Speaker
Description
This study aimed to develop an algorithm for measuring carotid artery stenosis using both deep learning and statistical models based on carotid ultrasound images. Carotid artery stenosis is a common clinical condition characterized by the narrowing of the carotid arteries, which are the major blood vessels that supply the brain. The condition is typically caused by atherosclerosis, a buildup of plaque in the arterial walls. As the plaque grows, it can obstruct blood flow to the brain, leading to a range of symptoms such as cognitive impairments, ischemic strokes, and dementias. The most popular method of diagnosing carotid stenosis is non-invasive imaging, specifically carotid ultrasound. This method uses high-frequency sound waves to produce detailed images of the carotid arteries and can detect the presence of plaque buildup and the degree of stenosis. Ultrasound-based diagnosis does not involve risks such as exposure to ionizing radiation, making it useful for routine examination of moderate to severe patients. The major difficulty in diagnosing carotid stenosis based on ultrasound is accurately determining the degree of stenosis. Ultrasound is operator-dependent, meaning that the quality of the images and the interpretation of the results can vary depending on the skill and experience of the technician or physician performing the ultrasound. This can lead to interobserver variability in the assessment of stenosis severity, particularly in cases where the degree of stenosis is borderline or near the threshold for intervention.
Therefore, we propose an automated algorithm to evaluate the degree of stenosis of the carotid artery. We obtained cine images of carotid ultrasonography from both the right and left carotid arteries of 18 ischemic stroke patients by continuously moving from the supraclavicular to the submandibular area. We collected a total of 13,586 raw axial images from the ultrasound images of all patients, and then we labeled the segmentation of the stenosis caused by atherosclerotic plaque from the vessel wall in each axial image by two experts. To develop the segmentation algorithm from the axial images, we applied a deep neural network to segment the vessel lumen from the vessel wall and a statistical model based on Gaussian mixture to segment the atherosclerotic stenosis from the vessel lumen. The accuracy of the algorithm to segment the lumen from the carotid vessel wall was a mean of 0.923 (±SD 0.46) of the dice coefficient, and the error of estimation for the stenotic area was a mean of 0.201 (±SD 0.137).
References
None
| Keywords | Carotid Artery Stenosis, Segmentation, Algorithm, Deep Learning, Statistical Model |
|---|
