DPA-HairNet: A Dual Encoder Attention Based Network for Hair Artifact Removal in Dermoscopic Images

Hair artifacts in dermoscopic images significantly hinder the accurate diagnosis of melanoma and other skin conditions by obscuring critical lesion details. To address this challenge, we introduce DPA-HairNet, a novel Dual Encoder Attention-Based Network designed specifically for effective hair artifact removal while preserving lesion integrity. The model features a dual encoder architecture to separately process hair-specific and lesion-specific features, an attention mechanism to prioritize diagnostically relevant regions, and a dual-output design that generates precise hair segmentation masks and high-quality reconstructed images. This study leverages the ISIC 2018 and HAM10000 datasets, incorporating advanced data augmentation techniques to enhance dataset diversity and ensure robust model training. DPA-HairNet was evaluated against six state-of-the-art segmentation models using comprehensive metrics, including Accuracy, Dice Coefficient, Jaccard Index, PSNR, MAE, Specificity, Precision, Recall, and F1-Score, where the proposed model outperformed all models. Furthermore, classification performance was assessed before and after hair artifact removal using twelve pre-trained classifiers, demonstrating significant improvements in diagnostic accuracy. Explainable AI techniques, Grad-CAM, UMAP and attention heatmaps, were employed to interpret and validate the model’s focus. These results underscore DPA-HairNet’s effectiveness and potential integration into clinical workflows to enhance dermoscopic image analysis and diagnosis. Future work will explore generalization to additional artifact types and optimization for real-time clinical deployment.