Deep Learning Models for Segmentation of Atherosclerotic Plaque Features in OCT Images

Introduction: We present a collection of deep learning models developed for the automated segmentation and quantification of plaque morphological features (PMFs) of atherosclerotic plaques in optical coherence tomography (OCT) images. These models enable precise delineation of plaque features such as the vascular lumen (LM), fibrous cap (FC), lipid core (LC), and vasa vasorum (VV), providing critical insights into plaque composition and vulnerability, which are essential for cardiovascular diagnostics and intervention planning.

Dataset: The dataset utilized in this study consists of OCT images obtained from 103 patients across multiple cardiovascular research centers. It includes 25,698 annotated slices representing plaque morphological features. The annotations highlight the lumen, fibrous cap, lipid core, and vasa vasorum, with verification by expert cardiologists to ensure accuracy and consistency. The dataset's diversity in imaging conditions and patient demographics ensures robust model evaluation. For detailed information about the dataset, please refer to the dataset documentation at https://doi.org/10.5281/zenodo.14478209.

Model Description: We employed nine advanced deep learning architectures — U-Net, U-Net++, DeepLabV3, DeepLabV3+, FPN, LinkNet, PSPNet, PAN, and MA-Net — to address the segmentation of atherosclerotic plaque features. To enhance performance, we applied a hybrid segmentation strategy, which involves designing specialized models for each plaque feature based on their prevalence and complexity:

• Lumen Segmentation: A single-class model trained exclusively for the lumen, leveraging its high representation in the dataset.
• Fibrous Cap and Lipid Core Segmentation: A two-class model trained jointly on these features due to their shared morphological characteristics and moderate dataset representation.
• Vasa Vasorum Segmentation: A separate single-class model focused on this rare feature to better capture its unique characteristics.

Extensive hyperparameter optimization using Bayesian techniques, combined with augmentation strategies like perspective distortion and brightness adjustments, enhanced the robustness of the models.

Results: The models exhibited varying levels of accuracy across features. U-Net++ excelled in lumen segmentation, achieving a Dice Similarity Coefficient (DSC) of 0.987. This performance reflects the advantage of a single-class model dedicated to the lumen, leveraging its high representation in the dataset to capture clear and distinct boundaries effectively.

For fibrous cap and lipid core segmentation, LinkNet performed strongly with DSCs of 0.736 and 0.751, respectively. These results highlight the effectiveness of a two-class model trained jointly on these features, allowing it to better handle their shared morphological characteristics, such as overlapping and intricate boundaries.

The segmentation of vasa vasorum, a rare and subtle feature, was tackled using a dedicated single-class U-Net model, achieving a DSC of 0.610. Despite the inherent challenges due to its sparse representation in the dataset, this approach ensured focused learning on the vasa vasorum's unique structure, yielding promising results.

An ensemble approach combining U-Net++, LinkNet, and U-Net further improved overall segmentation accuracy, achieving an average DSC of 0.882. These results demonstrate the effectiveness of tailoring model architectures and strategies to the specific requirements of each plaque feature. Performance metrics, including precision, recall, and IoU, are detailed in Table 1, with training dynamics and segmentation outcomes visualized in Figures 1 and Figure 2.

Access to the Study: Further information about this study, including curated source code, dataset details, and trained models, can be accessed through the following repositories:

• Source Code: https://github.com/ViacheslavDanilov/oct_segmentation
• Dataset: https://doi.org/10.5281/zenodo.14478209
• Models: https://doi.org/10.5281/zenodo.14481678

Table 1. Segmentation performance metrics for each plaque morphological feature, averaged over 5 folds during cross-validation.