作者：伍发，王鹏，蒋锐，冯莉娟，李建浩，陈娅，杜飞舟

单位：

1.610083四川省成都市，西部战区总医院放射诊断科 2.100020北京市，中日友好医院核医学科 通信作者：杜飞舟，E-mail：huanghejoker@aliyun.com

Units：

1.Department of Radiology, the General Hospital of Western Theater Command, Chengdu 610083, China 2.Department of Nuclear Medicine, China-Japan Friendship Hospital, Beijing 100020, China Corresponding author: DU Feizhou, E-mail: huanghejoker@aliyun.com

关键词：

脑出血; 自发性脑出血; 血肿扩大; 影像组学; 预测;

Keywords：

Cerebral hemorrhage; Spontaneous intracranial hemorrhage; Hematoma expansion; Radiomics;Forecasting

CLC：

DOI：

10.12114/j.issn.1008-5971.2022.00.173

Funds：

四川省科技计划项目（2018JY0804）

摘要：

目的 分析影像组学特征对自发性脑出血(sICH)患者发生血肿扩大(HE)的预测价值。方法 回顾性选取2017—2021年于西部战区总医院就诊的sICH患者232例为研究对象。收集患者一般资料,基于颅脑CT检查结果提取影像组学特征。根据HE发生情况,将患者分为HE组(39例)和非HE组(193例)。采用Lasso-Logistic回归模型选择最终影像组学特征并形成影像组学标签;采用ROC曲线分析影像组学标签对sICH患者发生HE的预测价值;按7∶3的比例将sICH患者分为训练集(162例)与验证集(70例),分别采用自适应提升算法(Adaboost)、K最近邻(KNN)分类算法、随机森林算法再次验证影像组学标签对sICH患者发生HE的预测价值。结果 共提取出107个影像组学特征。Lasso-Logistic回归模型结果显示,AUC最大时共包含11个影像组学特征。将Lasso-Logistic回归模型筛选出的11个影像组学特征与相应的加权系数乘积的线性组合作为影像组学标签。HE组患者影像组学标签高于非HE组(P<0.05)。影像组学标签预测sICH患者发生HE的AUC为0.780[95%CI(0.703,0.857)],最佳截断值为-1.34,灵敏度为66.7%,特异度为76.7%,约登指数为0.434,正确率为75.0%。Adaboost、KNN分类算法、随机森林算法分析结果显示,影像组学标签预测训练集sICH患者发生HE的AUC分别为0.881、0.873、0.904,预测验证集sICH患者发生HE的AUC分别为0.614、0.857、0.888。结论 由第10百分位数、第90百分位数、集群阴影、长行程高灰度优势、灰度差异度、短行程优势、标准化灰度不均匀度、区域性、力度、半轴长度、最小长度组成的影像组学标签对sICH患者发生HE有较好的预测价值,临床医生可根据该影像组学标签来筛选HE高风险的sICH患者,进而采取更为积极的治疗措施,从而改善患者预后。

Abstract：

【Abstract】　Objective To analyse the predictive value of radiomics features for hematoma expansion (HE) in patientswith spontaneous intracerebral hemorrhage (sICH) . Methods A total of 232 patients with sICH who were treated in the GeneralHospital of Western Theater Command from 2017 to 2021 were retrospectively selected as the research subjects. The generaldata of the patients were collected, and the radiomics features were extracted based on the CT examination results of the brain.According to the occurrence of HE, the patients were divided into HE group (39 cases) and non-HE group (193 cases) . LassoLogistic regression model were used to select the final radiomics features and form radiomics signatures; ROC curve was used toanalyze the value of radiomics signatures in predicting HE in patients with sICH; the sICH patients were divided into a training set(162 cases) and a validation set (70 cases) according to the ratio of 7∶3, and the adaptive boosting algorithm (Adaboost) , K-NearestNeighbor (KNN) classification algorithm and random forest algorithm were used to re-validate the value of radiomics signaturesin predicting HE in sICH patients, respectively. Results A total of 107 radiomics features were extracted. The results of theLasso-Logistic regression model showed that there were 11 radiomics features when AUC was the largest. The linear combinationof the 11 radiomics features screened out by the Lasso-Logistic regression model and the products of the corresponding weightingcoefficients was used as the radiomics signatures. The radiomics signatures of patients in HE group were higher than those in nonHE group (P <0.05) . The AUC of the radiomics signatures for predicting HE in sICH patients was 0.780 [95%CI (0.703, 0.857) ] ,the best cutoff value was -1.34, the sensitivity was 66.7%, the specificity was 76.7%, the Youden index was 0.434, and the correctrate was 75.0%. The analysis results of Adaboost, KNN classification algorithm, and random forest algorithm showed that the AUCof radiomics signatures for predicting HE in sICH patients in the training set was 0.881, 0.873 and 0.904, respectively, and theAUC of radiomics signatures for predicting HE in sICH patients in the validation set was 0.614, 0.857 and 0.888, respectively.Conclusion The radiomics signatures consisting of 10th percentile, 90th percentile, cluster shade, long-stroke high grayscaledominance, grayscale dissimilarity, short-stroke dominance, normalized grayscale unevenness, regionality, strength, half-axislength and the minimum length have a good predictive value for HE in sICH patients. Clinicians can screen sICH patients withhigh HE risk based on the radiomics signatures, and then take more active treatment measures to improve the prognosis of patients.

ReferenceList：