Thoracic diseases pose a significant threat to human health, and timely chest X-ray diagnosis is crucial for guiding clinical intervention. Traditional manual interpretation of chest X-rays is plagued by posture-related artifacts, subtle lesion missed detection, and heavy workload, resulting in low efficiency and high misdiagnosis rates. To address these issues, this study proposes SCH-Net, a novel hybrid network for the automatic classification of thoracic diseases on clinical chest X-rays, with a focus on balancing accuracy, real-time performance, and clinical interpretability. SCH-Net integrates four core components to achieve its goals: the STERN module corrects posture deviations in clinical chest X-rays and eliminates artifacts; a ViT-ResNet hybrid backbone captures both global lesion distribution and local subtle lesion details; cross-attention mechanism enhances focus on lesion-related anatomical regions; and a lightweight classification head reduces computational overhead to meet clinical real-time requirements. Experimental results on MIMIC-CXR and CheXpert datasets show that SCH-Net achieves 91.5% accuracy and 94.5% AUROC on MIMIC-CXR, and 88.8% accuracy and 91.8% AUROC on CheXpert. This work provides a reliable automated auxiliary tool for clinical chest X-ray interpretation, effectively solving key pain points in clinical manual diagnosis.
Xu et al. (Thu,) studied this question.