Segmentation and quantification of atherosclerotic plaques in optical coherence tomography

Cardiovascular disease remains a leading cause of mortality worldwide, with atherosclerosis at its core. Accurate identification of vulnerable plaques is critical for preventing acute cardiovascular events. While optical coherence tomography (OCT) provides high-resolution imaging of plaque features, manual analysis is labor-intensive and operator-dependent. This study addresses the need for automated, accurate segmentation of atherosclerotic plaques in OCT pullbacks using advanced machine learning (ML) techniques.

A comprehensive multi-center dataset of OCT pullbacks, encompassing 103 patients and annotated for key plaque morphological features (lumen, fibrous cap, lipid core, and vasa vasorum), was used to tune, train and evaluate nine ML models, including U-Net, U-Net++, DeepLabV3, DeepLabV3+, FPN, LinkNet, PSPNet, PAN, and MA-Net. To address dataset imbalances and optimize performance for each plaque feature, we introduced a hybrid segmentation strategy: single-class models were deployed for highly prevalent features (e.g., lumen) and underrepresented classes (e.g., vasa vasorum), while a multi-class model targeted morphologically complex features with overlapping boundaries (e.g., fibrous cap and lipid core). Hyperparameter tuning was performed using Bayesian optimization, and segmentation accuracy was assessed with the Dice Similarity Coefficient (DSC) and other metrics.

The models achieved high segmentation accuracy for lumen (DSC: 0.987), fibrous cap (DSC: 0.736), and lipid core (DSC: 0.751), demonstrating the potential of leveraging ML techniques to enhance OCT's diagnostic capabilities. While lumen segmentation showed exceptional precision, the moderate accuracy for fibrous cap and lipid core highlights challenges with complex morphologies. Satisfactory results for vasa vasorum (DSC: 0.610) suggest areas for further refinement. By integrating these models into a weighted ensemble, taking into account class prevalence and model confidence, the combined system achieved a weighted DSC of 0.882 across all plaque features, a significant improvement over individual models. These findings confirm the hybrid strategy's ability to balance computational efficiency with accuracy through rigorous optimization, tailored model selection, and ensemble integration.

This study presents a robust ML-driven framework for automated OCT segmentation that uses a hybrid approach and weighted ensemble learning to address class imbalance and feature complexity, significantly improving the accuracy and efficiency of atherosclerotic plaque analysis. The findings suggest potential clinical implications, including improved detection of high-risk plaques and enhanced decision-making in cardiovascular care. However, further prospective validation is required before clinical adoption. Future research should focus on expanding datasets, integrating multimodal imaging, and refining models for real-time clinical use, paving the way for transformative advancements in cardiovascular diagnostics.