Impact of Image Degradation on Melanoma Detection by Artificial Intelligence: A Patient Safety Perspective

Introduction The integration of artificial intelligence (AI) in teledermatology presents significant diagnostic potential, yet the vulnerability of these models to real-world image degradation remains a major patient safety concern. This study investigates the accuracy discrepancy, where global metrics obscure severe sensitivity drops in melanoma detection under visual stress. Methods We evaluated two deep learning architectures, ResNet-50 and MobileNetV2, using the HAM10000 dataset. An experimental pipeline was developed using the Python programming language to simulate three clinical photographic artifacts: Gaussian noise, focal blur, and low luminance. The models were subjected to stress tests to observe shifts in diagnostic classification and evaluate their structural stability. Results While baseline global accuracy remained deceptively stable, targeted degradation exposed catastrophic failures. Under moderate degradation, ResNet-50 sensitivity collapsed to near-zero levels, erroneously classifying malignant lesions as benign nevi. MobileNetV2 exhibited similar structural vulnerabilities, proving that standard evaluation metrics fail to capture critical diagnostic risks. Conclusions Deep learning models in dermatology exhibit significant vulnerability to common image degradation, posing a direct threat to patient safety. The implementation of automated image quality gates is essential before deploying these algorithms in clinical settings to prevent AI-induced diagnostic errors.