Deep Learning-Based Detection of Osteoporosis and Osteopenia Using Wrist X-rays: A Retrospective Analysis with Internal Validation (Preprint)

Abstract

<sec> <title>BACKGROUND</title> Osteoporosis is a common degenerative disorder in older adults that can lead to fractures, necessitating early diagnosis. </sec> <sec> <title>OBJECTIVE</title> This study trained and validated deep learning (DL) models to detect osteoporosis and osteopenia through wrist X-rays, aiming to offer a more convenient and time-saving alternative to the established measures for bone mineral density (BMD) assessment. </sec> <sec> <title>METHODS</title> This retrospective study gathered data from patients aged 20 years and older who underwent wrist X-rays and BMD testing via dual-energy X-ray absorptiometry (DEXA) at Tri-Service General Hospital in Taiwan. The Residual Network (ResNet) model were used to train and validate a deep learning model on wrist X-rays, differentiating osteoporosis and osteopenia from normal bone conditions. </sec> <sec> <title>RESULTS</title> A total of 3,825 wrist images with corresponding BMD measurements were utilized to train a DL model. The model demonstrated excellent performance in discriminating normal BMD from osteopenia/osteoporosis, achieving an area under the ROC curve (AUC) of 0.9696, with a sensitivity of 99.5% and a specificity of 86.1%. Further, in the subset of training data which excluded wrist images of osteopenia, the model excelled in differentiating osteoporosis from normal BMD, yielding an AUC of 0.9741, with 90.2% sensitivity and 96.5% specificity. </sec> <sec> <title>CONCLUSIONS</title> Our DL model excels at distinguishing normal BMD from osteoporosis using wrist X-rays. It provides a time-efficient and convenient alternative to traditional BMD measurements via DEXA, making it particularly beneficial in resource-limited environments. By integrating this model into standard wrist X-ray exams, we enable more accessible assessments, which support early detection and the development of effective treatment strategies, lowering resource expenditures. </sec>

Ähnliche Arbeiten