Transformer-Based RT-DETR Framework for Accurate Chest X-Ray Disease Detection

Abstract

Objectives

Chest X-ray (CXR) analysis is a vital tool for early disease detection, enabling timely diagnosis and treatment. However, achieving accurate CXR disease detection remains challenging due to low image contrast, overlapping anatomical structures, and variations in imaging conditions. This study aims to develop a robust and efficient disease detection framework using the RT-DETR model to address these limitations and improve diagnostic accuracy.

Material and Methods

The proposed framework integrates transformer-based components within the RT-DETR architecture for enhanced feature representation. The backbone incorporates HGBlock and HGStem modules to capture multi-scale spatial representations through hierarchical gradient flow. In the neck network, the Attention-Intensified Feature Interaction (AIFI) module and the Reparameterized Efficient Path Aggregation (REpc3) module refine feature fusion and strengthen contextual understanding. The detection head employs the RT-DETR decoder with an efficient query-based mechanism to improve localization and classification precision. Statistical validation was conducted using the Wilcoxon Test, Paired T-Test, and Kruskal-Wallis Test to ensure the reliability of performance outcomes.

Results

The proposed RT-DETR framework achieved a precision of 55.7%, outperforming YOLOv7x's 47.7% by 8%. The recall was comparable, with our model achieving 43.0% versus YOLOv7x's 43.1%. Importantly, the mean Average Precision (mAP) of our model reached 45.3%, representing a 3.7% improvement over YOLOv7x's 41.6%. These results confirm the model's superior performance and its statistical significance as validated by the applied tests.

Conclusion

The RT-DETR-based disease detection framework demonstrates improved accuracy and robustness in CXR analysis compared to state-of-the-art models. Its integration of advanced transformer-based modules enhances feature representation and detection precision. The results highlight its potential as a reliable and efficient tool for automated chest X-ray disease diagnosis, offering strong applicability in real-world clinical settings.