RGCA-Swin-Tiny CXRNet: A Residual Gated Convolutional Attention Guided Swin Transformer for Chest X-ray Classification | IJCT Volume 13 – Issue 4 | IJCT-V13I4P4

This research proposes RGCA-Swin-Tiny CXRNet, a hybrid deep learning model for five-class chest X-ray classification. ImageNet-pretrained Swin-Tiny Transformers and lightweight convolutional attention branches are used. The Swin-Tiny branch collects hierarchical global contextual representations, while the convolutional attention branch extracts supplementary local radiography characteristics. The Residual Gated Convolutional Attention (RGCA) technique, where the convolutional descriptor creates a centered residual gate to adaptively suppress or boost the Swin-Tiny feature representation before classification, is the fundamental innovation. The trial used a balanced five-class single-label ChestX-ray8 configuration with Atelectasis, Hernia, No Finding, Pneumonia, and Pneumothorax. The model has test accuracy of 64.67%, weighted precision of 64.48%, recall of 64.67%, F1-score of 64.44%, MCC of 0.5583, and AUC of 0.8236. The class-wise study indicated that the model performed best for Hernia, although visual overlap and slight radiographic changes made Atelectasis and No Finding harder. The confusion matrix and t-SNE visualization demonstrated that the model learnt valid feature representations, yet several illness categories had overlapping feature distributions. Transformer-based contextual modeling and convolutional attention-guided local feature refinement show promise for chest X-ray categorization. The proposed RGCA technique lets local convolutional cues dynamically control the pretrained Swin-Tiny representation, unlike CNN–Transformer feature concatenation. This allows radiographic image processing to combine local and global information with controlled and adaptive fusion. Despite these promising outcomes, this study has drawbacks. The experiment started using a five-class single-label subset of ChestX-ray8 instead of the full multi-label setup. Second, the study employed a balanced selection of pictures, which may not reflect real-world radiology dataset class imbalance. Third, while the proposed model shows beneficial feature learning, baseline comparisons and ablation experiments would increase the empirical validity of the RGCA process. The suggested approach may be expanded to multi-label ChestX-ray14 classification in the future. Use patient-wise data separation to improve clinical validity and prevent data leakage. Ablation of the convolutional attention branch and residual gate, comparison with CNN, Transformer, and CNN–Transformer baselines, and assessment on CheXpert or MIMIC-CXR may increase performance. To improve clinical interpretability, Grad-CAM, attention map visualization, and class activation analysis can be used. The model may also learn the residual gating intensity to automatically calculate the appropriate convolutional attention-based feature modulation during training.