Carotid Vessel Wall and Plaque Segmentation and Characterization in 3D Ultrasound Images by a Dual-stream Centreline-guided Network and Domain Adaptation
- Bernard Chi Yuen CHIU (Principal Investigator / Project Coordinator)Department of Electrical Engineering
- Rcza AZARPAZHOOH (Co-Investigator)
- John David SPENCE (Co-Investigator)
DescriptionIdentification of patients with high stroke risk is important as dietary and medical intervention can reduce their stroke risk significantly. Our group pioneered the use of three-dimensional ultrasound (3DUS) carotid plaque and vessel wall measurements in risk stratification and sensitive treatment monitoring. However, the requirement of manual plaque and vessel wall segmentation limits the clinical utility of 3DUS techniques, as manual segmentation is time-consuming and prone to observer variability. We therefore propose to develop a fully automated approach to vessel wall and plaque segmentation based on machine learning. Most existing automated vessel wall segmentation techniques delineate only the common but not the internal carotid artery (CCA and ICA, respectively). ICA segmentation is important for accurate risk stratification as plaques are more prevalent in ICA than CCA. Besides, these algorithms require users to provide seed points or a region-of-interest (ROI) for initialization, which is time-consuming even for a medium-scale study of ~300 patients. Therefore, there is a need to develop efficient algorithms for vessel wall and plaque segmentation at the CCA and ICA. The central hypothesis of our approach is that knowledge of the artery centerline will enable a neural network to focus on the artery without the need for manual ROI definition. But, because centerline extraction is itself a difficult problem, we propose to introduce a neural network capable of generating a “centerline heatmap” that indicates high probability regions for the centerline location. The heatmap can then be combined with the 3DUS image in a dual-stream centerline-guided segmentation network that is fully automated. We will also develop an adversarial network to tune the segmentation network trained in images acquired in one patient cohort to a different cohort (target cohort), which has a different level of disease burden, without manual segmentation in the target cohort. Plaque volume and texture measured from manual segmentation were shown to predict cardiovascular events in the target cohort consisting of ~300 patients. Besides evaluating segmentation accuracy, we will compare the event predicting ability of plaque volume and texture measured from automated segmentation with the corresponding measurements from manual segmentation. The proposed framework will improve the clinical utility of 3DUS carotid measurements in multi-center trials and in clinical practice. The adversarial network allows calibration for differences in 3DUS scanning parameters and patient characteristics. The domain adaption ability of the framework will especially benefit medical clinics with less availability of expertise in 3DUS interpretation.
|Effective start/end date||1/01/23 → …|