Enhancing Robustness of Cancer Detection by Addressing Adversarial Attacks: Towards A Comprehensive Assessment on Medical Images

Despite medical and technological enhancements, the interpretation, identification, and cancer treatment on the basis of imaging data continue to be important issues. The deep learning-aided image classification techniques have been ensured to perform relatively well in image classification. But, in the modern days, the experts have displayed that adding minimal noise that is not noticeable by the experts can affect these techniques to produce inaccurate solutions. These adversarial examples have proved to be dangerous concerning the hamper launching and security measures of deep learning-assisted approaches in the practical world. Hence, this review assesses the potential of adversarial networks to address several key challenges of cancer domain and dataset shifts, imaging, data access, and privacy, including data scarcity and imbalance, data annotation and quantification, as well as cancer detection, tumor profiling, and treatment planning. The networks used for the detection of adversarial attacks for robust cancer detection that are analyzed in this research work. Moreover, the type of attacks is detected using the recently developed approaches that are effectively discussed in this research work. The computational experiments are carried out using chest X-ray images, histopathological images, mammogram images, and Computed Tomography (CT) images are reviewed. The effectiveness of various defense strategies is evaluated and summarized with respect to diverse performance measures. Finally, this work assesses the advantages and disadvantages based on their efficiency and then examine the future orientations of these deep learning models for medical image recognition.