Adversarial Detection and Defense for Medical Ultrasound Images: From a Frequency Perspective

B-mode ultrasound imaging is a popular medical imaging technique, and deep neural networks (DNNs), like other image processing tasks, have become popular for the analysis of B-mode ultrasound images. However, a recent study has demonstrated that medical deep learning systems can be compromised by carefully-engineered adversarial attacks with small imperceptible perturbations. Therefore, the adversarial attacks of ultrasound images are analyzed from a frequency perspective for the first time, and an easy-to-use adversarial example defense method is proposed, which can be generalized to different attack methods with no retraining need. Extensive experiments on two publicly-available ultrasound datasets, i.e., Breast ultrasound and Thyroid ultrasound datasets, have demonstrated that the hereby proposed defense method can detect the adversarial attacks of ultrasound images with a high mean accuracy.