Deep Learning Methods for Super-Resolution of Biomedical Images: A Brief Survey, Methodology, Application, and Challenge

Deep learning–based super-resolution reconstruction has become a rapidly growing research topic in radiology, nuclear medicine, and biomedical imaging, driven by the increasing demand for high-resolution image under hardware, radiation-dose, and acquisition-time constraints. Despite substantial progress and a rapidly expanding body of literature, existing studies remain fragmented, with limited systematic analysis of methodological principles, experimental reproducibility, and unresolved challenges across different imaging modalities. This survey provides a comprehensive and structured review of deep learning methods for biomedical image super-resolution. First, we summarize commonly used datasets, imaging modalities, data characteristics, and preprocessing pipelines, highlighting their influence on model design and evaluation. Second, from a methodological perspective, we categorize deep learning–based super-resolution approaches into convolutional neural networks, adversarial learning frameworks, and diffusion-based models, and analyze their underlying principles and representative architectures. Third, we reproduce and experimentally evaluate several classic super-resolution methods to identify their performance bottle-necks, generalization limitations, and modality-specific failures, and discuss corresponding improvement strategies. Finally, we outline open challenges and future research directions, including data scarcity, cross-modality generalization, uncertainty modeling, and clinical reliability. Overall, this survey aims to bridge the gap between rapidly evolving deep learning techniques and their practical deployment in biomedical image super-resolution, providing researchers and clinicians with a unified reference and actionable insights.