Dermatoscopy is a common diagnostic tool used by dermatologists to examine skin lesions. However, existing classification methods often rely on spatial domain feature extraction. Dermatoscopic images are frequently affected by various sources of noise, such as hair and pigmented areas, resulting in non-robust feature extraction. Therefore, relying solely on spatial domain feature extraction is limited and cannot fully differentiate the subtle features of skin lesions. This paper proposes a multi-scale and local frequency enhancement network (MFENet) for skin lesion, which significantly improves the diagnostic accuracy and interpretability of dermatoscopic images. Specifically, we introduce the multi-scale and local of frequency (MLF) module to reduce noise interference and enhance edge, shape, and texture features. The module selects low-frequency information and suppresses high-frequency noise at different scales. It further enhances details and texture information in the image by processing local frequency blocks. This makes important morphological features, such as the shape and edge, more prominent in diagnosis. Additionally, a lesion channel attention (LCA) module is proposed to further optimize the extraction and amplification of key features by learning dependencies between channels. The results show that the algorithm achieves optimal performance on the ISIC 2017 and HAM10000 datasets.