基于术前MRI图像构建影像组学与深度学习的机器学习模型预测胶质瘤IDH-1基因表达的研究

目的探讨基于术前核磁共振成像（MRI）T2抑脂序列预测胶质瘤异柠檬酸脱氢酶（IDH)-1基因表达情况的价值。方法本研究共纳入2016年1月—2023年2月在济宁医学院附属医院经组织病理学证实的124例胶质瘤患者。使用ITK-SNAP软件勾画感兴趣区域(ROI)，使用Pyradiomics包实现影像组学特征的提取，使用经过预训练的ResNet50深度学习模型提取深度学习特征。使用Pearson相关系数和最小绝对收缩和选择算子(LASSO)回归模型进行特征筛选。最后进行10折交叉验证来评估模型效能。将传统影像组学、深度迁移学习以及融合模型基于支持向量机（SVM）、K近邻（KNN）以及随机森林（RF）三种机器学习模型分别建模。使用受试者工作特征（ROC）曲线评价各模型的预测效能。结果基于影像组学特征的机器学习模型SVM、KNN以及RF的曲线下面积（AUC）分别为0.699、0.628、0.616。基于深度迁移学习特征的机器学习模型SVM、KNN以及RF的AUC分别为0.853、0.753、0.807。基于融合特征的机器学习模型SVM、KNN以及RF的AUC分别为0.868、0.818、0.787。结论基于常规MRI序列中的T2WI抑脂序列的SVM融合模型对预测胶质瘤IDH-1基因表达情况具有较高的预测效能。

Objective To investigate value of preoperative magnetic resonance imaging(MRI) T2 fat suppression sequence in predicting isocitrate dehydrogenase(IDH)-1 gene expression in gliomas. Methods 124 patients with glioma, who were confirmed by histopathology, were collected. Regions of interest(ROI) was delineated using ITK-SNAP software. Pyradiomics package was used to extract radiomic features from the imaging data, while a pre-trained ResNet50 deep learning model was employed to extract deep learning features. Feature selection was performed using the Pearson correlation coefficient and the Least Absolute Shrinkage and Selection Operator(LASSO) regression model. Model performance was evaluated through 10-fold cross-validation. Traditional radiomics, deep transfer learning, and fusion models were separately constructed based on support vector machine(SVM), k-nearest neighbors(KNN), and random forest(RF) machine learning algorithms. The predictive performance of each model was assessed using receiver operating characteristic(ROC) curve. Results The area under curve(AUC) for the machine learning models SVM, KNN, and RF based on radiomic features were 0.699, 0.628, and 0.616, respectively. For the machine learning models SVM, KNN, and RF based on deep transfer learning features, the AUC values were 0.853, 0.753, and 0.807, respectively. The machine learning models SVM, KNN, and RF based on fusion features achieved AUC values of 0.868, 0.818, and 0.787, respectively. Conclusion The SVM fusion model based on the T2WI fat suppression sequence in routine MRI exhibits higher predictive performance in determining the IDH-1 gene expression status in gliomas.