基于影像组学和机器学习的脑膜瘤病理等级分类研究

【摘要】目的评估基于常规MRI的影像组学模型在预测脑膜瘤病理分级中的作用。方法回顾性分析227例接受术前常规磁共振扫描的脑膜瘤患者的临床资料(157例WHO Ⅰ级和70例WHO Ⅱ级)，按7∶3的比例随机分为训练组（158例）和验证组（69例）。将所有患者的T1WI增强和T2WI图像导入ITK-SNAP软件手动描绘肿瘤的感兴趣区域并提取影像组学特征。对特征数据进行降维处理后，再分别采用逻辑回归、高斯朴素贝叶斯、随机森林、K-邻近算法和支持向量机算法建立分类模型。采用受试者工作特征曲线来评价模型的预测性能。结果脑膜瘤病理等级的最佳分类模型在验证集中的曲线下面积为0.959(95% CI，0.878~1.000)。结论基于常规MRI的影像组学特征的机器学习分类器可以准确进行脑膜瘤病理等级的术前分类。

Abstract： Objective To evaluate the role of radiomics models based on conventional MRI in predicting pathological grade of meningiomas. Methods The clinical data of 227 patients with meningioma who underwent routine preoperative magnetic resonance scanning(157 who Ⅰ and 70 who Ⅱ) were analyzed retrospectively. They were randomly divided into training group(158 cases) and verification group(69 cases) according to the ratio of 7∶3. T1WI enhanced and T2WI images of all patients were imported into ITK-SNAP software to manually sketch tumor regions of interest and extract radiomics features. After dimensionality reduction of feature data, the classification model was established by logistic regression, Gaussian NB, random forest, K-nearest neighbor and support vector machine algorithms respectively. Receiver operating characteristic curve was used to evaluate the prediction performance of the model. Result The area under the curve of the best classification model of pathological grade of meningiomas in the validation set was 0.959(95% CI,0.878~1.000). Conclusion Machine learning classifier based on conventional MRI can accurately classify meningiomas by surgery.