Development and External Validation of a 3D Convolutional Neural Network for Automated Detection of Vertebral Fractures in CT Scans

Osteoporosis affects approximately 200 million people globally, resulting in 9 million fragility fractures annually. Vertebral fractures (VF) are the hallmark of the disease; the convey significant morbidity yet their identification lags. The growing number of abdomen and chest CT scans performand for unrelated indications provides an opportunity for the identification and reporting of VF. Convolutional Neural Networks (CNN) models have been successfully applied to patient images for various medical applications and a number of medical devices have been approved by health authorities for clinical use are leveraging this technology.

In this research project, designed and sponsored by UCB Pharma, we develop a 3D CNN model that is capable of automatically detecting VFs in abdomen and chest CT scans in men and women, age 50 years or older. We perform extensive external validation studies to examine the ability of the model to generalize on a number of representative data sets. Furthermore, we prepare for the regulatory approvals and commercialization of this CNN model in order to finally achieve an impact for patients living with osteoporosis.