High-Risk Carotid Lesion Segmentation: Advancing Stroke Risk Detection With Deep Learning

Abstract

Introduction:

While the benefits of carotid surgery for symptomatic carotid artery disease are well-established, the management of asymptomatic carotid lesions remains controversial, with wide variation in clinical practice. Recent studies suggest that plaque characteristics, such as thrombus versus calcium content, may predict cerebral events more accurately than the degree of stenosis alone. This study investigates the feasibility of segmenting carotid lesions on computed-tomography (CT) angiography using artificial intelligence (AI), and evaluates differences in plaque composition between symptomatic and asymptomatic patients.

Methods:

Carotid plaques were analyzed using 2 segmentation approaches: physician-controlled manual segmentation and fully automated segmentation with the AI-based software PRAEVAorta2 (Nurea). Thrombus content, calcium burden, and residual lumen were analyzed and compared between the 2 techniques. The AI-based software was pre-trained on 19 CT angiograms. Sensitivity, specificity, Dice similarity coefficient (DSC), and volumetric similarity were calculated to evaluate the performance of both methods. A total of 156 patients who underwent carotid artery surgery between February 2019 and February 2022 were included in the analysis, comprising 81 symptomatic and 75 asymptomatic lesions.

Results:

The DSC between fully automatic segmentation and physician-controlled manual segmentation was strong for lumen (0.83), calcification (0.68) and plaque (0.60) assessments but weaker for thrombus (0.33). Volume similarity, intra- and inter-observer reliability were high, with correlation coefficients of 0.98 for intra-observer, 1.00 for inter-observer analyses and 0.86 for fully automatic vs physician-controlled manual segmentation. Symptomatic carotid lesions exhibited significantly larger thrombus-to-total volume ratios (p<0.0001), higher raw thrombus volumes (p<0.0001), and greater total lesion volumes (p=0.003) than asymptomatic lesions. Conversely, asymptomatic lesions demonstrated higher calcification-to-total volume ratios (p=0.002).

Conclusion:

This study demonstrates the potential of PRAEVAorta2 to automate carotid lesion analysis, offering promise for identifying high-risk asymptomatic plaques and therefore aid surgical decision-making. Symptomatic carotid lesions displayed higher thrombus volume, lower calcium content, and larger plaque volumes than asymptomatic lesions.

Clinical Impact

This study introduces an AI-based tool, PRAEVAorta2, capable of automatically segmenting and quantifying carotid plaque components on CT angiography. By distinguishing between thrombus and calcification, the tool provides a more nuanced assessment of plaque vulnerability beyond stenosis grading. Clinically, this innovation could enhance risk stratification in asymptomatic carotid stenosis, supporting more individualized decisions for surgery. The demonstrated correlation between thrombus burden and symptoms highlights the potential to identify high-risk plaques before neurological events occur. This advancement may bridge the gap between imaging and clinical decision-making, promoting proactive and targeted management of carotid artery disease.