Abstract
The digital transformation of healthcare has accelerated the adoption of medical imaging systems interconnected through cloud platforms and hospital networks. While this connectivity enhances diagnostic accuracy and workflow efficiency, it also exposes imaging infrastructure to sophisticated cyber threats. Attacks such as ransomware, data manipulation, and unauthorized access jeopardize patient safety and the integrity of diagnostic outcomes. This review systematically examines recent ML and deep learning approaches employed for cyber threat detection in medical imaging systems. Key methodologies, including supervised anomaly detection, unsupervised clustering, hybrid ensemble techniques, and reinforcement learning-based intrusion detection frameworks, are analyzed. The study synthesizes findings from recent literature (2023-2025) across Scopus, IEEE Xplore, and SpringerLink, evaluating their detection accuracy, adaptability, and scalability in clinical environments. The review reveals that hybrid ML models, combining feature learning and adaptive optimization, outperform traditional classifiers in detecting zero-day attacks and abnormal traffic patterns. Deep neural architectures, particularly CNN-LSTM and autoencoder variants, demonstrate enhanced detection precision exceeding 95% in benchmark datasets. However, issues such as data imbalance, privacy preservation, explainability, and real-time deployment constraints remain major challenges. ML-driven cyber defense frameworks significantly advance the security posture of medical imaging systems. Also, to ensure resilient, transparent, and privacy-preserving threat detection solutions suitable for real-world healthcare settings.