Edge-Guided U-Net with Frequency Domain Refinement for Efficient Skin Lesion Segmentation

Precise segmentation of skin lesions is crucial to support the diagnosis of skin cancer within computer-aided diagnosis (CAD) systems. Although deep learning-based models like U-Net along with its variants have demonstrated strong segmentation performance, they often require substantial computational and energy resources. In this study, a lightweight and computationally efficient framework for skin lesion segmentation is proposed, leveraging information across the spatial and frequency domains. The proposed model, Edge-Guided U-Net (EGU-Net), integrates edge information directly into the input and incorporates a frequency-domain refinement step based on the discrete cosine transform (DCT). Through this integration, EGU-Net enables improved delineation of lesion boundaries while suppressing high-frequency noise in the predicted masks. The model underwent training and evaluation on the ISIC 2018 dataset, resulting in a Dice score of 87.85%, recall of 90.62%, precision of 89.02%, an intersection over union (IoU) of 80.35% and an accuracy of 95.47%. These results demonstrate that EGU-Net can approximate the performance of more complex architectures while maintaining significantly lower computational complexity. Consequently, EGU-Net establishes a strong baseline for interpretable and efficient pixel-wise skin lesion segmentation under constrained computational settings.