基于机器学习的多参数MRI影像组学预测喉鳞状细胞癌患者Ki-67表达水平的研究

Prediction of multi-parameter MRI radiomics based on machine learning in Ki-67 expression in laryngeal squamous cell carcinoma

目的探讨基于机器学习的多参数MRI影像组学在术前预测喉鳞状细胞癌(LSCC)患者Ki-67表达水平的价值。方法回顾性分析2016年1月至2024年2月在温州医科大学附属第五医院(中心1)和2019年1月至2023年12月丽水市人民医院(中心2)经病理检查证实为LSCC共271例患者的临床病理及MRI资料。将中心1的189例患者以7∶3比例随机分为训练集132例和验证集57例;中心2的82例患者作为外部测试集。在T2WI和对比增强T1加权成像(CE-T1WI)图像中提取病灶的影像组学特征,通过降维获得最佳特征并建立6种机器学习分类器。选择验证集和外部测试集中平均AUC最高的分类器作为最佳影像组学模型,并将其结果转换为影像组学评分(Rad-score)。将单因素分析中P<0.05的临床特征纳入多因素logistic回归分析,得到与Ki-67高表达相关的危险因素并建立临床模型。基于临床危险因素和影像组学特征构建联合模型,并绘制列线图。采用ROC曲线评价不同模型对Ki-67表达水平的预测效能。结果从T2WI和CE-T1WI图像中获得了12个最佳影像组学特征。在验证集和外部测试集中,6种机器学习分类器AUC范围分别在0.647~0.829和0.664~0.803,其中支持向量机具有最佳的预测效能(平均AUC为0.816)。进一步将临床T分期、MRI报告淋巴结状态与Rad-score相结合建立列线图模型。ROC曲线结果显示,列线图模型在训练集、验证集和外部测试集的AUC分别为0.923、0.870和0.822。结论基于机器学习的多参数MRI影像组学对LSCC患者Ki-67表达水平具有较好的预测效能,进一步结合临床特征建立的列线图模型能够更好地提升预测效能。

Objective To explore the value of machine learning-based multi-parameter MRI radiomics in predicting Ki-67 expression in laryngeal squamous cell carcinoma(LSCC)before surgery.Methods The clinicopathological and MRI data of 271 patients with LSCC confirmed by pathological examination in the Fifth Hospital Affiliated to Wenzhou Medical University(Center 1)from January 2016 to February 2024 and Lishui People's Hospital(Center 2)from Januany 2019 to December 2023 were retrospectively analyzed.A total of 189 patients from Center 1 were randomly divided into a training set of 132 cases and a validation set of 57 cases in a ratio of 7∶3;another 82 patients from Center 2 were used as an external test set.The radiomic features of lesions were extracted from T2WI and contrast-enhanced T1 weighted imaging(CE-T1WI)images.The optimal features were obtained through dimensionality reduction,and six machine learning classifiers were established.The classifier with the highest average AUC value in the validation set and the external test set was selected as the best radiomics model,and its results were converted to a radiomics score(Rad-score).Clinical features with P<0.05 in the univariate analysis were included in the multivariate logistic regression analysis to obtain the risk factors associated with high Ki-67 expression for establishing a clinical model.Finally,a combined model was constructed based on clinical features and radiomic features,and a nomogram was drawn.The ROC curve were used to evaluate the performance of different models in predicting Ki-67 expression.Results Twelve optimal radiomics features were obtained from T2WI and CE-T1WI images.In the validation set and external test set,the AUC ranges of the six machine learning classifiers were 0.647-0.829,and 0.664-0.803,respectively.Among them,random forest had the best predictive performance(average AUC of 0.816).The clinical T stage,MRI-reported lymph node status,and Rad-score were further combined to establish a nomogram.The ROC curve results showed that the AUC of the nomogram in the training set,validation set,and external test set were 0.923,0.870,and 0.822,respectively.Conclusion Multi-parameter MRI radiomics based on machine learning has good predictive value for the expression level of Ki-67 in LSCC patients,and the nomogram established by further combining clinical features can better improve performance.