Segmenting White Matter Hyperintensity in Alzheimer’s Disease using U-Net CNNs

• Kauê Tn Duarte University of Calgary / Foothills Medical Centre
• David G. Gobbi University of Calgary / Foothills Medical Centre
• Abhijot S. Sidhu University of Calgary / Foothills Medical Centre
• Cheryl R. McCreary University of Calgary / Foothills Medical Centre
• Feryal Saad University of Calgary
• Nita Das University of Calgary / Foothills Medical Centre
• Eric E. Smith University of Calgary
• Richard Frayne University of Calgary / Foothills Medical Centre

Resumo

White-matter hyperintensity (WMH) is associated with many disorders where it is suggestive of underlying cerebrovascular, small-vessel disease pathology. However, its role in Alzheimer’s disease (AD), mixed, and vascular dementia remains an open area of research. The fluid-attenuated inversion recovery (FLAIR) magnetic resonance (MR) imaging sequence is commonly used to visualize WMH because it provides good image contrast, not only between WMH and normal tissue, but also between WMH and cerebrospinal fluid. Manual segmentation of WMH lesions in brain volumes, on a slice-by-slice basis, is time-consuming with high inter-rater variability, however, this process remains the broadly accepted gold standard. In this study, variations of 2D, 2.5D and 3DU-shaped convolutional neural networks (U-Net CNNs) were used to perform semantic segmentation on FLAIR images. We evaluated these models in brain volumes obtained from 186 individuals from one of three disease classes: healthy normal (N = 94), mild cognitive impairment (N = 55), and AD (N = 37). Four common architectures (VGG16, VGG19, ResNet and EfficientNetBO) were employed as feature extractors. Results were assessed across the whole brain and by brain region (frontal, occipital, parietal, temporal lobes plus the insula) to identify differences in performance. In general, the predicted WMH volumes had an F-measure score >95% on the test data compared to manual segmentation. This work identified that 2. 5D with either VGG16 or VGG19 was the most suitable configuration when segmenting WMH. WMH segmentation performance and measured volume was found to vary between regions and disease classes. U-Net CNN architectures have good performance and may provide valuable insights about the white matter pathology.

Palavras-chave: Training, Pathology, Volume measurement, Manuals, Brain modeling, Feature extraction, Convolutional neural networks