The study of a dermoscopic image segmentation method based on U-Net structure with feature enhancement

Since the 20th century, the number of melanoma-related deaths caused by skin cancer has been numerous. In this regard, dermoscopic image segmentation technology has made significant progress in recent years and has become an important tool for assisting doctors in diagnosis and treatment. However, on the one hand, the characteristics of skin lesions are in-builtly complex and diverse; on the other hand, existing methods often design increasingly complex models based on U-Net to explore the deep features of dermoscopic lesions, thereby neglecting the interaction between local and global features within the network. This leads to blurry edges in the predicted images and lower segmentation accuracy. Based on this, we use the U-Net architecture as a baseline and design a feature enhancement method to enhance the lesion image features, capturing the detailed characteristics of the lesion areas to improve segmentation accuracy. In short, we augment the original U-Net architecture with a feature enhancement module, which strengthens local feature representations in both the spatial and channel dimensions while highlighting interdependent features. This enhances feature representation and enables the network to more accurately segment lesion shapes, edges, and textures. We conducted experiments on the dermoscopic lesion segmentation dataset (ISIC 2016). The experimental results show that this feature enhancement method achieves better performance compared to other U-Net-based models.