Robustness of ultrasound deep beamformers using low-energy adversarial perturbations

Itamar Salazar-Reque, Andrés Coila, Roberto Lavarello

Producción científica: Capítulo del libro/informe/acta de congresoContribución a la conferenciarevisión exhaustiva

1 Cita (Scopus)

Resumen

As the use of ultrasound beamformers based on deep neural networks (DNNs) continues to gain popularity, it becomes increasingly important to assess their robustness. Traditional model-based beamformers are typically evaluated by studying factors such as sound speed variations, off-axis clutter, and other variables that challenge the underlying model. However, these conventional evaluation methods may not be well-suited for assessing model-free DNN-based beamformers. To address this challenge, we propose employing adversarial perturbations as a means of evaluating their robustness. Thus, we computed low-energy adversarial perturbations for two deep beamformers using the basic iterative method. We then assessed their performance when exposed to perturbed inputs using a contrast metric. Our results indicate that both of these deep beamformers are susceptible to these perturbations, whereas the traditional Delay-and-Sum (DAS) beamformer was little affected. Furthermore, one of the deep beamformers exhibited greater vulnerability to these perturbations, resulting in reduced performance compared to the other. Experimental results corroborated these findings, showing similar trends in phantom acquisitions. In summary, our findings emphasize the utility of adversarial perturbations as a valuable tool in assessing the robustness of deep neural network-based beamformers.

Idioma originalInglés
Título de la publicación alojadaIUS 2023 - IEEE International Ultrasonics Symposium, Proceedings
EditorialIEEE Computer Society
ISBN (versión digital)9798350346459
DOI
EstadoPublicada - 2023
Evento2023 IEEE International Ultrasonics Symposium, IUS 2023 - Montreal, Canadá
Duración: 3 set. 20238 set. 2023

Serie de la publicación

NombreIEEE International Ultrasonics Symposium, IUS
ISSN (versión impresa)1948-5719
ISSN (versión digital)1948-5727

Conferencia

Conferencia2023 IEEE International Ultrasonics Symposium, IUS 2023
País/TerritorioCanadá
CiudadMontreal
Período3/09/238/09/23

Huella

Profundice en los temas de investigación de 'Robustness of ultrasound deep beamformers using low-energy adversarial perturbations'. En conjunto forman una huella única.

Citar esto