In MRI Diagnostic Systems using deep learning-based neuroimaging screening, it is a common practice
to use state-of-the-art neural networks designed initially for complex, large-scale natural image screening.
However, these networks, when used for brain imaging tasks, results in a sharp loss of performance efficiency
caused by use of an over-complex network on simple classification task; this decline of efficiency is particularly
marked under adversarial attacks. Indeed, these MRI diagnostic systems are much more vulnerable to adversarial
attacks than customary non-medical image analysis systems under attack. The research team observed that Deep learning-based
MRI diagnostic systems naively trained on neuroimaging scans do exhibit dramatic instability to small pixel-level
injected attacks, resulting in substantial decrease in diagnostic accuracy.
Accessibility Controls