A Systematic Review of Multimodal Deep-Learning Frameworks for Medical-Image Classification: Data Fusion, Augmentation, and Architectural Trends (2020–2026)

Background: Multimodal deep-learning systems can combine complementary information from different medical-imaging modalities, but their reported performance is difficult to compare because studies use different datasets, validation strategies, fusion mechanisms, and outcome definitions. Objective: This review examines recent deep-learning approaches for medical-image classification, emphasizing multimodal fusion, transformer-based architectures, data augmentation, computational burden, interpretability, and generalizability. Methods: The source manuscript searched Google Scholar, IEEE Xplore, ScienceDirect, PubMed, and SpringerLink for studies published between 2020 and 2026. The reported selection process identified 83 records, removed 18 duplicates, screened 65 records, assessed 35 full texts, and included 21 studies. Because the source manuscript contains inconsistent denominators and heterogeneous outcome definitions, this revision reports findings narratively rather than presenting an unverified pooled estimate. Results: The included literature covers MRI, CT, chest radiography, and other multimodal settings. The studies describe CNN, transformer, hybrid, ensemble, self-supervised, and GAN-based approaches. Multimodal and transformer-based systems frequently report strong performance, but differences in datasets, patient-level separation, external validation, class balance, and reporting of computational cost and interpretability limit comparisons. Conclusions: Multimodal deep learning is promising, but current evidence does not establish universal superiority over unimodal models. Future studies should use independent patient-level test sets, external multicenter validation, transparent reporting, clinically meaningful explainability assessment, and standardized computational benchmarks.