Katalog Plus
Bibliothek der Frankfurt UAS
Bald neuer Katalog: sichern Sie sich schon vorab Ihre persönlichen Merklisten im Nutzerkonto: Anleitung.
Dieses Ergebnis aus BASE kann Gästen nicht angezeigt werden.  Login für vollen Zugriff.

A prior-sampling conditional variational autoencoder for neuroimaging normative modelling: Benchmarking deep learning against statistical approaches

Title: A prior-sampling conditional variational autoencoder for neuroimaging normative modelling: Benchmarking deep learning against statistical approaches
Authors: Ho, MP; Song, Y; Sachdev, PS; Fan, L; Jiang, J; Wen, W
Source: urn:ISSN:2837-6056 ; Imaging Neuroscience, 4, imag.a.1098
Publisher Information: MIT Press
Publication Year: 2026
Collection: UNSW Sydney (The University of New South Wales): UNSWorks
Subject Terms: 46 Information and Computing Sciences; 4905 Statistics; 49 Mathematical Sciences; Machine Learning and Artificial Intelligence; Biomedical Imaging; Hypertension; Behavioral and Social Science; Neurosciences; Bioengineering; Basic Behavioral and Social Science; Networking and Information Technology R&D (NITRD); Neurological; Generic health relevance; UK Biobank; conditional variational autoencoders; deep learning; normative modelling; precision psychiatry; anzsrc-for: 46 Information and Computing Sciences; anzsrc-for: 4905 Statistics; anzsrc-for: 49 Mathematical Sciences
Description: Normative modelling in neuroimaging provides a powerful framework for quantifying individual deviations from expected brain measures as a function of relevant covariates. While earlier methods focused on analysing distinct variables in isolation, an increasing number of deep learning-based approaches are emerging to handle multiple response variables simultaneously. Conditional variational autoencoders (cVAEs) have previously been applied in this context and show promise for multivariate modelling. However, existing inference methods still face challenges in providing reliable probabilistic predictions, limiting their effectiveness as true normative models. In this study, we propose an enhanced cVAE-based framework that generates predictions directly from covariates through prior-sampling inference. This approach aligns with normative modelling principles while leveraging deep learning to handle high-dimensional data. We demonstrate the effectiveness of our approach using 195 imaging-derived phenotypes (IDPs), including morphometric features (cortical thickness, cortical volume, subcortical volume) and white matter hyperintensity (WMH) volumes, as a test case. Our dataset includes 8,551 normotensive and 18,180 hypertensive participants from the UK Biobank. We benchmarked our model against three well-established normative modelling techniques, including Generalised Additive Models for Location, Scale, and Shape (GAMLSS), Multivariate Fractional Polynomial Regression (MFPR), and Hierarchical Bayesian Regression (HBR), as well as the conventional posterior-sampling cVAE approach employed in existing autoencoder-based normative models. Through comparative analyses, our results show that the proposed cVAE-based framework achieves performance comparable with well-established normative models across various metrics, while appropriately capturing individual deviations associated with hypertension severity. Furthermore, our inference strategy demonstrates superior covariate sensitivity compared with those used in ...
Document Type: article in journal/newspaper
Language: unknown
Relation: http://purl.org/au-research/grants/arc/FT190100197; https://hdl.handle.net/1959.4/106997; https://doi.org/10.1162/IMAG.a.1098
DOI: 10.1162/IMAG.a.1098
Availability: https://hdl.handle.net/1959.4/106997; https://doi.org/10.1162/IMAG.a.1098
Rights: open access ; https://purl.org/coar/access_right/c_abf2 ; CC-BY ; https://creativecommons.org/licenses/by/4.0/ ; CC BY
Accession Number: edsbas.E72F0A1A
Database: BASE