HybridCATNet: an explainable CNN-Transformer framework with early multi-channel fusion and uncertainty estimation for...

Pneumonia remains a leading cause of morbidity and mortality worldwide, particularly among children and older adults, making rapid and reliable interpretation of chest X-rays essential for timely clinical decision-making. This study proposes HybridCATNet, an explainable and uncertainty-aware deep learning framework for automated pneumonia detection from chest radiographs.

HybridCATNet integrates a convolutional neural network and a vision transformer to capture both fine-grained local radiographic features and long-range anatomical dependencies. To incorporate clinically relevant image priors, a nine-channel early multi-channel fusion strategy combines raw grayscale chest radiographs, lung-segmented masks, and edge-enhanced structural maps. A Cascaded Spatial-Channel Attention Module further refines diagnostic feature representation, while Bayesian dropout is employed during inference to estimate predictive uncertainty. The model was trained and evaluated on a balanced multi-institutional dataset comprising 8,478 authentic chest radiographs without synthetic oversampling using 10-fold cross-validation. Generalizability was further assessed on two independent external validation cohorts. Explainability was evaluated using Grad-CAM and SHAP.

HybridCATNet achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.9953, an accuracy of 96.93%, a sensitivity of 98.13%, and a specificity of 95.72% in 10-fold cross-validation. External validation demonstrated robust cross-domain performance, with an accuracy of at least 94.8% across both independent cohorts. Explainability analysis showed strong agreement between model-generated attention maps and radiologist-annotated pathological regions, achieving an intersection-over-union of 0.812 and a faithfulness correlation of 0.921.

The results demonstrate that HybridCATNet provides accurate, interpretable, and uncertainty-aware pneumonia detection while maintaining strong generalization across external datasets. By integrating multi-channel image fusion, hybrid CNN-Transformer feature learning, attention-guided refinement, predictive uncertainty estimation, and explainability, the proposed framework offers a clinically transparent decision-support approach with potential to assist pneumonia screening and improve confidence in AI-assisted radiographic diagnosis.