Deep learning segmentation models are increasingly integrated into clinical imaging workflows, yet their robustness against adversarial perturbations remains incompletely characterized, particularly for ultrasound images. In a new study, researchers Nicholas Dietrich and David McShannon evaluated black-box adversarial attacks and inference-time defenses for thyroid nodule segmentation in B-mode ultrasound. The authors investigated the effects of two newly developed black-box attack methods and assessed three distinct inference-time mitigation techniques designed to protect segmentation performance against these targeted disruptions.
Experiments were conducted using a U-Net segmentation architecture trained on cine-clips drawn from a database of 192 thyroid nodules. Under baseline conditions evaluating unperturbed B-mode ultrasound images, the trained U-Net model achieved a mean Dice similarity coefficient of 0.76 with a standard deviation of 0.20. This baseline measurement served as the reference standard for quantifying performance degradation induced by adversarial perturbations and evaluating the mitigation potential of subsequent defense mechanisms.
To evaluate model vulnerability, the investigators developed two black-box adversarial attack strategies. The first technique, designated the Structured Speckle Amplification Attack, injects boundary-targeted noise into the images. This attack reduced the mean Dice similarity coefficient by 0.29, with a standard deviation of 0.20, dropping the segmentation metric from 0.76 down to 0.47 while maintaining high visual similarity at a structural similarity index measure of 0.94. The second technique, termed the Frequency-Domain Ultrasound Attack, applies bandpass-filtered phase perturbations in the Fourier domain. This frequency perturbation caused a smaller reduction in Dice similarity coefficient of 0.11, with a standard deviation of 0.09, shifting the score from 0.76 to 0.65, while exhibiting lower visual fidelity with a structural similarity index measure of 0.82.
To counteract these adversarial images, three inference-time mitigations were tested: randomized preprocessing with test-time augmentation, deterministic input denoising, and stochastic ensemble inference with consistency-aware aggregation. Against the Structured Speckle Amplification Attack, all three defenses demonstrated statistically significant improvements in Dice similarity coefficient after correction. Deterministic input denoising produced the largest recovery, achieving an increase of 0.10 (p < 0.001). Randomized preprocessing followed with a recovery of 0.09 (p < 0.001), while stochastic ensemble inference provided a recovery of 0.08 (p = 0.002).
In contrast, no defense achieved a statistically significant improvement in Dice similarity coefficient against the Frequency-Domain Ultrasound Attack, although recall consistently improved across evaluations at the cost of reduced precision. The experimental findings indicate that the tested inference-time defenses partially mitigated spatial-domain perturbations, whereas statistically significant recovery was not demonstrated against frequency-domain perturbations. The authors conclude that these differing defense responses highlight attack-specific challenges in evaluating adversarial robustness and securing deep learning segmentation models in clinical ultrasound workflows.
Subject of Research: Adversarial robustness of ultrasound thyroid nodule segmentation models
Article Title: Adversarial robustness of deep learning-based thyroid nodule segmentation in ultrasound
Article References: 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
Image Credits: AI Generated
DOI: 10.1371/journal.pdig.0001797
Keywords: thyroid nodule segmentation, deep learning, ultrasound imaging, adversarial robustness, black-box attacks, inference-time defenses, U-Net, medical image analysis, Adversarial, robustness, deep, learning-based
News Source: Ophelia Keating. (November 8, 2026). Defenses Only Partially Protect Thyroid Ultrasound Segmentation From Adversarial Attacks. Scienmag.



