The rapid identification of COVID-19 from chest computed tomography (CT) images remains challenging due to the diverse appearance, distribution, and scale of pulmonary abnormalities. Conventional convolutional neural networks (CNNs) can capture detailed local patterns but may have limited ability to model long-range spatial relationships, whereas Transformer-based models provide broader contextual representations but may overlook fine-grained image characteristics. To address these limitations, this study proposes a lesion-aware multi-scale CNN–Swin Transformer framework with adaptive attention fusion for explainable COVID-19 detection from chest CT images. The proposed framework first emphasizes pulmonary regions to reduce the influence of irrelevant image content and then extracts multi-scale local features using a CNN branch. In parallel, a Swin Transformer branch learns hierarchical contextual representations from the CT images. An adaptive attention fusion module is introduced to dynamically integrate local and global features rather than relying on direct feature concatenation. The resulting representation is used for binary classification of COVID-19 and non-COVID-19 cases. To improve model transparency, explainability techniques are incorporated to visualize the image regions contributing to the predictions. In addition to conventional classification metrics, the proposed approach is evaluated through comparative and ablation experiments to examine the contribution of individual components, while robustness analysis is performed under controlled image-quality variations. The proposed framework aims to provide an accurate, interpretable, and more robust approach for automated COVID-19 detection from chest CT images.

The proposed framework aims to provide an accurate, interpretable, and more robust approach for automated COVID-19 detection from chest CT images. The study further investigates whether focusing on lesion-relevant regions improves the discrimination of subtle COVID-19 patterns. The ablation analysis provides a systematic assessment of the contribution of multi-scale learning and adaptive feature fusion to the overall classification performance. The findings are intended to support the development of reliable computer-aided tools for CT-based respiratory disease assessment.

Keywords— COVID-19; chest CT; deep learning; convolutional neural network; Swin Transformer; multi-scale feature learning; adaptive attention; explainable AI.