OCT Dataset for Segmentation of Atherosclerotic Plaque Morphological Features

Objectives: The primary goal of this dataset is to enable the automated segmentation and quantification of atherosclerotic plaque features in OCT images. Cardiovascular disease, with atherosclerosis at its core, remains a global health challenge. Accurate identification of vulnerable plaques is crucial for preventing acute cardiovascular events such as myocardial infarction and stroke. OCT imaging provides high-resolution insights into plaque morphology but is often constrained by manual interpretation challenges. This dataset, curated with diverse annotations of key plaque morphological features, aims to facilitate the development and evaluation of machine learning models for precise plaque analysis. By advancing segmentation capabilities, this dataset contributes to improved diagnostics and therapeutic strategies in cardiovascular care. Ethical Approval: The dataset complies with ethical standards, adhering to the Declaration of Helsinki. Ethical approval was granted by the Local Ethical Committee of the Research Institute for Complex Issues of Cardiovascular Diseases (Kemerovo, Russia) under protocol code 2022/06 (approved on June 30, 2022). All participants provided informed consent. Data collection involved patients aged 18 years or older, ensuring balanced gender representation and inclusion of various comorbid conditions for comprehensive clinical relevance (refer to Table 1). Description: The dataset consists of OCT images acquired from 103 patients across two cardiovascular research centers. These images, collected over one year, represent a diverse array of imaging devices and patient demographics. The dataset includes 25,698 annotated slices, each capturing key plaque morphological features. These features include lumen (LM), fibrous cap (FC), lipid core (LC), and vasa vasorum (VV). The images vary in dimensions from 704 x 704 to 1024 x 1024 pixels, reflecting differences in anatomical characteristics and imaging conditions. Annotations were performed using Supervisely, with meticulous double-verification processes to ensure accuracy. Annotation Method: Two cardiologists annotated the dataset, identifying plaque features using binary masks. The annotations underwent a review and double-verification by a senior cardiologist and technical specialist, enhancing precision and consistency. The morphological features segmented include the vascular lumen, fibrous cap, lipid core, and vasa vasorum, each providing critical insights into plaque stability and cardiovascular risk. Dataset Split: A 5-fold cross-validation technique was employed for dataset splitting, ensuring robust model evaluation while preventing data leakage. Approximately 80% of images were allocated for training in each fold, with the remaining 20% reserved for testing (refer to Table 2). This method allowed a balanced and comprehensive assessment of segmentation performance across the dataset. 