Abstract Background and aims White matter hyperintensities (WMH) as part of the small vessel disease score are increasingly subjected to automated WMH segmentation to enhance objectivity and comparability across studies. In this study, we investigated whether a deep-learning based automated segmentation model, TrUE-Net performs better than 0. 85 thresholded BIANCA (Brain Intensity AbNormality Classification Algorithm), a k-nearest neighbour approach in identifying WMH lesions Methods We compared the robustness and performance of BIANCA₀. 85 vs. TrUE-Net using data of the prospective Berlin Long-term Observation of Vascular Events=BeLOVE, () cohort and the WMH Segmentation Challenge dataset () using manually delineated WMH masks as ground truth. We assessed the overall performance, precision, sensitivity using three datasets. One consisting of a mixed dataset (n=109), data from the BeLOVE cohort (n=59) with mixed vascular lesions, examined at a 3T Siemens and 3 T Philips scanner and data from the Challenge dataset (n=50) with a variety of MRI scanners, comparing their DICE score (DC) and accounting for factors such as lesion volumes Results TrUE-Net showed better performance, precision, sensitivity in all 3 datasets (DC: 0. 66±0. 19/0. 64±0. 22/0. 69±0. 15) than BIANCA₀. 85 (DC: 0. 5±0. 22/ 0. 57±0. 24/0. 41±0. 14) (figure1) and greater robustness towards different scanner types than BIANCA ₀. 85 (figure2). Out of the investigated factors (figure3), lesion volume differences had the most impact on the performance of TrUE-Net Conclusions TrUE-Net proves to be more accurate and robust in segmenting WMH than BIANCA₀. 85, even with different MRI scanners, making it more reliable for analysing large amounts of data, both cross-sectional and longitudinal. Conflict of interest nothing to disclose regarding the topic of the submitted abstract: Chiara-Sophia Hübotter, Uchralt Temuulen, Joachim E. Weber, Katharina Schönrath, Ira Rohrpasser-Napierkowski, Tuncer, Mehmet, Matthias Endres, Ivana Galinovic, Kersten Villringer Figure 1 - belongs to Results Figure 2 - belongs to Results Figure 3 - belongs to Results
Huebotter et al. (Fri,) studied this question.