摘要目的 探讨高通量纹理分析鉴别脑内单发转移瘤( SBM)和高级别胶质瘤( HGG)的可行性,并验证建立的诊断模型.方法 回顾性收集病理确诊的SBM和HGG患者各43例,患者术前均行常规头部MRI扫描.从86例患者的MRI图像中选取236幅包含肿瘤信息的FLALR图像,将每幅图像视为1个样本.训练组200幅,其中SBM 106幅,HGG 94幅,用于建立诊断模型.验证组36幅,其中SBM 19幅,HGG 17幅,用于验证所建立的诊断模型.经过图像预处理、图像分割、特征萃取、特征选择,最终建立影像组学诊断模型.采用受试者工作特征( ROC)曲线评价诊断模型的诊断效能.采用分层聚类分析评价所提取图像特征数据的质量和诊断模型的分类效果.采用独立的验证组对模型进一步检验.结果 每个样本有629个图像特征被萃取和量化,筛选出其中的41个特征建立特征子集,构建了诊断模型模型,分类决策函数为f(x)＝sign [∑N i＝1 aiyiK(x,xi)+ b] ,核函数为K(x,xi)＝exp(－‖x－xi‖2/2σ2).在训练组中,诊断模型鉴别SBM和HGG的准确性为0.845,灵敏度为0.849,特异度为0.840,阳性预测值为0.857,阴性预测值为0.832,ROC曲线下面积达0.939.模型在验证组中具有相似的结果 .结论 高通量纹理分析对SBM和HGG的鉴别诊断具有较高准确性.

abstractsObjective To explore the feasibility of high-throughput texture analysis in the distinction of single brain metastases ( SBM) from high-grade gliomas ( HGG) and validate the established model. Methods A total of 86 patients who were histologically diagnosed with SBM or HGG were retrospectively collected, including 43 patients with SBM and 43 with HGG. All of patients were performed preoperative conventional head magnetic resonance imaging ( MRI) scans. A total of 236 fluid-attenuated inversion recovery (FLALR) images containing the information of tumors were selected from the MRI images and each image was considered as an object.The training set had 200 images, including 106 from SBM group and 94 from HGG group, whereas the validation set had 36 images, including 19 from SBM group and 17 from HGG. After images preprocessing, images segmentation, features extraction, and features selection, a radiomic diagnostic model was finally established using the training set. The diagnostic performance of the diagnostic model was evaluated using a receiver operating characteristic ( ROC ) curve. Hierarchical clustering analysis was used to evaluate the quality of the extracted feature data and the classification effect of the model. The model was further validated using the independent validation set. Results A total of 629 features were extracted and quantified from each sample, and 41 features were selected to establish feature subsets and the diagnostic model. The classification decision function of the model is f ( x )＝sign [∑Ni＝1 aiyiK(x,xi)+b] and the kernel function of the model is K(x,xi)＝ exp(－‖x－xi‖2/2σ2). In the training set, the diagnostic accuracy, sensitivity, specificity, positive predictive value and negative predictive value were 0.845, 0.849, 0.840, 0.857 and 0.832, respectively. The area under the ROC curve reached to 0.939. Similar results were obtained in the validation set. Conclusion The high-throughput texture analysis shows high accuracy in differentiating SBM from HGG.