On the transferability of photometric augmentations across medical imaging domains

Data augmentation is widely used to improve robustness in medical image classification, yet the transferability of photometric augmentation strategies across imaging domains remains underexplored. Channel-wise perturbations are commonly assumed to promote invariance to illumination and acquisition variability, although medical imaging modalities differ substantially in their acquisition physics and diagnostic color semantics. In this study, we systematically evaluate channel-wise Gaussian perturbations across retinal fundus imaging (APTOS), dermoscopic skin lesion classification (ISIC), and gastrointestinal endoscopic imaging (HyperKvasir). Using a fixed Xception backbone, deterministic preprocessing, and five-fold cross-validation, we assess red-, green-, blue-, and RGB-channel perturbations under low, medium, and high intensity levels. The results show that channel-wise perturbation is not universally beneficial. Low-intensity perturbations generally preserve performance and occasionally provide small gains, whereas stronger perturbations frequently degrade performance. The clearest positive mean shift is observed in dermoscopic imaging, where medium-intensity blue-channel perturbation yields the highest mean performance among the evaluated configurations. In contrast, retinal and endoscopic tasks show limited or task-specific benefits. Across domains, high-intensity RGB perturbation is consistently harmful, causing substantial Macro-F \(_1\) degradation in several tasks. The results demonstrate that photometric augmentation in medical image classification is strongly channel-, intensity-, and modality-dependent, and should therefore be treated as a controlled design choice rather than a default robustness strategy.