Deep volumetric super‐resolution improves the detection of amyloid‐related cross‐sectional group differences in MCI: Neuroimaging / New imaging methods. (7th December 2020)

• Record Type: Journal Article
• Authors: Hesterman, Jacob; Avants, Brian; Greenblatt, Elliot; Tustison, Nicholas J.
• 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.039961
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Super‐resolution (SR) methodology has emerged as a practical tool that may improve the interpretability and reliability of clinical MRI [Zhao2019, Xie2019] for structural measurements. Despite promise in neurodegenerative disease populations, relatively little work has been performed to validate SR image analysis. We hypothesize that deep learning SR methodology will improve inference about the relationship between surrogate measurements of AD pathology and in vivo measurements of cortical thickness. We test this hypothesis in ADNI subjects with baseline measurements of AV45 PET and diagnosis of mild cognitive impairment. Method: We employ a 3D variant of the deep back projection network optimized with both distortion minimizing and perceptual losses [Haris2018]. The network was not optimized in any way for ADNI data. ANTsR and ANTsRNet [Muschelli2019] were used onT1w MRI for brain extraction, registration, anatomical labeling, bias correction, segmentation, and cortical thickness estimation. Segmentation and thickness estimation were done at the resolution of the native or SR image, respectively (Figure 1). Demographic data for the ADNI cohort (n=120) used in this study are shown in Figure 2. We interrogate p‐values of both the AV45 PET SUVR predictor and the data collection site predictor, enabling evaluation of the effect of SR on both predictors of interest and nuisance variables. We also control for age at baseline, gender and APOE status. Result: (1) The model relating APOE, gender, age and SUVR in left temporal lobe explained 42.9% of the variance using native resolution estimates compared to 47.1% for SR. All predictors were significant. The p‐value for SUVR was 5e‐4 for native resolution and 9e‐5 for SR. (2) MR sequence effect mitigated by super‐resolution (p = 6.7e‐7 ). (3) Main biological effects are improved (gender, APOE4, AV45, age) while nuisance (site and quality) effects are reduced. (4) SR thickness predicts amyloid level ‐ after controlling for low resolution estimate of thickness. Conclusion: Volumetric super‐resolution methodology may provide a useful preprocessing step in the structural analysis of neuroimages in MCI populations. Future work should evaluate impact on longitudinal processing, quantification of functional data, optimization for clinical signal detection, and additional validation efforts in other AD and non‐amyloid pathology.