Deep learning-based classification of cerebrovascular lesions on computed tomography images

Abstract

Background and objective: Computed tomography (CT) is a widely used and cost-effective imaging modality for the assessment of cerebrovascular disorders. Although many artificial intelligence and machine learning methods have been developed for automated image classification, most studies have focused on binary or limited multi-class tasks. This study proposes a novel artificial intelligence based classification framework for the diagnosis of cerebrovascular lesions using a newly designed convolutional neural network inspired by the Swin Transformer. Materials and methods: A retrospective in-house dataset consisting of 10,865 CT images obtained from 733 individuals was used in this study. The dataset included ten classes of cerebrovascular disorders. Two models, namely SwinNeXt and SwinNeXt-Deep, were developed. Deep features extracted by SwinNeXt were optimized using four iterative feature selection methods: Chi-squared, Neighborhood Component Analysis, minimum redundancy maximum relevance, and ReliefF. The selected features were then classified using the k-nearest neighbor algorithm. Results: The introduced SwinNeXt-Deep model achieved 98.54% recall, 98.59% precision, 98.56% F1-score, and 98.56% accuracy. In addition, it outperformed the baseline SwinNeXt model by approximately 2% and exceeded several well-known convolutional neural network models by 2% to 6% in terms of classification accuracy. Conclusions: The computed results and obtained findings demonstrate that the proposed models are effective for the multi-class classification of cerebrovascular lesions on CT images. These findings indicate that the proposed framework has strong potential for use in computer-aided clinical decision support systems in neuroimaging.