Transformers and beyond: next-generation deep learning architectures in neuroimaging

Neuroimaging is an interdisciplinary field at the intersection of neuroscience, computer science, and biomedical engineering, aiming to decode the structure and function of the human brain through advanced computational techniques. Traditional machine learning and early deep learning architectures, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have significantly improved image classification, segmentation, and functional mapping tasks. However, despite their success, these methods are often limited in their ability to handle the complex spatial-temporal dependencies and multimodal information inherent in neuroimaging data. Recent breakthroughs in transformer architectures, originally developed for natural language processing, have shown remarkable potential in capturing global contextual information and learning from large, diverse datasets, positioning them as a powerful tool for next-generation brain analysis.This Research Topic aims to examine the emerging role of transformer-based, foundation, generative, and other next-generation deep learning models in neuroimaging. We welcome contributions that advance methodological innovation while also critically evaluating model validity, interpretability, robustness, and translational relevance. Key areas include the integration of spatial and temporal information; multimodal learning across structural MRI, diffusion MRI, functional MRI, PET, EEG, MEG, and clinical or molecular data; self-supervised and transfer-learning approaches; and the development of models that generalize across scanners, acquisition protocols, institutions, and patient populations. Particular emphasis will be placed on rigorous validation, comparison with appropriate conventional and deep-learning baselines, prevention of data leakage, uncertainty estimation, transparent reporting, and reproducible implementation. Studies addressing data scarcity, domain shift, bias, fairness, privacy, computational efficiency, and clinically meaningful evaluation are also encouraged. To gather further insights into the advances and challenges in deep learning for neuroimaging, we welcome articles addressing, but not limited to, the following themes:refine the topic list to:o Transformer and attention-based architectures for structural, diffusion, functional, molecular, and electrophysiological brain imaging o Self-supervised learning, foundation models, transfer learning, and generative models for brain data o Multimodal and cross-domain integration of imaging, clinical, cognitive, genomic, and molecular data o Longitudinal modeling, disease progression, prognosis, and treatment-response prediction o Explainability, uncertainty quantification, robustness, and biologically informed model interpretation o Clinical translation, biomarker discovery, and decision-support applications o Federated learning, privacy-preserving methods, and adaptation across sites, scanners, and acquisition protocols o Benchmarking, external validation, data-efficient learning, and comparison with appropriate baseline models o Bias, fairness, reproducibility, reporting standards, and responsible use of neuroimaging AI o Critical reviews, perspectives, and negative or null findings that clarify the limitations of next-generation architectures Topic editors: Veena A Nair, Nagesh Adluru, Burak Acar Source: https://www.frontiersin.org/research-topics/83899/transformers-and-beyond-next-generation-deep-learning-architectures-in-neuroimaging