Sun Kui, Shi Rongchao, Yu Xinxin, Wang Ying, Zhang Wei, Yang Xiaoxia, Zhang Mei, Wang Jian, Jiang Shu, Li Haiou, Kang Bing, Li Tong, Zhao Shuying, Ai Yu, Qiu Jianfeng, Wang Haiyan, Wang Ximing
Department of General Surgery, Peking University Third Hospital, Beijing, China.
Department of Radiology, Beijing Friendship Hospital, Capital Medical University, Beijing, 100050, China.
Sci Rep. 2025 Jan 30;15(1):3743. doi: 10.1038/s41598-025-88016-1.
To develop and validate non-contrast computed tomography (NCCT)-based radiomics method combines machine learning (ML) to investigate invisible microscopic acute ischaemic stroke (AIS) lesions. We retrospectively analyzed 1122 patients from August 2015 to July 2022, whose were later confirmed AIS by diffusion-weighted imaging (DWI). However, receiving a negative result was reported by radiologists according to the NCCT images. Patients in five institutions (n = 592) were combined to generate training and internal validation sets, remaining in three institutions as external validation sets (n = 204, 53 and 273). Through a series of procedures: head alignment, co-registration of NCCT and DWI, the volume of interest delineation and feature extraction. Multiple ML models (random forest, RF; support vector machine, SVM; logistic regression, LR; multilayer perceptron, MLP) were used to discriminate microscopic AIS and non-AIS. Among 1122 patients included (760 men [67.7%]; median [range] age, 64 [21-96] years). After least absolute shrinkage and selection operator (LASSO) algorithm, 44 optimal features were remained. The radiomics combined ML models were yielded similar mean areas under the receiver operating characteristic curve of 0.808 (95% CI 0.754 to 0.861) for RF, 0.802 (95% CI 0.748 to 0.856) for radial basis kernel function-based SVM, 0.792 (95% CI 0.737 to 0.847) for MLP, 0.792 (95% CI 0.736 to 0.848) for Linear-SVM and 0.787 (95% CI 0.730 to 0.844) for LR, respectively. Combining radiomics with ML models can be an efficient, noninvasive, economical, and reliable technique for evaluating invisible microscopic AIS on NCCT and assisting radiologists to make clinical decisions.
开发并验证基于非增强计算机断层扫描（NCCT）的放射组学方法，结合机器学习（ML）来研究隐匿性微观急性缺血性卒中（AIS）病变。我们回顾性分析了2015年8月至2022年7月的1122例患者，这些患者后来通过扩散加权成像（DWI）确诊为AIS。然而，放射科医生根据NCCT图像报告结果为阴性。来自五个机构的患者（n = 592）被合并以生成训练集和内部验证集，其余来自三个机构的患者作为外部验证集（n = 204、53和273）。通过一系列步骤：头部对齐、NCCT和DWI的配准、感兴趣区勾画和特征提取。使用多种ML模型（随机森林，RF；支持向量机，SVM；逻辑回归，LR；多层感知器，MLP）来区分微观AIS和非AIS。纳入的1122例患者中（760例男性[67.7%]；年龄中位数[范围]为64[21 - 96]岁）。经过最小绝对收缩和选择算子（LASSO）算法后，保留了44个最佳特征。放射组学结合ML模型在接受者操作特征曲线下的平均面积相似，RF为0.808（95%CI 0.754至0.861），基于径向基核函数的SVM为0.802（95%CI 0.748至0.856），MLP为0.792（95%CI 0.737至0.847），线性SVM为0.792（95%CI 0.736至0.848），LR为0.787（95%CI 0.730至0.844）。将放射组学与ML模型相结合可以成为一种有效、无创、经济且可靠的技术，用于在NCCT上评估隐匿性微观AIS并协助放射科医生做出临床决策。
