光学技术, 2019, 45 (6): 749, 网络出版: 2020-01-08
基于3D CNN的脑胶质瘤分类算法
Brain glioma classification algorithm based on 3D CNN
摘要
基于目前脑胶质瘤在低分级胶质瘤和高分级胶质瘤上分类不准确的问题, 提出了基于3D CNN特征提取的磁共振成像脑胶质瘤分类方法。在3D CNN模型中加入Batch Normalizing层和Dropout层, 降低了过度拟合并加速了网络收敛速度; 使用N4ITK和数据增强后的数据进行训练, 进一步降低了过度拟合, 提升了模型分类效果; 构建特征融合层, 实现自动分类。实验结果表明, 算法在BraTS 2018数据集中的分类效果具有明显优势, 分类准确率高达91.67%, 明显高于AlexNet、VGG和ResNet三大主流网络模型, 算法具有较好的鲁棒性和泛化性, 对脑胶质瘤的分类和诊断具有临床应用潜力。
Abstract
Based on the inaccurate classification of gliomas in low-grade gliomas and high-grade gliomas at present, this thesis proposed the classification of gliomas by Magnetic Resonance Imaging in the basis of 3D CNN. Firstly, the Batch Normalizing layer and the Dropout layer are added to the 3D CNN model, which reduces the over-fitting and accelerates the convergence speed of the network. Then, the model has trained with N4ITK and the data that were augmented, which further reduces the over-fitting and improves the effect of model classification. Finally, the feature fusion layer is constructed to realize automatic classification. The experiment results show that the algorithm has obvious advantages in the classification of BraTS 2018 dataset, and the classification accuracy is as high as 91.67 percent, which is significantly higher than the three mainstream network models-AlexNet, VGG and ResNet. The algorithm has better robustness and generalization as well as has the clinical potential of the classification and diagnosis of glioma.
赵尚义, 王远军. 基于3D CNN的脑胶质瘤分类算法[J]. 光学技术, 2019, 45(6): 749. ZHAO Shangyi, WANG Yuanjun. Brain glioma classification algorithm based on 3D CNN[J]. Optical Technique, 2019, 45(6): 749.