Automatic Abdominal Aortic Aneurysm Detection from Ultrasound Imaging using Deep Learning

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
An abdominal aortic aneurysm (AAA) is an enlargement of the abdominal aorta. If an AAA ruptures, it leads to death in 48.5% to 81% of the cases. Detection and monitoring of AAAs is therefore vital and is currently performed by a trained sonographer at a hospital. A general practitioner is not able to perform an ultrasound, since it requires months of training. In this study, we present a method based on deep learning to automatically detect an aorta from ultrasound (US) imaging and to automatically measure the aortic diameter, so untrained people would be able to measure the aortic diameter without extensive training. The method consists of two steps. In the rst step, a deep learning model with a U-Net architecture segments the aorta for each acquired US frame. In the second step, connected-component labeling (CCL) is used to nd the segmented aorta, and a direct least-squares ellipse t is performed to measure the aortic diameter. Data from 100 patients was acquired. A handheld US device was used to make an axial sweep of 7 seconds from the xiphoid process up until the umbilicus. Data of 80 patients was used to train the algorithms. 20 patients were used as a test set which showed a median Dice of 0.88 (IQR = 0.78 - 0.92). The segmentation model was included into a smartphone application, which was used to acquire data from 44 additional patients which also received a computed tomography (CT) as ground truth. The results show that the CT-US maximum diameter di erences had a median of 6.0 mm (IQR = 4.0 - 9.6 mm) and 73.8% of the measurements fell within within the clinically acceptable limits of agreement of 5 mm.
Faculteit der Sociale Wetenschappen