Deep learning pipeline for cortical gray matter segmentation and thickness analysis in Ultra High Resolution T2w 7 Tesla Ex vivo MRI across neurodegenerative diseases reveals associations with underlying neuropathology. (20th December 2022)

• Record Type: Journal Article
• Authors: Khandelwal, Pulkit; Sadaghiani, Shokufeh; Chung, Eunice; Lim, Sydney A; Duong, Michael Tran; Ravikumar, Sadhana; Arezoumandan, Sanaz; Peterson, Claire; Bedard, Madigan L; Capp, Noah; Ittyerah, Ranjit; Migdal, Elyse; Choi, Grace; Kopp, Emily; Patino, Bridget Loja; Hasan, Eusha; Li, Jiacheng; Prabhakaran, Karthik; Mizsei, Gabor; Gabrielyan, Marianna; Schuck, Theresa; Robinson, John; Ohm, Daniel T; Lee, Eddie B; Trojanowski, John Q; McMillan, Corey T; Grossman, Murray; Irwin, David J.; Tisdall, Dylan M; Das, Sandhitsu R.; Wisse, Laura EM; Wolk, David A.; Yushkevich, Paul A.;
• Is Part Of: Alzheimer's & dementia. Volume 18(2022)Supplement 5
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-12-20
• 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.065737
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Ex vivo magnetic resonance imaging (MRI) enables detailed characterization of neuroanatomy (Augustinack et al. 2013), such as hippocampal subfields in the medial temporal lobe (MTL) (Yushkevich et al. 2021, Ravikumar et al. 2021). However, automated cortical segmentation methods in ex vivo MRI are not well developed due to limited data availability and heterogeneity in scanners and acquisition. Here, we investigate a deep learning framework to parcellate the cortical mantle, compute thickness and link them with neuropathology ratings across 16 cortical regions in 7 Tesla MRIs of 38 ex vivo brain specimens spanning Alzheimer Disease and Related Dementias. Method: A deep learning method, nnU‐Net (Isensee et al. 2021), was trained on manually segmented 3D image patches (Figure 1C) to obtain automated cortical segmentations across 38 subjects (Table 1). We identified 16 landmarks (Figure 1A) for localized quantitative signatures of cortical morphometry and used the pipeline in Wisse et al. 2021 to measure local thickness (Figure 1B). Associations were computed between cortical thickness from manual and automated segmentations via Pearson's correlation and average fixed‐raters Intra‐class Correlation Coefficient (ICC) for 16 locations (Figure 3). We also correlated thickness from both automated and manual segmentations with neuropathological ratings of tau and neuronal loss in corresponding contralateral regions and global Braak staging (Figures 4 and 5). Result: Figure 2 depicts cortical mantle segmentation across brain hemispheres. Figure 3 shows good agreement between ground truth and automated thickness, with 15 regions with significant associations (p<0.05) and 8 regions having r>0.6. We observe high ICC scores with 9 regions where ICC>0.7, confirming that automated segmentations accurately measure thickness. Figure 4 shows significant correlations between thickness and Tau ratings for Brodmann Area 35 (BA35) and midfrontal regions and trends between neuronal loss and thickness in entorhinal cortex (ERC), anterior temporal pole and anterior insula. Figure 5 shows significant correlations between thickness and Braak staging in ventrolateral temporal cortex and ERC, with trends in other regions. Conclusion: Our automated ex vivo neuroimaging framework accurately segments the cortical mantle, provides thickness measurements that concur with user‐supervised thickness and links morphometry with underlying neurodegeneration, thus suggesting the strengths of ex vivo MRI.