摘要
目的:探讨基于常规MRI的影像组学模型对预测软组织肉瘤(STS)复发的价值。方法:回顾性分析2012年1月-2021年6月在本院经手术病理证实的92例STS患者的临床和影像资料。术后每3个月进行一次影像学检查,随访时间至少12个月以上,根据随访结果有无复发或远处转移分为复发组(27例),无复发组(65例)。采用完全随机方法将所有患者按7:3的比例分为训练集(n=65)和验证集(n=27)。使用ITK-SNAP软件,分别在T_(1)WI和压脂T_(2)WI上逐层沿肿瘤边缘手动勾画ROI并进行三维融合(VOI),然后使用AK软件提取纹理特征,使用最小冗余最大相关(mRMR)和最小绝对值收敛和选择算子(LASSO)回归分析方法分别对T_(1)WI序列、压脂T_(2)WI序列和联合序列的纹理特征进行降维和筛选,并建立影像组学模型,根据各个组学特征的权重系数计算影像组学评分(Radscore),运用100次留组交叉验证(LGOCV)方法来评估模型的可靠性。将临床病理、常规MRI特征与预测效能最高的影像组学模型的Radscore相结合,采用多因素logistic回归(LR)、随机森林(RF)和支持向量机(SVM)三种机器学习算法分别建立机器学习模型。采用受试者工作特征(ROC)曲线评价各模型的预测效能,应用决策曲线分析(DCA)评估模型的临床应用价值。结果:临床模型在训练集和验证集中预测STS复发的ROC曲线下面积(AUC)分别为0.71(95%CI:0.58~0.85)和0.74(95%CI:0.52~0.97)。基于T_(1)WI、压脂T_(2)WI和联合序列的影像组学模型在训练集中预测STS复发的AUC分别为0.81(95%CI:0.70~0.93)、0.92(95%CI:0.86~0.99)和0.91(95%CI:0.84~0.99),在验证集中分别为0.84(95%CI:0.63~1.00)、0.92(95%CI:0.81~1.00)和0.86(95%CI:0.72~1.00)。采用机器学习算法构建的LR、RF和SVM模型在训练集中预测STS复发的AUC分别为0.93(95%CI:0.87~0.99)、0.91(95%CI:0.84~0.99)和0.77(95%CI:0.63~0.91),在验证集中分别为0.93(95%CI:0.83~1.00)、0.86(95%CI:0.71~1.00)和0.83(95%CI:0.66~1.00)。DCA分析结果表明,压脂T_(2)WI和联合序列的影像组学模型、以及LR和RF模型的临床受益均较好。结论:基于常规MRI序列中的压脂T_(2)WI和联合序列构建的影像组学模型对预测STS复发具有较高的预测效能和较好的临床受益,基于不同机器学习算法构建的预测模型的预测效能并无明显提高。
Objective:To investigate the value of conventional MRI imaging diagnostic model in predicting the recurrence of soft tissue sarcoma(STS).Methods:The clinical and imaging data of 92 patients with STS confirmed by surgery and pathology in Yijishan Hospital of Wannan Medical University from January 2012 to June 2021 were retrospectively analyzed.Imaging examination was performed every 3 months after operation,and each of the follow-up interval was at least 12 months.According to the follow-up results,the patients were divided into recurrence group(27 cases)and non-recurrence group(65 cases).Patients were divided in a ratio of 7:3 into the training set(n=65)and the validation set(n=27)using a completely randomized approach.Using ITK-SNAP software,ROIs were manually delineated layer by layer along the tumor edge on T_(1)WI and fat suppressed(FS)T_(2)WI,and then 3D fusion(VOI)was performed.Then the AK software was imported to extract the texture features.The minimum redundancy maximum correlation(mRMR)and least absolute shrin-kage and selection operator(LASSO)regression analysis methods were used to reduce the dimension of texture features of T_(1)WI sequence,fat suppressed T_(2)WI sequence and combined sequence,respectively,to selected out the best feature subset,and establish the radiomics model.Then the radiomics score(Radscore)was calculated according to the feature weights,and the reliability of the model was evaluated by 100 leave-group-out cross validation(LGOCV).The clinicopathological indexes and conventional MRI features with the Radscore based on T_(2)WI sequence were combined to establish machine learning model using three machine learning algorithms,including multivariate logistic regression(LR),random forest(RF)and support vector machine(SVM).The predictive efficacy of diffe-rent model was evaluated by receiver operating characteristic(ROC)curve,and the clinical value of the model was evaluated by decision curve analysis(DCA).Results:The area under the ROC curve(AUC)of the clinical model for predicting STS recurrence was 0.71(95%CI:0.58~0.85)in the trai-ning set and 0.74(95%CI:0.52~0.97)in the validation set.The AUC of the radiomics model of T_(1)WI sequence,FS-T_(2)WI sequence and combined sequence in predicting STS recurrence in the training set were 0.81(95%CI:0.70~0.93),0.92(95%CI:0.86~0.99)and 0.91(95%CI:0.84~0.99),respectively.In the validation set,they were 0.84(95%CI:0.63~1.00),0.92(95%CI:0.81~1.00)and 0.86(95%CI:0.72~1.00),respectively.The AUC of LR,RF and SVM model constructed by machine learning algorithm for predicting STS recurrence in training set were 0.93(95%CI:0.87~0.99),0.91(95%CI:0.84~0.99)and 0.77(95%CI:0.63~0.91),respectively.In the validation set,they were 0.93(95%CI:0.83~1.00),0.86(95%CI:0.71~1.00)and 0.83(95%CI:0.66~1.00),respectively.The results of DCA analysis showed that the radiomics model of FS-T_(2)WI sequence and combined sequence,LR and RF models had better clinical benefits.Conclusion:Radiomics models based on fat suppression T_(2)WI and combined sequences of conventional MRI sequences have high predictive efficiency and good clinical benefit for predicting STS recurrence,while prediction models based on different machine learning algorithms have no significant improvement in the predictive efficiency.
作者
周慧
陈基明
吴莉莉
邵颖
范海云
陈亮亮
ZHOU Hui;CHEN Ji-ming;WU Li-li(Imaging Center,Yijishan Hospital,Wannan Medical College,Anhui 241001,China)
出处
《放射学实践》
CSCD
北大核心
2022年第12期1561-1567,共7页
Radiologic Practice
关键词
软组织肉瘤
肿瘤复发
磁共振成像
影像组学
预测模型
Soft tissue sarcoma
Tumor recurrence
Magnetic resonance imaging
Radiomics
Prediction model