AG PVM‐UNet: An Efficient Attention‐Enhanced Network for Skin Lesion Segmentation in Computer‐Aided Diagnosis

Abstract

Skin lesion segmentation is a key component of computer‐aided dermatological analysis and supports early detection of skin cancer. However, many existing segmentation approaches rely on computationally intensive architectures and have difficulty modeling both fine‐grained local structures and long‐range contextual information. These limitations restrict their practical use in resource‐constrained clinical environments. This study proposes a lightweight segmentation framework termed AG PVM‐UNet, which integrates convolutional feature extraction with a Parallel Vision Mamba mechanism for efficient global context modeling. The architecture incorporates attention‐guided feature refinement and adaptive multi‐scale representation to improve lesion boundary delineation while maintaining low computational complexity. The proposed method is evaluated on the ISIC 2017 and ISIC 2018 dermoscopic image data sets and compared with several representative segmentation models. Experimental results show that AG PVM‐UNet achieves improved segmentation accuracy while significantly reducing model complexity. Compared with VM‐UNet, the proposed model increases the Dice Similarity Coefficient and accuracy by 2.34% and 0.94%, respectively, while reducing the number of parameters and computational complexity by 99.87% and 98.44%, respectively. These results demonstrate that the proposed framework provides accurate and computationally efficient skin lesion segmentation, indicating its potential for integration into practical medical imaging analysis systems.