EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Adversarial robustness of deep learning-based thyroid nodule segmentation in ultrasound

Nicholas Dietrich iD, David McShannon iD

PLOS Digital Health · 04.11.2026 · 5 (11) · e0001797

DOI: 10.1371/journal.pdig.0001797 ↗

Kurz gefasst

Fachartikel in der Zeitschrift PLOS Digital Health, online erschienen am 04.11.2026. Verfasst von Nicholas Dietrich und 1 weiteren Autor:innen. Beteiligte Einrichtung: Temerty Faculty of Medicine, University of Toronto, Toronto, Ontario, Canada. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Deep learning-based segmentation models are increasingly integrated into clinical imaging workflows, yet their robustness to adversarial perturbations remains incompletely characterized, particularly for ultrasound images. We evaluated adversarial attacks and inference-time defenses for thyroid nodule segmentation in B-mode ultrasound. Two black-box adversarial attacks were developed: (1) Structured Speckle Amplification Attack (SSAA), which injects boundary-targeted noise, and (2) Frequency-Domain Ultrasound Attack (FDUA), which applies bandpass-filtered phase perturbations in the Fourier domain. Three inference-time mitigations were evaluated on adversarial images: randomized preprocessing with test-time augmentation, deterministic input denoising, and stochastic ensemble inference with consistency-aware aggregation. Experiments were conducted on a U-Net segmentation model trained on cine-clips from a database of 192 thyroid nodules. The baseline model achieved a mean Dice similarity coefficient (DSC) of 0.76 (SD 0.20) on unperturbed images. SSAA reduced DSC by 0.29 (SD 0.20), from 0.76 to 0.47, while maintaining high visual similarity (SSIM = 0.94). FDUA resulted in a smaller DSC reduction of 0.11 (SD 0.09), from 0.76 to 0.65, with lower visual fidelity (SSIM = 0.82). Against SSAA, all three defenses significantly improved DSC after correction, with deterministic denoising showing the largest recovery (+0.10, p < 0.001), followed by randomized preprocessing (+0.09, p < 0.001), and stochastic ensemble inference (+0.08, p = 0.002). No defense achieved statistically significant improvement against FDUA, although recall consistently improved at the cost of reduced precision. The tested inference-time defenses therefore partially mitigated spatial-domain perturbations, whereas statistically significant DSC recovery was not demonstrated against frequency-domain perturbations, highlighting attack-specific challenges in adversarial robustness evaluation.

Autor:innen und Institutionen

  • Temerty Faculty of Medicine, University of Toronto, Toronto, Ontario, Canada.
  • Faculty of Engineering, McMaster University, Hamilton, Ontario, Canada.

ORCID-Profile: Nicholas Dietrich, David McShannon

Zitieren

APA

Dietrich, N., & McShannon, D. (2026). Adversarial robustness of deep learning-based thyroid nodule segmentation in ultrasound. PLOS Digital Health, 5(11), e0001797. https://doi.org/10.1371/journal.pdig.0001797

Vancouver

Dietrich N, McShannon D. Adversarial robustness of deep learning-based thyroid nodule segmentation in ultrasound. PLOS Digital Health. 2026;5(11):e0001797. doi:10.1371/journal.pdig.0001797

BibTeX
@article{Dietrich2026,
  author = {Dietrich, Nicholas and McShannon, David},
  title = {Adversarial robustness of deep learning-based thyroid nodule segmentation in ultrasound},
  journal = {PLOS Digital Health},
  year = {2026},
  volume = {5},
  number = {11},
  pages = {e0001797},
  doi = {10.1371/journal.pdig.0001797}
}
Kostenlos starten →