Abstract: Carotid artery stenosis is one of the primary culprits behind ischemic stroke, a leading cause of morbidity and mortality worldwide. Precise diagnosis of this condition is crucial for effective patient management and treatment. Currently, Digital subtraction angiography (DSA) is the gold standard for diagnosing carotid artery stenosis, offering clear and detailed images that enable radiologists to visualize the degree of arterial narrowing. However, traditional diagnostic approaches often rely on radiologists to manually scrutinize and interpret these images—an inherently time-consuming process that is prone to human error and heavily dependent on professional expertise. In this context, the advent of artificial intelligence has brought forth promising auxiliary diagnostic tools aimed at enhancing the efficiency and accuracy of medical image analysis. Despite significant advancements, many current recognition systems have limited capacity, typically detecting only a single area of stenosis within an image. Such limitations can lead to incomplete or inaccurate diagnoses, potentially compromising patient outcomes. To address these challenges and improve diagnostic performance for carotid artery stenosis, this study introduces a novel approach: the Non-local channel attention network (NLCANet), an advanced network architecture specifically designed to enhance the classification accuracy of carotid artery stenosis. By leveraging non-local attention mechanisms and channel-wise feature extraction, NLCANet provides a more robust and nuanced approach to image analysis, ensuring more precise and reliable diagnostic results. This innovation aims to advance the field of medical image classification and significantly improve clinical outcomes by enabling faster and more accurate diagnoses. The proposed model is built upon two integral components that operate synergistically: the Non-local multi-scale fusion (NLMSF) module and the Multi-level channel attention (MLCA) module. The NLMSF module draws inspiration from non-local networks, which are capable of capturing long-range dependencies and contextual relationships within an image. By simulating spatial attention operations, the NLMSF module enables the model to focus on critical regions across the entire image, thereby incorporating both local detail and global context. Furthermore, it integrates a multiscale feature fusion strategy—particularly important in medical imaging, where stenotic regions often vary in size and complexity. In parallel, the MLCA module enhances the model’s ability to prioritize and utilize channel-wise information. Medical images typically contain rich multidimensional data distributed across channels, with each channel contributing differently to the classification task. The MLCA module is designed to effectively identify and weight channels that carry the most diagnostically relevant semantic information. This attention mechanism improves the model’s capacity to differentiate between varying degrees of stenosis, thereby reducing the risk of misclassification and improving diagnostic accuracy. By combining both spatial and channel-level attention mechanisms, NLCANet achieves a more comprehensive understanding of medical images, resulting in superior classification performance. We constructed a carotid artery stenosis dataset by applying key frame-extraction techniques to DSA sequences and conducted comparative experiments against several mainstream medical image classification models. Experimental results show that NLCANet achieves the best performance, with classification accuracy at least 2% higher than that of the other benchmark models. These findings demonstrate the effectiveness and clinical potential of NLCANet as an accurate, efficient, and reliable diagnostic tool for carotid artery stenosis.