PhaseAT: Fourier Phase Adversarial Training for Medical Image Domain Generalization
Medical imaging models often fail when deployed across scanners or hospitals because they overfit to domain-specific textures and appearance statistics. PhaseAT addresses this by adversarially perturbing the Fourier phase of the luminance channel while keeping the amplitude spectrum fixed, generating hard training views that stress spatial organization under matched style. By encouraging the model to ignore brittle amplitude-driven correlations and focus on stable geometric structure, PhaseAT achieves large gains on both single-source and multi-source domain generalization benchmarks.
PhaseAT is a creative and well-executed contribution to medical domain generalization. The core mechanism—phase-constrained adversarial perturbations in YCbCr luminance—is technically sound, supported by a thorough ablation study, and yields strong empirical gains on Camelyon17-WILDS and a multi-site diabetic retinopathy benchmark, especially in the challenging single-source regime. However, the scope is currently limited to 2D color images, and the paper lacks direct mechanistic evidence that the induced robustness specifically translates to reduced texture bias in the learned representations.
The ablation study in Table 3 compellingly validates the design choices: perturbing only phase while fixing amplitude outperforms both amplitude-only and joint phase-amplitude updates, and the phase-saliency mask is essential—without it average accuracy on Camelyon17 collapses from 85.7% to 56.5%. The method is also architecture-agnostic, improving ResNet-18 and ViT-Base alike, and exhibits a broad stable operating region around the chosen hyperparameters.
While the gains are promising, the paper does not demonstrate that adversarial phase noise actually disentangles texture from shape in the model's internal representations; it relies on the plausible but unproven assumption that phase perturbations serve as a faithful proxy for clinically relevant structure-preserving shifts. The headline claim of 'over 20% improvement in single-source domain generalization' is technically true for the C2 source on Camelyon17, but could mislead readers into thinking it reflects a uniform average gain across all settings rather than a best-case single-source scenario. Additionally, the theoretical justification for why fixing amplitude and perturbing phase specifically improves out-of-domain generalization—beyond generic adversarial regularization—remains informal.
PhaseAT is evaluated against a reasonable but narrow set of baselines (ERM, AdverIN, ConDiSR, RAS4DG) using the standard DomainBed protocol, and the reported paired t-test significance markers add credibility. That said, the omission of broader Fourier augmentations (e.g., FDA, MixStyle) and more recent DG methods makes it hard to position PhaseAT precisely within the landscape. The ImageNet-pretrained experiment also reveals diminishing returns: PhaseAT adds only about one point of accuracy when strong pretraining is used, suggesting its benefit may be most pronounced in limited-data or from-scratch training regimes.
The authors release code and describe most hyperparameters ($K=5$ inner steps, $\lambda=0.5$, $\eta=0.2$), but the phase perturbation budget $\epsilon$ is not reported, and key optimizer settings (learning rate, weight decay) are omitted. The custom Wirtinger-calculus gradient for phase updates (Eq. 2) is non-standard and would benefit from a reference implementation or pseudocode to avoid reproduction bugs. Moreover, training for only 30 epochs with a 5-epoch warm-up schedule is unusually short for scratch training on medical images, raising questions about whether the reported gains are fully converged or sensitive to early stopping.
Reliable clinical deployment of deep medical image models is hindered by distribution shifts across scanners, sites, and acquisition protocols. Existing domain generalization (DG) methods often focus on style or intensity diversification, but they can still leave networks dependent on domain-specific texture correlations. Inspired by evidence that Fourier phase encodes semantic structure, we introduce PhaseAT, a phase-aware adversarial training framework for medical DG. PhaseAT forms phase-perturbed training views in the Fourier domain by iteratively updating a bounded phase perturbation while keeping the amplitude spectrum unchanged, thereby stressing spatial organization under matched appearance statistics. Perturbations are applied only to the luminance channel in YCbCr color space to avoid chromatic artifacts. Additionally, a simple phase-saliency mask concentrates updates on the most influential frequencies. The model is trained with a weighted combination of losses on clean and phase-perturbed samples, supporting both single-source and multi-source DG. We validate our method on two challenging medical datasets and demonstrate that PhaseAT achieves over 20% improvement in single-source domain generalization, outperforming several state-of-the-art DG methods. The code implementation is available at: https://github.com/ahmed-sharshar/PhaseAT.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.