Multimodal brain image fusion for neurological disease detection and classification: a comprehensive review

The fast development of machine learning (ML) and deep learning (DL) has significantly impacted the automated methods of brain disease detection and classification. Specifically, multimodal brain image fusion has been identified as a powerful tool for the combination of complementary structural, functional, metabolic, and clinical information from a variety of modalities including MRI, PET, CT, EEG, MEG, genetic, and neuropsychological studies. This review offers a thorough analysis of the traditional image fusion methods, classical ML models, CNN-based models, 3D DL models, attention-driven and Transformer-based models, as well as recent generative and incomplete multimodal learning models. The review also discusses publicly available datasets, evaluation metrics, and performance comparisons in detail to highlight trends and limitations. Multimodal fusion often improves performance, but it can fail when the modality is misaligned or noisy, especially for Alzheimer's disease (AD), Parkinson's disease (PD), Brain Tumor (BT), epilepsy, and stroke diagnosis. In addition, the review critically examines the current challenges of modality alignment, data unavailability, computational cost, cross-center generalization, and clinical applicability. The review concludes by pointing out the open research areas for the development of robust, interpretable, scalable, and clinically applicable multimodal DL models for neurological disorder diagnosis.