Image Resolution vs. Accuracy Trade-Off in Dermatology AI

Abstract

Deploying AI-powered dermatology tools on mobile and edge devices requires a critical balance between classification accuracy and computational efficiency. While higher-resolution images provide granular details necessary for identifying skin conditions, they impose significant computational costs. This paper investigates the trade-off between image resolution (64 x 64, 128 x 128, 224 x 224, 256 x 256) and performance metrics (Accuracy, AUC, Model Size) for skin disease classification using CNNs. We evaluate MobileNet and ResNet50 architectures on a dataset of Acne vs. Normal skin. Our results demonstrate that MobileNet achieves a superior balance, maintaining high accuracy (>90%) at lower resolutions (128 x 128) while consuming significantly less memory (12.5 MB) compared to ResNet50 (90 + MB), identifying it as the optimal choice for resource-constrained deployment.