ResNet-based deep learning for atopic dermatitis diagnosis using the SCIN dataset

Summary: Abstract Body: The diagnosis of atopic dermatitis (AD), a chronic inflammatory skin condition, is primarily reliant on clinical expertise, which can introduce variability and limit accessibility in resource-constrained settings. This study aimed to evaluate the potential of ResNet-based deep learning models to enhance diagnostic accuracy and efficiency for AD using the Skin Condition Image Network (SCIN), a publicly available dermatological image dataset. Data preprocessing involved resizing, normalization, and augmentation to optimize image quality and mitigate class imbalances. Three pre-trained architectures, including ResNet18, ResNet50, and ResNet152, were fine-tuned via transfer learning, with performance assessed using the area under the receiver operating characteristic curve (AUROC). Among the tested architectures, ResNet18 demonstrated the best generalization with a test AUROC of 0.58, outperforming deeper models that showed overfitting tendencies. ResNet50 and ResNet152 achieved superior training metrics but underperformed on the test set, reflecting the trade-offs between model complexity and dataset limitations. Notably, the SCIN dataset's underrepresentation of eczema-positive cases constrained the models’ performance and underscored the need for larger, class-balanced datasets. Our findings emphasize the promise of deep learning in automating AD diagnosis while identifying critical areas for improvement, including dataset curation, interpretability, and scalability. Future work will explore lightweight architectures for mobile deployment, aiming to democratize access to diagnostic tools and improve outcomes in underserved populations. This research represents a step towards integrating artificial intelligence into clinical dermatology, paving the way for equitable and precise care.