Towards Robust Eczema Region Identification in Clinical Skin Images via Domain-Adaptive Deep Hashing

Abstract Accurate identification of eczema-affected regions is a critical step toward reliable computer-aided dermatological analysis, yet the scarcity of labeled data in clinical settings limits model performance. To address this challenge, we explore domain adaptation as a strategy to leverage richly annotated data from related skin conditions while adapting to target domains with sparse labels. We introduce a novel deep learning framework that integrates labeled source data with unlabeled target samples, enabling the model to learn transferable representations that remain robust across domain shifts. A curated dataset of dermatoscopic images, including eczema regions and other dermatological conditions, is constructed to evaluate the effectiveness of the proposed approach. By combining domain-adaptive feature extraction with deep hashing mechanisms, the framework enhances cross-domain generalization and accurately localizes eczema regions in previously unseen images. Experimental results demonstrate substantial improvements over conventional baselines in both recognition accuracy and robustness, highlighting the potential of the method to advance automated eczema detection for real-world clinical practice.