% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Gicquel:283099,
author = {Gicquel, Malo and Flood, Gabrielle and Zhao, Ruoyi and
Wuestefeld, Anika and Spotorno, Nicola and Strandberg, Olof
and Xiao, Yu and Åström, Kalle and Wisse, Laura E M and
van Westen, Danielle and Berron, David and Hansson, Oskar
and Vogel, Jacob W},
title = {{AI} {S}uperresolution: {C}onverting {T}1‐weighted {MRI}
from 3{T} to 7{T} resolution toward enhanced imaging
biomarkers for {A}lzheimer’s disease},
journal = {Alzheimer's and dementia},
volume = {21},
number = {Suppl 8},
issn = {1552-5260},
reportid = {DZNE-2025-01506},
pages = {e109817},
year = {2025},
abstract = {High-resolution (7T) MRI facilitates in vivo imaging of
fine anatomical structures selectively affected in
Alzheimer's disease (AD), including medial temporal lobe
subregions. However, 7T data is challenging to acquire and
largely unavailable in clinical settings. Here, we use deep
learning to synthesize 7T resolution T1-weighted MRI images
from lower-resolution (3T) images.Paired 7T and 3T
T1-weighted images were acquired from 178 participants (134
clinically unimpaired, 48 impaired) from the Swedish
BioFINDER-2 study. To synthesize 7T-resolution images from
3T images, we trained two models: a specialized U-Net, and a
U-Net mixed with a generative adversarial network
(U-Net-GAN) on $80\%$ of the data. We evaluated model
performance on the remaining $20\%,$ compared to models from
the literature (V-Net, WATNet), using image-based
performance metrics and by surveying five blinded MRI
professionals based on subjective quality. For n = 11
participants, amygdalae were automatically segmented with
FastSurfer on 3T and synthetic-7T images, and compared to a
manually segmented 'ground truth'. To assess downstream
performance, FastSurfer was run on n = 3,168 triplets of
matched 3T and AI-generated synthetic-7T images, and a
multi-class random forest model classifying clinical
diagnosis was trained on both datasets.Synthetic-7T images
were generated for images in the test set (Figure 1A). Image
metrics suggested the U-Net as the top performing model
(Figure 1B), though blinded experts qualitatively rated the
GAN-U-Net as the best looking images, exceeding even real 7T
images (Figure 1C). Automated segmentations of amygdalae
from the synthetic GAN-U-Net model were more similar to
manually segmented amygdalae, compared to the original 3T
they were synthesized from, in 9/11 images (Figure 2).
Classification obtained modest performance
$(accuracy∼60\%)$ but did not differ across real or
synthetic images (Figure 3A). Synthetic image models used
slightly different features for classification (Figure
3B).Synthetic T1-weighted images approaching 7T resolution
can be generated from 3T images, which may improve image
quality and segmentation, without compromising performance
in downstream tasks. This approach holds promise for better
measurement of deep cortical or subcortical structures
relevant to AD. Work is ongoing toward improving
performance, generalizability and clinical utility.},
month = {Jul},
date = {2025-07-27},
organization = {Alzheimer’s Association
International Conference, Toronto
(Canada), 27 Jul 2025 - 31 Jul 2025},
keywords = {Humans / Alzheimer Disease: diagnostic imaging / Magnetic
Resonance Imaging: methods / Female / Male / Aged / Deep
Learning / Neuroimaging: methods / Image Processing,
Computer-Assisted: methods / Brain: diagnostic imaging /
Sweden},
cin = {AG Berron},
ddc = {610},
cid = {I:(DE-2719)5000070},
pnm = {353 - Clinical and Health Care Research (POF4-353)},
pid = {G:(DE-HGF)POF4-353},
typ = {PUB:(DE-HGF)1 / PUB:(DE-HGF)16},
pubmed = {pmid:41434097},
pmc = {pmc:PMC12725268},
doi = {10.1002/alz70862_109817},
url = {https://pub.dzne.de/record/283099},
}