A Transformer-Based Method for Automated Segmentation and Quantitative Evaluation of Lipid Plaques in IVOCT Images

The formation and rupture of coronary lipid-rich plaques represent key pathophysiological mechanisms underlying acute coronary syndrome, and their accurate identification is of substantial clinical value for guiding interventional strategies and assessing prognosis. Intravascular optical coherence tomography (IVOCT), owing to its high spatial resolution, enables detailed characterization of plaque microstructures. However, manual interpretation of IVOCT images is time-consuming and highly subjective, highlighting the urgent need for efficient and accurate automated analysis methods. This study proposes an automated framework for lipid plaque segmentation and quantitative assessment based on a Transformer architecture. Specifically, a TransDeepLab model integrating Swin Transformer with DeepLabv3+ is developed. In addition, a lumen-boundary-guided cropping augmentation strategy and a global–local attention mechanism (GLA) are introduced to enhance the model’s sensitivity to lesion-specific details. Based on the obtained segmentation results, key quantitative indicators, including lipid arc, fibrous cap thickness, plaque area, and plaque length, are further calculated automatically. The reliability and clinical applicability of the proposed method are validated through Bland–Altman agreement analysis against manual annotations. Experimental results demonstrate that the proposed approach outperforms existing mainstream models across multiple evaluation metrics, suggesting its promising potential for clinical application.