DAMMD-Net: A lightweight and enhanced deep segmentation network for skin lesion detection

Abstract

Early and accurate diagnosis of skin cancer is critical to improving survival rates. Dermoscopy is one of the most important imaging techniques for this purpose. However, manual examination of dermoscopic images is laborious, time-consuming, and error-prone due to variations in the color, shape, location, texture, and size of skin lesions. Therefore, developing automatic segmentation models is crucial for assisting physicians in both qualitative and quantitative assessments. Although numerous deep learning-based segmentation models have produced satisfactory results in skin lesion detection, their backbone architectures still face intrinsic limitations and extrinsic challenges. In light of this motivation, this paper proposes a lightweight and enhanced segmentation network (DAMMD-Net) based on the DeepLabV3+ model, with an attention mechanism (AAC) and modified decoder to improve segmentation performance. The AAC is used as a local feature enhancement tool to address the interference of useless information related to healthy skin. The modified decoder module enhances the network's ability to capture spatial details and integrate contextual information by leveraging multi-level feature maps from the encoder. The proposed segmentation pipeline has been evaluated on two well-known benchmark datasets: ISIC2018 and PH2. The experimental results showed that DAMMD-Net achieved an average Dice similarity coefficient (DSC) of 0.887 for the ISIC2018 dataset and 0.929 for the PH2 dataset, outperforming the backbone network. The overall results revealed that the proposed DAMMD-Net not only achieved satisfactory performance compared to existing models but also demonstrated significant potential for clinical practice due to its lightweight architecture.