Susceptibility of Skin Lesion CNN Classifiers to Increasing Gaussian Noise

Noise is a realistic source of corruption in dermoscopic imaging, stemming from sensor limitations, low-light acquisition, and uncontrolled capture conditions. This study examines how this type of degradation alters the behavior of CNN-based skin lesion classifiers by systematically perturbing a balanced ISIC 2018 subset with increasing noise intensities. Four representative CNN architectures are compared under a unified training protocol, and their performance is analyzed across evaluation setups that separate matched and mismatched noise conditions between training and inference. Beyond standard Accuracy, we report Precision, F1 score, and Matthews Correlation Coefficient to capture changes in both overall and class-sensitive reliability. The analysis shows that the main performance loss occurs when models are tested at noise levels higher than those encountered during training, whereas degradation is less disruptive when the noise distribution is consistent across both phases. These results underscore the importance of explicitly accounting for noise variability when deploying a skin lesion classifier in real-world clinical settings.