Vertebral Compression Fractures (VCF) are brittle fractures that result from the deterioration of the bone structure, mainly when the anterior portion of the vertebra (the vertebral body) breaks and loses height, while the posterior portion is maintained [1,2]. Based on the percentage of vertebral height loss (VHL), fracture severity is characterized as grade 1 (20% - 25% collapse), grade 2 (26% - 40% collapse), or grade 3 (>40% collapse), according to the Genant semiquantitative classification scheme [3]. They are mostly caused by osteoporosis, but they can also arise from trauma, infection or neoplasm [1,4].

VCF is one of the most common healthcare issues worldwide, accounting for 1.4 million new fractures annually [5]. They are responsible for a 2-fold increase in mortality and a 3-fold increase in subsequent fractures [6]. Approximately, 30–50% of women and 20–30% of men will develop vertebral fractures during their lives with half experiencing multiple VCF [7]. In addition, VCF tend to remain clinically silent and only 15–30% are symptomatic, hence, early signs are frequently undetected leading to high underdiagnosis rates [4]. In fact, more than two-thirds of patients are incidentally diagnosed with VCF [1]. 

In order to reduce underdiagnosis rates, opportunistic screening aims to identify individuals at high risk of fractures by retrieving bone mineral density data from CT scans performed for other reasons, so that no additional costs or radiation exposure are required [8]. Automated tools for VCF detection could be integrated into an opportunistic detection framework to assess CT images and reduce the percentage of VCF undetected by clinicians. Hence, the purpose of this study was to evaluate the performance of a deep learning (DL)-based application designed to automatically screen CT scans with unsuspected VCF.