
%Aigaion2 BibTeX export from HES SO Valais Publications
%Thursday 19 November 2026 10:09:02 PM

@ARTICLE{,
    author = {Fatima, Ez-Zahraa Bazay and Martin, Barry and Widmer, Antoine and M{\"{u}}ller, Henning and Rachid, Jennane and glocker, ben and El Maliani, Ahmed Drissi},
  keywords = {Meta-learning {\textperiodcentered} Few-shot domain generalization {\textperiodcentered} Brain net work classification {\textperiodcentered} Multi-site neuroimaging {\textperiodcentered} Autism spectrum disorder},
     month = oct,
     title = {Few-Shot Cross-Site Domain Generalization for Multi-Site Autism Brain Network Classification},
   journal = {MICCAI Society},
      year = {2026},
      note = {Conf{\'{e}}rence MICCAI 2026 - 27.09-01.10.26 Strasbourg},
  abstract = {Multi-site functional Magnetic Resonance Imaging (fMRI)
datasets offer unprecedented opportunities for developing diagnostic biomark
ers for neuropsychiatric disorders such as Autism Spectrum Disorder
(ASD). However, site-induced distributional shifts across acquisition cen
ters severely compromise model generalizability, rendering classifiers trained
at one institution ineffective when deployed at others. In this work, we
propose a few-shot cross-site domain generalization framework based
on Model-Agnostic Meta-Learning (MAML) for multi-site Brain Net
work classification. By treating each acquisition site as a distinct meta
learning task, our framework meta-trains a linear classifier on top of a
pre-trained backbone, learning an initialization that rapidly adapts to
unseen sites using only k = 3 labeled support examples, achieving a
1.5\% improvement in AUC across different backbone architectures. In
comparison, achieving similar accuracy with classical approaches would
require up to 15 labeled examples, demonstrating that MAML provides
a strong site-generalizable initialization that enables rapid few-shot clas
sifier adaptation to each unseen site. Our method consistently outper
forms CORrelation ALignment (CORAL) and Meta-Learning for Do
main Generalization (MLDG) across multiple backbone architectures,
and performance gains further scale with the number of training sites.
This study shows the necessity of focusing research on inter-site variabil
ity for fMRI studies, as classical methods poorly generalize to new sites.
The implementation is available at https://github.com/fezz-bazay/
MAML-TTA-Brain-Network.}
}

