Automated Analysis of Abdominal Aortic Calcification in Vertebral Fracture Assessment Images

Abstract

Cardiovascular diseases are the most common cause of death globally. For more than half of those who die of a cardiovascular event, the disease has been clinically silent until that point, indicating a need for more targeted intervention. Abdominal aortic calcification (AAC) is an independent predictor of CVD and can be used as a measure of atherosclerotic extent within the arterial system, allowing more accurate risk stratification and monitoring ahead of a major cardiovascular event. Dual energy X-ray absorptiometry (DXA) vertebral fracture assessment (VFA), performed on a densitometer can visualise calcifications in the abdominal aorta. These images represent an opportunity to obtain clinically informative data on cardiovascular risk in a noninvasive manner. Despite these advantages, AAC is time consuming to annotate and not routinely reported; it is not commonly used to affect treatment decisions. This work investigates the automation of AAC measurement in VFA images. Approaching from the perspective of a semantic segmentation problem, two major strategies are compared to automatically identify AAC. Both random forest classification and convolutional neural networks are applied to the problem of AAC segmentation on VFA images for the first time. Additionally, an automated method to locate the abdominal aorta within VFA images using skeletal landmarks, and subdivide the aorta to produce clinically informative semi-quantitative AAC scores is presented, unique in VFA images. This is the first deep learning work in this area, and this segmentation strategy is demonstrated to outperform the random forest, and previous work on AAC segmentation in other x-ray images. This work also presents the first automated attempt at recreating a semi-quantitative clinical measure of AAC. Automated scoring shows good correlation with expert scoring of images, indicating the potential for its use as a clinically informative screening tool.

Date of Award	31 Dec 2020
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Timothy Cootes (Supervisor) & Carole Twining (Supervisor)