GuideNet: Learning Inter- Vertebral Guides in DXA Lateral Spine Images

Abstract

Cardiovascular Disease (CVD) is the leading cause of death worldwide. Calcification in the Abdominal Aorta is a stable marker of CVD development and, hence, it's early detection is considered crucial to saving lives. Imaging techniques such as Computed Tomography (CT) and Digital X-Ray Imaging can be used to accurately predict and localize Abdominal Aortic Calcification (AAC), however, these methods are not only expensive but also expose the patients to high ionizing radiation. In contrast, Dual Energy X-ray Absorptiometry (DXA) is an efficient, cost-effective and low radiation exposure-based imaging alternative, but with challenges like low resolution and vague vertebral boundaries. This poses a bottleneck in identifying the vertebrae and their boundaries which is crucial in manual as well as automatic scoring of AAC from DXA scans. In this paper, we address this research gap by proposing a framework which first localizes the vertebrae T12, L1, L2, L3, L4 and L5 and then generates Inter-Vertebral Guides (IVGs) between them. Our deep model is trained on lateral view DXA spine images and shows promising results in generating IVGs with high accuracy, which we believe can greatly reduce inter-observer variability in AAC scoring in DXA imaging domain.