Adaptive collaborative feature fusion and shape-aware optimization for multi-scale chest lesion detection

Chest diseases remain a major cause of global morbidity and mortality, and accurate detection from chest X-ray images is critical for early diagnosis and clinical decision-making. However, large variations in lesion scale, morphology, and spatial distribution pose significant challenges for automated detection systems, particularly in identifying small lesions and achieving precise localization. To address these issues, we propose a multi-scale chest lesion detection method based on adaptive collaborative feature fusion and shape-aware optimization. The method enhances multi-scale feature modeling and introduces shape structural constraints to improve detection accuracy and localization robustness in complex anatomical environments. Specifically, a Context-Embedded Feature Enhancement Network is designed to jointly capture global anatomical context and local lesion characteristics, strengthening lesion representation. An Adaptive Feature Focusing Network further improves multi-scale feature representation through adaptive spatial feature aggregation, enabling more effective detection of small lesions. In addition, a shape-aware optimization strategy integrating normalized Wasserstein distance with shape-weighted constraints improves localization stability and bounding box regression accuracy for irregular lesions. Compared with state-of-the-art methods, the proposed method achieves improvements of 6.3% in mAP, 6.4% in small-lesion mAP, and 8.6% in mean recall on VinDr-CXR, as well as improvements of 4.2% in mAP and 7.8% in mean recall on ChestX-ray8, demonstrating its effectiveness and generalization capability for multi-scale chest lesion detection.