimaging research.