<p>Laryngeal cancer has become a global public health challenge due to its high incidence and low early diagnosis rate. Early lesions are difficult to detect with traditional white-light endoscopy, while Narrow Band Imaging (NBI) technology can improve the accuracy of vascular lesion identification, but faces the challenges of low equipment popularity and the absence of dual-modality image collaborative display. To address this technical bottleneck, we propose an unsupervised throat Narrow Band Imaging translation method(NBICUT), which enables precise conversion of white-light laryngoscopic images into NBI images through unpaired image translation. We constructed a multi-scale cross-layer dual attention module (CLDA), which uses a hierarchical feature fusion approach to obtain richer white light image information, thereby enhancing the correlation between multi-scale spatial features and channel features. Additionally, a parallel multi-scale feature extraction module (PMFE) with a parallel convolutional architecture is designed to dynamically adjust the weights of features at different scales while reducing convolutional kernel parameters, enabling adaptive focused extraction of features. Finally, a texture perception loss function is employed to constrain the texture features distribution discrepancy between the source and target domains, thereby optimizing the generation quality of vascular edges. Experimental results show that NBICUT achieves SSIM up to 0.9150, SSIM(s) up to 0.9412, LPIPS as low as 0.1659, PIQE as low as 2.41, and BRISQUE as low as 5.17 on two hospital datasets, outperforming state-of-the-art methods in structure preservation, vascular detail retention, and NBI style conversion.</p>