Abstract

Abstract

Deep learning has soon taken medical image processing to the cutting-edge, with state-of-the-art performance in classification, segmentation, and anomaly detection. Convolutional neural networks (CNNs) led early breakthroughs, then generative adversarial networks (GANs) for data augmentation and super-resolution, and vision transformers (ViTs) and self-supervised learning (SSL) more recently for global context modeling and label-efficient training. Federated learning (FL) has become a privacy-preserving framework for multi-institutional collaboration. In spite of these developments, translation to clinical practice continues to be limited by issues of interpretability, data variability, regulatory affairs, and ethical review.

This review offers a critical integration of 2024–2025 advances in medical imaging deep learning, organized along a three-axis taxonomy: (1) architectural innovation, (2) paradigms for training, and (3) integration with clinical practice. In contrast to previous surveys, quantitative performance benchmarks are associate with particular datasets, compare explainable AI (XAI) tools to the criterion of clinical usability, and place technical advancement within the contemporary debates over regulation and ethics, such as the EU AI Act (2024) and FDA developments.

By taking advantage of the synergy of technology innovations and translation thinking, this overview delineates current research challenges and lists directions—such as multimodal data fusion, edge-based and light-weighted architecture, standardized benchmarks, and bias auditing—needed to enable the equitable, safe, and scaleable deployment of AI for healthcare imaging.

Statistics