基于影像组学与集成学习的脑胶质瘤分级预测

脑胶质瘤的术前分级对治疗决策和预后评估至关重要. 为了提高分级精度, 提出了一种基于影像组学和集成学习的无创胶质瘤术前分级方法. 首先, 从不同序列的感兴趣区域提取428个影像组学特征, 采用递归特征消除算法进行特征选择, 采用6种不同的机器学习算法对脑胶质瘤进行分级, 并对各自的性能进行评估; 然后, 根据评估结果, 选取逻辑回归、决策树和多层感知机3种分类器作为脑胶质瘤分级预测的机器学习算法; 最后, 将3种分类器的输出采用投票方式进行集成, 并评估硬投票机制与软投票机制的性能. 实验结果表明, 对于数据集BraTS2019, 基于硬投票机制的集成学习算法的性能较好, 受试者工作特性曲线下面积为0.933±0.031, 准确度为0.886±0.048, 敏感度为0.872±0.077, 特异度为0.905±0.105. 该方法不仅能增加胶质瘤分级模型的可解释性, 而且可以提高分级精度.

Glioma grading prediction based on radiomics and ensemble learning

Glioma grading before surgery is very critical for the treatment planning and prognosis. In order to improve the grading accuracy, a non-invasive method for predicting the glioma grades based on radiomics and ensemble learning is proposed. First, 428 radiomics features are obtained from the region of interest (ROI) with different sequences. Feature selection is executed using Recursive Feature Elimination (RFE) algorithm, and 6 different machine learning algorithms are used to predict the glioma grade. Then, according to the evaluation results, three best classifiers, that is, Logistic Regression (LR), Decision Tree (DT) and Multilayer Perceptron (MLP), are selected as the machine learning algorithm for Glioma grading. Finally, these three classifiers are used for ensemble classification with a voting mechanism. The performance of hard and soft voting mechanism is also evaluated. Experimental results show that on the dataset BraTS2019, the hard voting mechanism based ensemble learning algorithm achieved the best performance, with the AUC value of 0.933±0.031, the accuracy of 0.886±0.048, the sensitivity of 0.872±0.077, and the specificity of 0.905±0.105. The presented work not only increases the interpretability of glioma grading model, but also ameliorates the grading accuracy.