RGCA-Swin-Tiny CXRNet: A Residual Gated Convolutional Attention Guided Swin Transformer for Chest X-ray Classification

Automatic chest X-ray categorization is a crucial research topic for radiological screening and clinical decision-making. However, subtle disease-specific patterns, overlapping thoracic abnormalities, and considerable visual resemblance across chest illnesses make correct categorization difficult. RGCA-Swin-Tiny CXRNet, a hybrid deep learning model for chest X-ray classification, combines an ImageNet-pretrained Swin-Tiny Transformer with a lightweight convolutional attention branch to solve these issues. The convolutional attention branch catches local radiography signals such texture fluctuations, boundary changes, and opacity-like patterns, whereas the Swin-Tiny branch recovers hierarchical global contextual representations. The proposed model's Residual Gated Convolutional Attention (RGCA) method is novel, in which the convolutional descriptor creates a centered residual gate to adaptively suppress or boost the Swin-Tiny feature representation before classification. The suggested model is tested on a balanced five-class single-label ChestX-ray8 classification setting: Atelectasis, Hernia, No Finding, Pneumonia, and Pneumothorax. The dataset has 7500 images—5000 training, 1000 validation, and 1500 testing. Experimental findings demonstrate that RGCA-Swin-Tiny CXRNet 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 0.823. The confusion matrix and t-SNE visualization show that the proposed model learns meaningful discriminative representations, although visually overlapping classes like Atelectasis and No Finding remain difficult. Residual gated convolutional refinement is promising for chest X-ray classification using transformer-based global feature learning and local convolutional attention.

Chest X-ray classification, Swin Transformer, convolutional attention, residual gated attention, medical image classification, deep learning.