EfficientNet with attention and global blocks for accurate pulmonary disease detection in chest CT scans

The rapid and accurate diagnosis of COVID-19 from computed tomography (CT) scans is critical for effective disease management. However, existing deep learning architectures struggle to balance high diagnostic accuracy with computational efficiency, often suffering from immense parameter overhead and suboptimal specificity. To address this, we introduce the EfficientNet-B0 Non-Local Attention (EN-NL) model. Evaluated on a consolidated, near-symmetrically balanced dataset comprising 746 chest CT scans, this novel architecture strategically appends a Non-Local attention block immediately following the terminal convolutional activation of the EfficientNet-B0 backbone. This approach effectively captures long-range dependencies without altering the internal structure of the base network. Evaluated on an independent, unseen test subset, the proposed EN-NL model achieved an Accuracy of 98.00%, a Precision of 100.00%, and a perfect Specificity of 100.00% (zero false positives). Crucially, comparative analysis reveals that while the EN-NL model matches the high accuracy (98.00%) of the leading baseline architecture (VGG-SA), it demonstrates profound computational superiority: achieving a 95.8% reduction in network parameters (5.8 M vs. 138.4 M) and accelerating training time by 66.2% (4.15 vs. 12.30 min). These findings confirm that the proposed framework not only eliminates false positive diagnoses but also provides a highly lightweight, scalable, and clinically viable solution for rapid COVID-19 screening in resource-constrained medical environments.