AI-based Detection of Vertebral Compression Fractures within Frontal Chest Radiographs

Vertebral compression fractures (VCFs) are prevalent among older adults, especially those with osteoporosis or other bone-weakening conditions. VCFs can lead to severe pain and decreased mobility, significantly impacting the quality of life for elderly individuals. “AI      Undiagnosed or untreated VCFs can result in progressive spinal deformities, such as kyphosis (abnormal curvature of the spine), height loss, and reduced lung capacity, which can further exacerbate respiratory issues and increase the risk of falls and fractures. Furthermore, VCFs are associated with an increased mortality risk in the elderly population. 

This study aims to identify VCFs in posteroanterior (PA) chest X-rays (CXRs), focusing on the thoracolumbar junction, typically ranging from T12 to L2, where approximately 60% to 75% of VCFs occur. Given the widespread use of CXRs across various medical contexts, we hypothesize that CXR represents a promising modality for VCF screening due to its high accessibility.

We obtained 1,981 CXRs from patients diagnosed with VCFs through the Health Screening and Promotion Department of a tertiary university hospital. Given the challenge of identifying compression fractures within CXRs, we initially screened participants using thoracolumbar spine lateral X-rays to confirm VCF presence and subsequently acquired corresponding CXRs taken on the same day. These data were utilized to train a classification model based on the InceptionV3 architecture, aimed at distinguishing CXR images with and without compression fractures. To validate whether our model operates based on regions of interest (ROIs) corresponding to VCFs, we conducted a comparison between class activation maps (CAM) and clinical readings.

The area under the curve (AUC) for the binary classification task was calculated to be 0.938, with an associated accuracy of 86.45%. Further examination of the CAM revealed that our model primarily emphasized the integrity of the lumbar spine rather than focusing on the ROIs corresponding to VCFs. Also, Our analysis revealed that the model primarily focuses on assessing the integrity of the lumbar spine to differentiate normal CXRs from those with VCFs.

Our research addresses the challenge of detecting VCFs in PA CXRs, traditionally diagnosed using lateral spine X-rays, to enhance the accessibility of VCF screening. Our developed model achieved impressive performance metrics, with an AUC exceeding 0.93 and accuracy surpassing 86% in binary classification. This study represents a significant advancement in VCF detection, as it enables the identification of fractures solely through CXR imaging, which is more accessible compared to lateral spine X-rays. To our knowledge, our study is the first to employ deep learning for VCF detection in PA CXRs, further highlighting the novelty and potential utility of our approach as an efficient screening tool for VCFs.

Figure: