基于3D卷积网络和多模态MRI的脑胶质瘤自动分割

Auto-segmentation of glioma based on 3D convolutional network and multimodal MRI

目的:基于三维(3D)卷积神经网络和多模态MRI实现脑胶质瘤的自动分割。方法:首先对来自BraTS2020公共数据集的369例脑胶质瘤的4个模态MRI数据进行3D剪裁、重采样、去伪影、归一化的预处理。其次将MRI数据和脑胶质瘤标注信息输入到基于U-net的3D卷积神经网络模型进行训练和测试。利用相似性系数、召回率和精确率评价整体肿瘤区域、核心肿瘤区和增强肿瘤区的分割结果。结果:在74例测试数据集上,整体肿瘤区域、核心肿瘤区域和增强肿瘤区域的相似系数平均值分别为0.88、0.77和0.73,中位值分别为0.90、0.84和0.81,召回率平均值分别为0.88、0.78和0.78,中位值分别为0.90、0.84和0.84,精确率平均值分别为0.89、0.83和0.75,中位值分别为0.91、0.89和0.79。结论:基于U-net的3D卷积神经网络在多模态MRI数据集上获得了较好的分割结果,显示其在脑胶质瘤自动分割方面的潜力,可为临床诊断分级和治疗策略选择提供参考。

Objective To realize the automatic segmentation of glioma based on 3D convolutional neural network and multimodal magnetic resonance imaging(MRI).Methods The 4-modal MRI data of 369 cases of gliomas from the BRATS2020 public dataset were preprocessed by 3D clipping,resampling,artifacts removal and normalization.The preprocessed MRI data and label information of glioma were input into 3D convolutional neural network based on U-net for training and testing.The segmentation results of the whole tumor region,the core tumor region and the enhanced tumor region were evaluated using Dice similarity coefficient(DSC),recall rate and precision rate.Results For the 74 cases in testing dataset,the mean DSC of the whole tumor region,the core tumor region and the enhanced tumor region were 0.88,0.77 and 0.73,respectively,and the median values were 0.90,0.84 and 0.81,respectively.The mean recall rates were 0.88,0.78 and 0.78,respectively,and the median values were 0.90,0.84 and 0.84,respectively.The mean precision rates were 0.89,0.83 and 0.75,respectively,and the median values were 0.91,0.89 and 0.79,respectively.Conclusion The 3D convolutional neural network based on U-net achieves good segmentation results on multimodal MRI dataset,showing its potential for automatic segmentation of glioma,and it would be beneficial for clinical use in diagnosis and treatment planning.