Automated scheme of plaque classification based on segmentation in carotid ultrasound images using transformer approach.

Purpose: Carotid plaque is a major risk factor for cerebral infarction. Ultrasonography (US) is extensively used for screening carotid plaque, but US images contain more noise than those of computed tomography and magnetic resonance imaging, and the edges of the plaque regions are unclear. In addition, B-mode echogenicity evaluation, which is important for plaque risk assessment, has challenges involving the subjectivity of the evaluator. Although previous studies on carotid plaque assessment have included plaque segmentation, most studies involved manual operations. In this study, we propose an automated scheme of plaque classification based on segmentation in carotid US images using the transformer approach, to resolve the issues of previous studies and to perform plaque echogenicity classification.

Methods: The B-mode video captured in the long-axis cross-section was converted to still images, and region extraction and echogenicity classification were performed using TransUNet. The results of the TransUNet output and US images were fed into the Vision Transformer (ViT) for classification into hypoechoic or isoechoic-hyperechoic plaques.

Results: The Dice index, which indicates the accuracy of plaque region extraction, was 0.592. The Dice indices by echogenicity were 0.200, 0.493, and 0.542 for the hypoechoic, isoechoic, and hyperechoic regions, respectively. The balanced accuracy, which indicates the classification accuracy, was 79.6%. The correct classification rate for high-risk hypoechoic plaques was 95.2%.

Conclusion: These results suggest that the proposed method is useful for evaluating the echogenicity classification of carotid artery plaques.