Wu Fang, Wei Hai-Ning, Zhang Miao, Ma Qing-Feng, Li Rui, Lu Jie
Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, 45 Changchun Street, Beijing, 100053, China.
Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, 45 Changchun Street, Beijing, 100053, China.
Transl Stroke Res. 2025 Mar 19. doi: 10.1007/s12975-025-01345-1.
The rupture of vulnerable plaques is the principal cause of luminal thrombosis in acute ischemic stroke. The identification of plaque features that indicate risk for disruption may predict cerebrovascular events. Here, we aimed to build a high-risk intracranial plaque model that differentiates symptomatic from asymptomatic plaques using radiomic features based on high-resolution magnetic resonance imaging (HRMRI). One hundred and seventy-two patients with 188 intracranial atherosclerotic plaques (100 symptomatic and 88 asymptomatic) with available HRMRI data were recruited. Clinical characteristics and conventional plaque features on HRMRI were measured, including high signal on T1-weighted images (HST1), the degree of stenosis, normalized wall index, remodeling index, and enhancement ratio (ER). Univariate and multivariate analyses were performed to build a traditional model to differentiate between symptomatic and asymptomatic plaques. Radiomic features were extracted from pre-contrast and post-contrast HRMRI. A radiomic model based on HRMRI was constructed using random forests, ridge, least absolute shrinkage and selection operator, and deep learning (DL). A MIX model was constructed based on the radiomic model and the traditional model. Gender, HST1, and ER were associated with symptomatic plaques and were included in the traditional model, which had an area under the curve (AUC) of 0.697 in the training set and 0.704 in the test set. The radiomic model achieved an AUC of 0.982 in the training set and 0.867 in the test dataset for identifying symptomatic plaques. In the training set, the MIX model showed an AUC of 0.977. In the test set, the MIX model exhibited an improved AUC of 0.895, which outperformed the traditional model (p = 0.032). Radiomic analysis based on DL and machine learning can accurately identify high-risk intracranial plaques.
易损斑块破裂是急性缺血性卒中管腔内血栓形成的主要原因。识别提示斑块破裂风险的特征可能有助于预测脑血管事件。在此，我们旨在构建一个基于高分辨率磁共振成像（HRMRI）的放射组学特征的颅内高危斑块模型，以区分有症状斑块和无症状斑块。招募了172例患有188个颅内动脉粥样硬化斑块（100个有症状斑块和88个无症状斑块）且有可用HRMRI数据的患者。测量了临床特征和HRMRI上的传统斑块特征，包括T1加权图像上的高信号（HST1）、狭窄程度、标准化管壁指数、重塑指数和强化率（ER）。进行单因素和多因素分析以构建区分有症状斑块和无症状斑块的传统模型。从增强前和增强后的HRMRI中提取放射组学特征。使用随机森林、岭回归、最小绝对收缩和选择算子以及深度学习（DL）构建基于HRMRI的放射组学模型。基于放射组学模型和传统模型构建了一个混合模型。性别、HST1和ER与有症状斑块相关，并被纳入传统模型，该模型在训练集中的曲线下面积（AUC）为0.697，在测试集中为0.704。放射组学模型在训练集中识别有症状斑块的AUC为0.982，在测试数据集中为0.867。在训练集中，混合模型的AUC为0.977。在测试集中，混合模型的AUC提高到0.895，优于传统模型（p = 0.032）。基于深度学习和机器学习的放射组学分析能够准确识别颅内高危斑块。
