(P270) A NOVEL APPROACH IN VASCULAR IMAGING: 3D RECONSTRUCTION FROM 2D ULTRASOUND VIA ADVANCED DEEP LEARNING

Background: Stroke remains a leading global cause of death and disability, often attributed to the rupture of high-risk atherosclerotic plaques in the carotid arteries. The standard approach to assessing stroke risk involves 2D ultrasound examination, which, due to operator dependency, a lack of 3D plaque structure assessment, and subjective interpretation, leads to 28% misclassifications and >40% re-scans1,2. To address these challenges, we developed Vaso3DTM, an artificial-intellgence (AI)-powered software that leverages 2D ultrasound images to generate precise 3D models of the carotid arteries and automate the detection of high-risk plaques.

METHODS AND RESULTS: Vaso3DTM leverages a multi-class U-Net AI model derived from >800 2D ultrasound images acquired from 113 North American patients (≥45 years old) with atherosclerotic risk factors. Independent vascular experts (n=3) annotated the images using CoreSlicer, labeling important carotid artery structures (outer wall, inner wall, lumen, plaque). 3D reconstructions were performed by combining the 2D ultrasound images with associated positional data acquired via an electromagnetic sensor (Northern Digital Inc, Canada). This was followed by automated vessel diameter, plaque stenosis, and risk classification (mild, moderate, severe) measurements. We validated the 3D ultrasound reconstruction and automated analysis methods against a carotid artery phantom with a predefined 70% stenosis (R.G. Shelley Ltd, Ontario, Canada)3. Prospectively, 13 patients were recruited to the study to compare the performance of Vaso3DTM to the standard of care. The AI models performed with an accuracy of >90% in detecting 3 classes: carotid artery outer wall, inner wall, and atherosclerotic plaque. 3D models of the three carotid artery classes were successfully generated by integrating 2D ultrasound data with positional information. Key vascular metrics, including artery stenosis and lumen diameters, were derived from reconstructions and validated to the carotid artery phantom system demonstrating 99.9% agreement to known stenosis using a carotid phantom. In 13 prospective patients, Vaso3DTM achieved a 90% reduction in ultrasound scan with the same diagnostic accuracy when compared to a vascular radiologist with >15 years of experience.

Conclusion: Our ground-breaking innovation significantly streamlines the carotid imaging workflow, achieving a remarkable ten-fold improvement in scan time efficiency while maintaining comparable accuracy to experts. Additionally, diagnostic information can be given to the health care provider in real-time. Through the integration of AI, our solution eliminates subjectivity in manual interpretation, ensuring consistent and reliable results within a fraction of the time traditionally required.