Multi-scale parallel perception and adaptive feature fusion for chest x-ray multi-disease detection

Objective.Chest x-ray (CXR) imaging is widely used for screening thoracic diseases. However, accurate multi-disease detection remains challenging because lesion sizes vary substantially, multiple abnormalities often coexist, and small or low-contrast lesions can be obscured by anatomical structures. This study aims to improve multi-disease detection in CXR images by enhancing multi-scale lesion representation and cross-level feature consistency.Approach.We propose a CXR multi-disease detection framework based on multi-scale parallel perception and adaptive feature fusion. The Multi-scale Parallel Feature Sensing module captures complementary lesion features under different receptive fields. The Smooth Residual Fusion Block refines lesion-related spatial responses and suppresses irrelevant background interference. The adaptive pyramid feature fusion module dynamically integrates multi-level features to reduce semantic discrepancies and preserve fine-grained lesion details.Main results.On the VinDr-CXR dataset, the proposed method achieved mean Average Precision scores of 0.389 forand 0.186 for, corresponding to absolute improvements of 5.3% and 1.9% points over the baseline, respectively. The gains were particularly evident for small and subtle abnormalities, such as nodules/masses and pneumothorax. Additional validation on the bounding-box annotated subset of ChestX-ray8 further confirmed robustness, with,, and Recall values of 0.186, 0.093, and 0.240, respectively.Significance.The proposed framework provides an effective solution for structurally consistent multi-disease detection in CXR images. Results on two public datasets suggest that multi-scale parallel perception and adaptive feature fusion improve lesion localization robustness, showing potential as a supportive tool for computer-aided screening workflows.