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. Baseline characteristics of patients included in the study. Parameter Value Sex: Male, n (%) 77 (74.7) Female, n (%) 26 (25.3) Median Age, years [min – max] 69 [43 – 83] Arterial hypertension, n (%) 92 (89.3) Diabetes Mellitus, n (%) 22 (21.4) Myocardial Infarction, n (%) 22 (21.4) Polyvascular Disease, n (%) 29 (28.2) Angina Pectoris: Silent ischemia, n (%) 9 (8.7) Functional class 1, n (%) 24 (23.3) Functional class 2, n (%) 55 (53.4) Functional class 3, n (%) 15 (14.6) Table 2. Image and plaque morphological feature distributions across folds and subsets. Fold Subset LM FC LC VV Total objects Total images 1 Train 17264 5610 5576 328 28778 16901 1 Test 4544 1616 1616 122 7898 4492 2 Train 17554 5709 5690 237 29190 17207 2 Test 4254 1517 1502 213 7486 4186 3 Train 17220 5600 5565 407 28792 16962 3 Test 4588 1626 1627 43 7884 4431 4 Train 17813 5724 5686 416 29639 17473 4 Test 3995 1502 1506 34 7037 3920 5 Train 17381 6261 6251 412 30405 17029 5 Test 4427 965 941 38 6371 4364# 研究目标 本数据集的核心目标是实现光学相干断层成像（Optical Coherence Tomography, OCT）图像中动脉粥样硬化斑块特征的自动分割与量化。以动脉粥样硬化为核心病理的心血管疾病仍是全球公共卫生挑战。精准识别脆弱斑块对于预防心肌梗死、中风等急性心血管事件至关重要。OCT成像可提供斑块形态的高分辨率洞察，但手动解读存在诸多局限。本数据集针对关键斑块形态学特征进行了多样化注释，旨在推动用于精准斑块分析的机器学习模型开发与评估。通过提升分割性能，本数据集将助力心血管护理领域诊断与治疗策略的优化。 ## 伦理合规说明 本数据集符合赫尔辛基宣言的伦理标准，已通过俄罗斯克麦罗沃心血管疾病复杂问题研究所地方伦理委员会审核批准，审批编号为2022/06（批准日期：2022年6月30日）。所有受试者均已签署知情同意书。数据采集对象为18岁及以上患者，兼顾性别比例均衡，并纳入多种合并症以确保研究的临床全面性（详见表1）。 ## 数据集概况 本数据集包含来自两家心血管研究中心的103名患者的OCT图像，采集周期为一年，涵盖多样化的成像设备与患者人口统计学特征。数据集共计25698张带注释的图像切片，每张切片均标注了关键斑块形态学特征，包括管腔（lumen, LM）、纤维帽（fibrous cap, FC）、脂质核心（lipid core, LC）与血管滋养管（vasa vasorum, VV）。图像尺寸范围为704×704至1024×1024像素，反映了解剖特征与成像条件的差异。注释工作基于Supervisely平台完成，并经过严格的双重验证流程以保障标注准确性。 ## 注释流程 本数据集由两名心脏病医师完成标注，采用二值掩码识别斑块特征。标注结果经资深心脏病医师与技术专家双重审核验证，进一步提升了标注精度与一致性。本次待分割的形态学特征包括血管管腔、纤维帽、脂质核心与血管滋养管，上述特征均可为斑块稳定性评估与心血管风险分层提供关键依据。 ## 数据集划分方案 本数据集采用5折交叉验证（5-fold cross-validation）策略进行划分，以保障模型评估的稳健性并避免数据泄露。每一折中约80%的图像用于模型训练，剩余20%用于测试（详见表2）。该划分方案可实现对全数据集分割性能的平衡且全面的评估。 ## 研究资源获取 有关本研究的更多信息，包括整理后的源代码、数据集详情与训练模型，可通过以下渠道获取： - 源代码：https://github.com/ViacheslavDanilov/oct_segmentation - 数据集：https://doi.org/10.5281/zenodo.14478209 - 训练模型：https://doi.org/10.5281/zenodo.14481678 --- ### 表1 研究纳入患者的基线特征 | 参数 | 数值 | | --- | --- | | 性别 | | | 男性[例数（占比）] | 77 (74.7%) | | 女性[例数（占比）] | 26 (25.3%) | | 年龄中位数[最小值–最大值] | 69 [43–83] | | 高血压[例数（占比）] | 92 (89.3%) | | 糖尿病[例数（占比）] | 22 (21.4%) | | 心肌梗死[例数（占比）] | 22 (21.4%) | | 多血管病变[例数（占比）] | 29 (28.2%) | | 心绞痛： | | | 无症状心肌缺血[例数（占比）] | 9 (8.7%) | | 功能分级1级[例数（占比）] | 24 (23.3%) | | 功能分级2级[例数（占比）] | 55 (53.4%) | | 功能分级3级[例数（占比）] | 15 (14.6%) | --- ### 表2 各折次与子集的图像及斑块形态学特征分布 | 折次 | 子集 | LM | FC | LC | VV | 总目标数 | 总图像数 | | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | 训练集 | 17264 | 5610 | 5576 | 328 | 28778 | 16901 | | 1 | 测试集 | 4544 | 1616 | 1616 | 122 | 7898 | 4492 | | 2 | 训练集 | 17554 | 5709 | 5690 | 237 | 29190 | 17207 | | 2 | 测试集 | 4254 | 1517 | 1502 | 213 | 7486 | 4186 | | 3 | 训练集 | 17220 | 5600 | 5565 | 407 | 28792 | 16962 | | 3 | 测试集 | 4588 | 1626 | 1627 | 43 | 7884 | 4431 | | 4 | 训练集 | 17813 | 5724 | 5686 | 416 | 29639 | 17473 | | 4 | 测试集 | 3995 | 1502 | 1506 | 34 | 7037 | 3920 | | 5 | 训练集 | 17381 | 6261 | 6251 | 412 | 30405 | 17029 | | 5 | 测试集 | 4427 | 965 | 941 | 38 | 6371 | 4364 |