Advances in Detecting and Mitigation Adversarial Manipulations in Medical Imaging: A Review

Abstract

Medical Image Analysis has seen remarkable improvements with Deep Learning (DL), which have helped in refining the accuracy of disease detection, diagnosis, and clinical decision support. But these models are still very vulnerable to Adversarial Attacks, in which subtle yet carefully engineered perturbations can lead to false predictions that are not easily visually noticeable, which poses serious concerns about the dependability and safety of Artificial Intelligence (AI) in healthcare. This review summarizes and discusses the latest developments in medical imaging application Adversarial Attacks, detection methods and defense. It describes Adversarial Attacks like evasion, poisoning, universal, physical, transfer-based, and multimodal attacks and reviews their effect on imaging modalities, including Magnetic Resonance Imaging (MRI), Computed Tomography (CT), X-ray, and ultrasound, as well as dermoscopy and histopathology. Besides, the review provides an overview of existing Detection Techniques based on different statistical approaches, feature representation, density estimation, reconstruction, and explainable AI and defense techniques such as adversarial training, feature squeezing, denoising networks, robust optimization, and trusted AI frameworks. Future challenges such as Federated Learning (FL), multimodal medical AI, model generalization, computational complexity, and clinical implementation are also addressed. Lastly, future research directions focused on robust, interpretable and clinically valid DL systems are presented to facilitate the safe use of AI in Medical Image Analysis.