Two -stage contrastive learning framework for vertebral compression fracture screening in frontal chest X-ray

Vertebral Compression Fractures (VCFs) are early signs of osteoporosis and the most common type of osteoporotic fracture. In clinical practice, the pipeline of VCFs diagnosis consists of two examinations while most only routinely perform the first examination when the absence of clinical manifestations, leading to an underdiagnosis. In this study, we aim to explore the potential of VCFs screening in initial frontal Chest X-rays (CXRs) examination, which is more challenging for radiologists due to difficulties such as organ occlusion and inter-class similarity. We develop a two-stage framework: in the first stage, a segmentation framework and a post-processing algorithm are proposed to obtain individual vertebrae; in the second stage, the TADC-Net is proposed to perform accurate recognition of VCFs by designing a Triplet Aggregation module for contextual feature extraction and aggregation, and a Dual Contrastive Loss module for overcome inter-class similarity. Experiments conducted on three datasets collected from multi-center hospitals demonstrate that the proposed method outperforms existing approaches quantitatively and qualitatively. Moreover, a comparative analysis with two radiologists demonstrates that the proposed method exhibits optimal performance while effectively enhances radiologists' accuracy and sensitive in VCFs screening on CXRs by helping them identify additional fracture cases, verifying the potential of early screening.