An intelligent decision support framework for automated pneumonia classification from chest X-ray images using hybrid vision transformers and CNNs

Background Accurate and timely differentiation between multiple pneumonia subtypes, including COVID-19, remains a critical challenge in clinical chest radiography due to overlapping visual patterns in chest X-ray (CXR) images. Intelligent decision support systems based on medical informatics can assist clinicians by improving diagnostic consistency and efficiency. Objective This study aims to develop an efficient deep learning–based decision support framework that integrates global contextual modeling and local feature extraction for robust pneumonia classification from CXR images. Methods We propose two hybrid architectures designed for clinical decision support. The first model, MSAF-Net, employs a Multi-Scale Attention Fusion mechanism to effectively integrate Vision Transformer and EfficientNet representations. The second model, DAR-Net, introduces a Dynamic Attention Refinement module within a ResNet backbone to recalibrate channel-wise features prior to multimodal fusion, improving early-stage feature discrimination. Both models were evaluated on a curated dataset of 5200 chest X-ray images covering four diagnostic categories. Results DAR-Net achieved a balanced accuracy of 81.75% and an AUC of 0.9521, outperforming strong attention-based baselines. MSAF-Net attained comparable diagnostic performance while reducing model complexity by approximately 3.5 × , demonstrating suitability for resource-constrained clinical environments. Conclusions The proposed frameworks enhance automated pneumonia screening by supporting accurate and efficient diagnostic decision making. These results highlight the potential of hybrid deep learning models as reliable components of clinical decision support systems in medical imaging.