DSSViT: Multi-Scale Adaptive Fusion Vision Transformer With Dense Feature Reuse for Robust Pneumonia Detection in Chest Radiography

Abstract

Accurate pneumonia diagnosis using chest x-rays (CXR) remains a critical challenge due to the need for precise extraction of fine-grained local features and effective multi-scale spatial pattern recognition. While Vision Transformer (ViT) models have demonstrated strong performance in medical imaging, they often struggle with these aspects, limiting their effectiveness in clinical applications. This study proposes Dense-SEA ViT (DSSViT), an enhanced Vision Transformer architecture, to address these limitations by improving fine-grained feature representation and multi-scale spatial information capture for pneumonia detection. DSSViT integrates DenseNet121 as a feature extractor to enhance feature reuse and improve information flow, thereby compensating for ViT's weakness in capturing low-level visual details. Additionally, the Squeeze-Excitation and Adaptive Fusion (SEA) mechanism is introduced to calibrate channel attention and enable multi-scale adaptive fusion, enhancing the model's ability to extract critical diagnostic features while reducing noise interference. The proposed architecture was evaluated on a chest X-ray dataset for pneumonia classification. Experimental results demonstrate that DSSViT achieves superior feature extraction capability, leading to a test accuracy of 97.69%, outperforming baseline models such as EfficientNet (93.90%) and VGG19 (96.57%). These findings suggest that DSSViT is a promising approach for improving automated pneumonia diagnosis in clinical settings.