Automated hippocampal segmentation improved by convolutional neural network approach in participants with a history of cerebrovascular accident: Neuroimaging / New imaging methods. (7th December 2020)

• Record Type: Journal Article
• Authors: Zavaliangos‐Petropul, Artemis; Tubi, Meral A; Zhu, Alyssa; Haddad, Elizabeth; Jahanshad, Neda; Thompson, Paul M; Liew, Sook‐Lei
• Is Part Of: Alzheimer's & dementia. Volume 16(2020)Supplement 4
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-12-07
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.041634
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Cerebrovascular accidents (CVA) are risk factors for dementia, including Alzheimer's disease (AD; Arvanitakis et al., 2016). Hippocampal atrophy is observed in both CVA (Gemmel et al., 2012) and AD (Braak & Braak, 1991) patients. Thus, the hippocampus may be an important biomarker for CVA patients with high AD‐risk. FreeSurfer, a commonly used automated hippocampal segmentation method to study post‐CVA hippocampal volume (Khlif et al., 2019), uses an atlas‐based approach, making it more likely to fail in the presence of CVA lesions (Yang et al., 2016). Two recent convolutional neural network‐based (CNN) hippocampal segmentation algorithms, Hippodeep (Thyreau et al., 2018) and Hippmapp3r (Goubran et al., 2019), are more flexible for post‐CVA anatomical changes since they do not rely on a single atlas. However, they have not been widely tested in CVA populations. Here, we compare the segmentation accuracy of FreeSurfer, Hippodeep, and Hippmapp3r in participants with a history of CVA. Method: 30 T1‐weighted structural brain MRIs from CVA participants with varying lesion sizes (0.43‐181.5 cc) in the Anatomical Tracings of Lesions After Stroke dataset (Liew et al., 2018) were used to compare FreeSurfer, Hippodeep, and Hippmapp3r to manual segmentations. Accuracy was measured as spatial overlap between each automated and manual segmentation using the Dice Coefficient (DC). An ANOVA was used to test for DC differences across methods. Result: DC significantly differed by segmentation method (F‐value=178.3; p‐ value<2.10x ‐16 ). Ipsilesional results: Hippodeep DC was significantly greater than FreeSurfer DC ( p ‐value=1.03x10 ‐17 ; t ‐value=18.7) but not significantly different from Hippmapp3r DC. Contralesional results: Hippmapp3r DC was significantly greater than both Hippodeep ( p ‐value=9.53x10 ‐4 ; t ‐value=4.52) and FreeSurfer ( p ‐value=4.68x10 ‐21 ; t ‐value=24.8;Figure 1). Conclusion: Hippodeep and Hippmapp3r segmentations had significantly better accuracy than FreeSurfer, although all methods had a DC >0.7 (recommended threshold for good segmentation overlap; Zou et al., 2004). For ipsilesional segmentation accuracy, the two CNN methods did not significantly differ, while for contralesional segmentations, Hippmapp3r performed best. Future studies may benefit from using CNN approaches to more accurately estimate hippocampal volumes in patients with CVA with high AD risk.