基于多序列MRI影像组学与深度迁移学习的脑胶质瘤分级预测研究
DOI:
作者:
作者单位:

1.南京医科大学附属泰州人民医院信息处;2.南京医科大学附属泰州人民医院脑外科;3.南京医科大学附属泰州人民医院影像科;4.南京医科大学泰州临床医学院

作者简介:

通讯作者:

中图分类号:

基金项目:


Grading prediction of glioma based on multi-sequence MRI deep learning radiomics features
Author:
Affiliation:

1.Department of Information,Taizhou People'2.'3.s Hospital Affiliated to Nanjing Medical University;4.Institute of Clinical Medicine,Taizhou People'

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    目的:探讨基于多序列MRI手工影像组学(hand-crafted radiomic,HCR)和深度迁移学习(deep transfer learning,DTL)特征的机器学习模型在术前预测脑胶质瘤分级的效能。方法:选取BraTS2019数据集中332例患者的影像数据[高级别胶质瘤(high-grade glioma,HGG):258例,低级别胶质瘤(low-grade glioma,LGG):74例],随机抽取30例HGG和8例LGG作为测试数据,其余294例作为训练集和验证集。从T1、T2 、T1c和Flair序列中提取病灶的HCR特征和DTL特征,并筛选出影响力前十的特征子集,基于HCR特征、DTL特征和两者组合的深度学习影像组学(deep learning radiomics,DLR)特征,分别建立7种的机器学习(machine learning,ML)模型,评估模型的预测效能。选择最佳模型后,使用SHAP法来对模型特征重要性进行量化及归因分析。结果:基于HCR和DTL组合的DLR特征构建的机器学习模型预测效能最高,当使用支持向量机的递归特征消除(SVM-recursive feature elimination,SVM-RFE)筛选特征后,使用T2+T1c+Flair序列组合的支持向量机(support vector machine,SVM)分类器的预测模型效果最佳。在验证集上,AUC达到0.996,约登指数(youden index,YI)、准确度(accuracy,ACC)、灵敏度(sensitivity,SEN)和特异性(specificity,SPE)分别为 0.920,0.976,0.988和0.932,在测试集上同样具有较高的分级预测效能。SHAP特征权重分析显示Flair序列的特征贡献较大,其次为T2及T1c序列,HCR和DTL特征均有重要贡献。结论:基于多序列MRI的DLR特征构建的机器学习模型可有效预测脑胶质瘤的肿瘤分级,其中经过SVM-RFE筛选后的T2+T1c+Flair序列组合的SVM分类器模型效能最佳。

    Abstract:

    Objective: This study aimed to develop a glioma grading prediction model utilizing multi-sequence MRI hand?crafted radiomics (HCR) and deep transfer learning (DTL) features, and assess its efficacy in preoperative glioma grading prediction. Methods: We utilized image data from 332 patients (HGG: 258, LGG: 74) in the BraTS2019 dataset. A subset of 30 HGG and 8 LGG were randomly selected as test data, while data from the remaining 294 patients were served as training and validation sets. Initial steps involved extracting HCR and DTL features from T1, T2, T1c, and Flair sequences. Seven machine learning (ML) prediction models were then constructed using HCR features, DTL features, and combined DLR features. The predictive accuracy of each model for HGGs and LGGs was assessed to determine the optimal ML model. The SHAP method was subsequently employed to quantify the importance of model features and for attribution analysis. Results: The ML model based on DLR features exhibited the most superior predictive performance.Specifically, when features were filtered using SVM-RFE, the SVM classifier model integrating T2, T1c, and Flair sequences emerged as the best, achieving an AUC of 0.996 (95% CI: 0.991, 1.000) on the validation set, with a Youden index, accuracy, sensitivity, and specificity of 0.920, 0.976, 0.988, and 0.932, respectively. On the test set, the model displayed an AUC of 0.986 and accuracy of 0.921, indicating robust predictive capacity.SHAP analysis highlighted that the Flair sequence had the greatest feature contribution, followed by the T2 and T1c sequences. Conclusion: The DLR model, based on multi-sequence MRI, proves effective in glioma tumor grade prediction. The optimal model is the SVM classifierusing T2, T1c, and Flair sequences post SVM-RFE selection.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2023-09-22
  • 最后修改日期:2023-12-16
  • 录用日期:2024-02-27
  • 在线发布日期:
  • 出版日期: