Adversarial Perturbation Analysis for Robustness Testing of Deep Learning–Based Radiation Therapy Dose Prediction
Abstract
Purpose
Deep learning–based three‑dimensional (3D) dose prediction models are increasingly becoming viable for radiotherapy workflows. However, their sensitivity to small variations in input data remains insufficiently characterized. This study presents a framework to quantitatively evaluate the robustness of dose prediction by a 3D convolutional neural network for head‑and‑neck(HN) patients under controlled perturbations of CT image intensities.
Methods
A 3D dose prediction model was implemented using the Open Knowledge-Based Planning code and dataset of 340 HN patients (200 training, 40 validation, 100 testing). Perturbations were applied to CT intensity (Hounsfield units) using the Fast Gradient Sign Method with magnitudes ε = 0.001, 0.005, 0.01, and 0.05 (HU). Resulting dose predictions were compared with unperturbed outputs using standard planning target volume (PTV) metrics (D99, D95, D1) and organ‑at‑risk (OAR) metrics (D0.1cc, V30, V50). Dosimetric changes were evaluated as a function of perturbation magnitude.
Results
Even the smallest perturbation (ε = 0.001) produced measurable deviations in predicted dose, with PTV D95 decreasing by 1–2% and D1 increasing by 2–4%. Increasing perturbation levels led to progressive degradation in target coverage: D95 reductions reached 6–9% at ε = 0.01 and exceeded 15–20% at ε = 0.05. Corresponding increases in dose hotspots were observed. OAR doses demonstrated similar sensitivity, with spinal cord and larynx D0.1cc increasing 2–3% at ε = 0.001 and more than 25% at ε = 0.05. Volume‑based metrics also increased systematically, indicating reduced sparing robustness with higher perturbation magnitudes.
Conclusion
Small adversarial perturbations to CT input data can induce systematic and clinically meaningful changes in deep learning–predicted dose distributions. These findings emphasize the importance of incorporating robustness evaluation, uncertainty analysis, and perturbation‑based testing into the development and clinical assessment of deep learning based dose prediction models.