Deep learning strategies for multimodal image fusion in neuroimaging

Neuroimaging has become a cornerstone of modern neuroscience, enabling detailed investigation of brain structure, function, and pathology. With advances in imaging modalities such as magnetic resonance imaging (MRI), functional MRI (fMRI), positron emission tomography (PET), and magneto- and electroencephalography (M/EEG), researchers now gather complementary data capturing diverse aspects of brain physiology. However, combining these heterogeneous sources into unified, meaningful representations remains a critical challenge due to differences in spatial resolution, contrast mechanisms, and signal characteristics.Recent advances in artificial intelligence (AI), particularly deep learning, have opened new opportunities for multimodal neuroimaging fusion by enabling automated feature extraction, hierarchical representation learning, and complex cross-modal integration. These approaches have demonstrated considerable potential for improving disease characterization, early diagnosis, treatment planning, and biomarker discovery in neurological and neuropsychiatric disorders. Yet a central limitation persists: deep learning models — and multimodal models in particular — often behave as opaque “black boxes,” offering predictions without revealing the neurobiological signals that drive them. Alongside challenges of robustness, reproducibility, data harmonization, and generalizability, this lack of transparency is one of the field’s most pressing barriers to clinical and scientific adoption.This Research Topic aims to highlight state-of-the-art deep learning methodologies and emerging AI paradigms for effective multimodal image fusion in neuroimaging, with a particular emphasis on visualization and explainability as a path to mechanistic insight. The objective is to identify new architectures, training strategies, and validation pipelines that maximize data complementarity while preserving — and actively surfacing — biological interpretability. We especially encourage contributions that go beyond predictive performance to show what fused representations encode, where in the brain the signal originates, and how it may relate to putative underlying mechanisms of disease. Researchers are invited to investigate how novel deep learning frameworks can improve diagnostic accuracy, biomarker identification, and understanding of neurological and neuropsychiatric disorders, and to explore the trade-offs between model complexity, computational cost, and clinical applicability.Emphasis will be placed on clinically translatable and explainable AI approaches that bridge the gap between advanced computational models and real-world neuroimaging applications. We are particularly interested in work that uses visualization, attribution, and interpretability tools — such as saliency and attention maps, concept-based explanations, counterfactual analyses, latent-space probing, and mechanistic interpretability — to generate testable neuroscientific hypotheses from multimodal fused representations. We also welcome studies addressing the balance between model complexity, computational efficiency, scalability, and deployment in clinical environments. Key questions include:• How can multimodal fusion models achieve robust performance across heterogeneous acquisition protocols and imaging centers?• How can prior neuroscientific and clinical knowledge be effectively incorporated into deep learning architectures?• How can fused representations be made interpretable, explainable, and trustworthy for clinicians and researchers — and used to reveal putative underlying mechanisms of brain function and disease?• Which visualization and explainability techniques best expose the cross-modal interactions that drive model predictions?• What strategies can improve reproducibility, fairness, and ethical deployment of AI-driven neuroimaging systems?To advance understanding of AI-driven multimodal neuroimaging integration, we welcome original research, reviews, and methodological contributions addressing, but not limited to, the following themes:• Novel deep learning architectures for multimodal image fusion, including transformers, graph neural networks (GNNs), multimodal variational autoencoders (VAEs), diffusion models, and foundation models• Cross-modal and cross-domain learning strategies, such as transfer learning, self-supervised learning, federated learning, and domain adaptation• Explainable and interpretable AI methods for fused neuroimaging representations, including visualization techniques (saliency, attention, activation, and attribution maps), concept- and prototype-based explanations, counterfactual reasoning, and mechanistic interpretability approaches that link model behavior to neurobiological substrates• Visualization frameworks that translate fused multimodal representations into clinically and scientifically meaningful insights about underlying disease mechanisms• Benchmark datasets, evaluation frameworks, and standardized metrics for multimodal fusion performance and explanation quality• Applications in neurological and psychiatric disorders, including Alzheimer’s disease, Parkinson’s disease, epilepsy, stroke, autism spectrum disorders, multiple sclerosis, depression, and schizophrenia• Fusion of structural, functional, metabolic, and electrophysiological imaging modalities• Real-time, lightweight, or resource-efficient AI frameworks for clinical translation and edge deployment• Data harmonization and robustness strategies for multi-center and longitudinal neuroimaging studies• Ethical, reproducibility, privacy, and fairness considerations in AI-based neuroimaging fusion• Clinical decision support systems and precision medicine applications enabled by multimodal neuroimaging AI Topic editors: Jürgen Dammers, Tie-Qiang Li Source: https://www.frontiersin.org/research-topics/81800/deep-learning-strategies-for-multimodal-image-fusion-in-neuroimaging