Abstarct: Despite the remarkable advancements of deep neural networks in medical image analysis, two-dimensional chest X-ray radiography persistently poses significant challenges for automated computer-aided diagnosis systems due to anatomical tissue compression, dense bone shadows, and the high visual similarity of infectious patterns. Classical convolutional models suffer from a limited local receptive field, while standard vision transformers exhibit quadratic computational complexity, hindering an optimal balance between clinical accuracy and processing throughput. In this dissertation, a novel multi-scale hybrid architecture is proposed, baxsed on a dual processing pipeline of a convolutional neural network and a transformer equipped with an efficient linear attention mechanism. In the developed methodology, feature maps are simultaneously extracted across different frequency levels and, after dimension alignment, are fed into the transformer laxyers. The computational frxamework of the transformer reduces space and time complexity from quadratic to linear by generating a compressed query vector through a global average pooling operator. Experimental evaluation of the model on two international benchmark datasets, Cohen (adult multi-class) and Kermany (imbalanced pediatric), demonstrates high network stability in extracting statistical metrics. The proposed architecture achieved an overall accuracy of 96.15% and an F1-score of 95.61% on the Cohen dataset, and a recall of 96.70% and an F1-score of 96.07% on the Kermany dataset. Activation map analysis verified that dynamic tensor compression prior to the transformer acts as a soft spatial filter, suppressing irrelevant rib bone signals and minimizing Type II errors (false negatives), thereby delivering superior robustness and generalizability for real-world clinical applications.