Hierarchical Transformer Residual Model for Pneumonia Detection and Lesion Mapping

Abstract

Pneumonia, a potentially fatal infection and a common disease-causing culprit among children and the elderly, still remains as a prevalent threat even after years of research on tackling it. Rapid and proper identification is crucial for timely treatment and improved results. While thoracic radiographs are widely employed in pneumonia diagnosis, real-world clinical assessment is frequently questioned by factors such as subtle radiographic patterns, overlapping symptoms, subjective manual judgement and dependency on expert radiologists. The study proposes a hybrid deep learning model integrating ResNet50 and the Swin Transformer, coupled with an auxiliary segmentation decoder to facilitate both classification and lesion localization in chest X-ray images. ResNet50 acts as the backbone for hierarchical spatial feature extraction, capturing fine-grained local textures indicative of pulmonary abnormalities, and the Swin Transformer serves as the global attention-driven feature aggregator. The shifted window mechanism of the Swin Transformer maintains spatial hierarchy while facilitating effective contextual modelling. Global Average Pooling (GAP) and Multilayer Perceptron (MLP) form the classification head, yielding accurate predictions in classifying the images, while the segmentation decoder utilizes multiscale features to generate pixel-wise masks for pneumonia lesion regions. The model outperformed conventional methods with 98.4% classification accuracy, 98.2% precision, 99.2% recall and an F1-score of 98.7% with a 0.88 Dice Coefficient in segmentation. These results reflect the hybrid architecture's superior performance and its dual capacity for diagnostic prediction and lesion interpretability. The proposed model demonstrates promising results for deployment in real-world clinical workflows, especially in resource-constrained or high-patient-load environments.