Challenging cases for WMH segmentation comparatively processed by seven automated methods

Presented at

Clinical Neuroradiology, 31(Supplement 1), 40-41. (2021)


Aruci, M., Dünnwald, M., Schreiber, F., Sciarra, A., Maass, A., Schreiber, S., Oeltze-Jafra, S.



White matter hyperintensities of presumed vascular origin (WMH), a hallmark feature of cerebral small vessel disease (CSVD), are FLAIR/T2-hyperintense lesions that predict various clinical readouts, e.g., stroke or dementia [6, 8].WMH are commonly determined according to their volume load [1], while capturing or classifying more subtle features such as different WMH patterns or “WMH mimics” is demanding.We aimed to compare automatic methods to segment challenging WMH (e.g. multifocal spots, peri-basal ganglia WMH or WMH “mimics” surrounding lacunes/large hemorrhages) in clinical CSVD cases.

Materials & Methods

We applied seven different automatic WMH segmenting methods (LGA and LPA [7], SLS [5], MDbrain (Mediaire GmbH), BIANCA [2], FreeSurfer[4] and PGS – a Deep Learning approach [3]) in T1/FLAIR MRI sequences and compared their performance against gold standard manual segmentations in 10 CSVD patients with challenging WMH aiming to identify the most suitable method to segment them. Segmentation accuracy was determined through Dice similarity coefficient and other metrics measuring sensitivity or precision.


In our dataset, the PGS (DSC:0.6), LPA (DSC:0.59) and MDbrain (DSC:0.57) were superior in detecting periventricular and deep “multifocal spot WMH challenges” with a high sensitivity and precision rate. However, similarly to other methods, on a variable scale, they falsely segmented “WMH mimics” surrounding CSVD-related lesions as “true WMH” resulting in an overestimation of WMH volume.The volume of these “WMH mimics” segmentations is fluctuate on the method, but PGS performs better as it detects less false-positive-WMH-volume compared to other tools.

Boxplot for DICE similarity coefficient for different white matter lesion detection tools - mdbrain achieves best values (plotting 25% and 75% percentile, median and all individual values).


Future WMH segmentation tools will need to detect more accurately challenging WMH, as specific patterns that are highly relevant, to move forward in the understanding of different CSVD subtypes and pathophysiology.


These results show weakpoints of common WMH segmentation methods in complex FLAIR/T2-hyperintense lesions and can be insightful in choosing the best performing tool in segmenting challenging WMH.



[1] Charidimou, A., et al. Neurology. 86.6 (2016)

[2] Griffanti, L., et al. Neuroimage.141 (2016)

[3] Park,G., et al. Neuroimage. 237 (2021)

[4] Puonti, O., et al. Neuroimage. 143 (2016)

[5] Roura, E, et al.Neuroradiology. 57.10 (2015)

[6] Scheumann, V., et al. JNS. 419 (2020)

[7] Schmidt, P., et al.NeuroimageClin. 23 (2019)

[8] Wardlaw, J.M., et al. The Lancet. Neurology. 12.8 (2013)


*bald verfügbar 

mediaire GmbH

Möckernstrasse 63
10965 Berlin, Germany

P +49 30 286 490 67
F +49 30 286 490 66

© 2022 mediaire GmbH


Experience mdprostate, the first fully-automated AI-tool including longitudinal assessment for the evaluation of prostate MRIs.

mediaire GmbH

Möckernstrasse 63
10965 Berlin, Germany

P +49 30 286 490 67
F +49 30 286 490 66

© 2022 mediaire GmbH