Abstract: Carotid artery stenosis is one of the primary causes of ischemic stroke. Currently, digital subtraction angiography (DSA) is considered the gold standard for diagnosing carotid stenosis. However, traditional diagnostic methods require pathologists to manually screen and analyze DSA images, which are slow and prone to errors. To address this, we have developed a carotid stenosis dataset using key frame extraction technology. While general models for image classification have excelled in the field of computer vision, there is still significant potential for growth in models classifying medical images. Therefore, we propose a Non-Local Channel Attention Network (NLCANet) to assist in the accurate classification of carotid stenosis. The NLCANet primarily consists of two modules: the Non-Local Multi-Scale Fusion module (NLMSF), which utilizes the concept of non-local networks to simulate spatial attention operations, and incorporates a multi-scale fusion module to enhance feature extraction across different scales, playing a crucial role in classifying carotid images; and the Efficient Channel Attention (ECA) module, which effectively utilizes the channel features of images to provide additional semantic information for classification. We compared NLCANet with other mainstream medical image classification models on our carotid stenosis dataset. Our model achieved the best results, with a classification accuracy at least 2% higher than other mainstream models.