Intelligent Models Equivalent to the Level of Radiologists for Multi-Label Lesion Detection in Chest X-Rays

In the field of radiology, the analysis of X-ray images serves as an essential tool for diagnosing thoracic diseases. With the rapid evolution of deep learning technologies, automated image analysis methods have been extensively integrated, aiming to significantly enhance the work efficiency and diagnostic precision of radiologists. Nonetheless, existing deep learning models still encounter a series of challenges when processing abnormal chest X-ray images, including the similarity of multiple categories of lesions, the diversity of scales, and the difficulty in detecting small target lesions. These issues severely hinder the advancement of model performance. To surmount these limitations, this study introduces an model for the detection and localization of multiple diseases from abnormal chest X-ray images. The model incorporates an advanced CNN-Swin Transformer module, which integrates global and local context information and employs a pixel-level self-attention mechanism through a moving window mechanism, effectively capturing multi-scale features of the images. Furthermore, the model introduces a channel and spatial permutation attention allocation mechanism, significantly enhancing the recognition of salient features in key areas and thereby improving detection accuracy. We evaluated the proposed model on the VinDr-CXR dataset, and compared with the state-of-the-art baselines, it achieved an increase of 2.3% in mean Average Precision (mAP), demonstrating the effectiveness of our approach in advancing the field of automated chest disease detection.