NSEC-YOLO: Real-time lesion detection on chest X-ray with adaptive noise suppression and global perception aggregation

Abstract

Chest diseases significantly threaten respiratory health, making early and accurate diagnosis essential for improving patient survival rates. Traditional detection methods face difficulties in accurately identifying and localizing chest lesions, attributed to the intricate morphology of lung diseases and interference from background noise, subsequently elevating the risk of misdiagnosis. To address these challenges, we propose a novel method called NSEC-YOLO, which significantly enhances the efficiency and accuracy of chest disease detection. Firstly, we introduce an adaptive noise suppression module during the visual feature extraction stage, effectively reducing background noise interference and improving the clarity and precision of feature representation. Secondly, we employ a global perceptual aggregation detection head that strengthens the model’s performance in classification and regression tasks, thereby improving the accuracy and reliability of detection results. Finally, we incorporate a well-designed AccurEIOU-Loss to fine-tune the training process, thereby augmenting the detection accuracy and efficiency. To comprehensively validate the performance of NSEC-YOLO, extensive experiments were conducted on the public VinDr-CXR dataset and systematically compared it with popular detection models such as YOLO-v5, YOLO-v9, and SSD. The experimental findings indicate that NSEC-YOLO excels in detecting lung diseases in chest X-ray images, achieving a precision of 0.416 under the [email protected] standard and an accuracy of 0.194 under the more stringent [email protected]:0.95 standard. Notably, NSEC-YOLO maintains a high processing speed of 163 frames per second while delivering high detection accuracy, outperforming mainstream detection models in both precision and efficiency. These results underscore the strong application potential and practical value of NSEC-YOLO in real-time chest X-ray lesion detection.