Abstract: Early diagnosis of skin diseases relies heavily on accurate segmentation of lesions in dermoscopic images. However, lesions often have blurred boundaries and irregular shapes. Furthermore, images are often affected by interference from factors such as hair and shadows, making it difficult to effectively extract key features. These factors significantly impact the segmentation results. To address these issues, a skin disease segmentation network (MPUNet) is proposed. This network integrates a multi-scale feature enhancement module (MS-Module) and a pooling attention module (PL-Module). The MS-Module incorporates discriminative spatial prior information and adaptively aligns and fuses channels in both height and width, effectively enhancing the representation of multi-semantic information and the ability to extract features from lesion edges. The PL-Module employs a pooling attention mechanism to enhance the extraction of global contextual features. By activating relevant channels, it suppresses the impact of interference from factors such as hair and blood vessels on segmentation performance, while also alleviating the vanishing gradient phenomenon during training. Experimental results on the ISIC2018, ISIC2017, and PH2 public skin disease segmentation datasets demonstrate that MPUNet achieves an average improvement of 1.57 percentage points in the mean intersection over union (mIoU) metric.