Automated Deep-learning-based Vertebral Body Localization and Instance Segmentation for Osteoporosis Assessment using CT

Abstract

Osteoporosis is an important disorder that is underdiagnosed. We aim to develop a robust automated CT-based vertebral label-independent approach for localization and instance segmentation of the vertebral bodies (VB) which permits assessment of bone mineral density (BMD) and generates VB center data for subsequent fracture assessment tools. We utilize an nnU-Net adapted for segmentation of the surface vs interior of the VBs vs background. This allows delineation of individual VBs, segmentation of the cortical surface encompassing the cancellous bone for BMD assessment, and localization of VB centers. After training the performance was evaluated on an external test data set. Our approach detected 97.5% of the vertebral bodies showing robustness to spinal degenerations like osteophytes. VB centers were determined with residual errors of 3.5±0.9 mm, sufficiently accurate as input for our fracture detection tool. BMD evaluated in Hounsfield Units (HU) correlated with ground truth values with r2 = 0.96, RMS error = 11.3 HU, sufficiently accurate for automated diagnosis and CT-based opportunistic screening for osteoporosis.