Early detection and diagnosis of skin cancers is essential to improve patient survival. However, traditional diagnostic methods have limitations due to the complexity and diversity of skin lesions. Although deep learning-based skin disease detection methods are available, the ambiguity of the boundaries of skin lesion regions may lead to model neglect and misclassification, generating suboptimal results and affecting clinical decisions. To address this problem, this paper proposes a hybrid network based on Adaptive Grouped Transformer (AGT) and curvature information fusion for skin lesion detection, called AGTCNet. AGTCNet enhances the network’s adaptive multi-scale learning capability by introducing AGT. In addition, a curvature-based guidance enhancement module (CGEM) is proposed in this paper, which utilizes the curvature information to effectively guide the model in enhancing its capture of complex lesion edge information. To further optimize the model performance, the deep supervision mechanism is used to dynamically calculate the loss at each stage and adjust the learning strategy based on the loss feedback. Through comprehensive experimental validation on the ISIC2016, ISIC2017, and PH2 skin lesion segmentation datasets, the results show that AGTCNet significantly outperforms the existing mainstream methods on all datasets, and especially exhibits excellent performance in detailed feature processing and fuzzy region segmentation.