摘要
目的:利用3D深度残差网络和多模态MRI实现对脑胶质瘤的自动分级。方法:利用BraTS2020公共数据集的293例高级别胶质瘤(HGG)和76例低级别胶质瘤(LGG)的多模态MRI数据训练和测试3D深度残差卷积网络模型。多模态MRI图像经过3D剪裁、重采样和归一化的预处理,随机分组为训练(64%)、验证(16%)和测试(20%)样本,将预处理后的多模态MRI图像和分级标注输入到网络模型进行训练、验证和测试。利用准确率(ACC)和受试者工作特征(ROC)曲线下面积(AUC)评价分级结果。结果:在59例(48例HGG和11例LGG)验证数据集上,ACC和AUC分别为0.93和0.97,在75例(62例HGG和13例LGG)测试数据集上,ACC和AUC分别为0.89和0.93。结论:3D深度残差网络在多模态MRI数据集上获得了较好的脑胶质瘤自动分级结果,可以为确定治疗方案和预测预后方面提供重要参考。
Objective To achieve automated glioma grading using 3D deep residual network and multimodal MRI. Methods The multimodal MRI data of 293 high grade gliomas(HGG) and 76 low grade gliomas(LGG) from the BraTS2020 public dataset were used to train and test the 3D deep residual network. After being preprocessed by 3D clipping, resampling and normalization, the multimodal MRI images and grade labeling were input into the network model for training(64%),validation(16%) and testing(20%). The glioma grading performance of the 3D deep residual network was evaluated in terms of accuracy(ACC), area under the receiver operating characteristic(ROC) curve(AUC). Results The ACC and AUC were0.93 and 0.97 in the 59 cases from validation dataset(48 cases of HGG and 11 cases of LGG), and 0.89 and 0.93 in the 75cases from testing dataset(62 cases of HGG and 13 cases of LGG). Conclusion The 3D deep residual network has a good performance in automated glioma grading on multi-modal MRI dataset, and it can provide significant reference for treatment option determination and prognosis prediction.
作者
王瑞
刘志强
齐崇
孟蓝熙
李少武
WANG Rui;LIU Zhiqiang;QI Chong;MENG Lanxi;LI Shaowu(Beijing Neurosurgical Institute/Beijing Tiantan Hospital,Capital Medical University,Beijing 100070,China;National Cancer Center/National Clinical Research Center for Cancer/Department of Radiotherapy,Cancer Hospital,Chinese Academy of Medical Sciences and Peking Union Medical College,Beijing 100021,China)
出处
《中国医学物理学杂志》
CSCD
2022年第10期1236-1243,共8页
Chinese Journal of Medical Physics
基金
中国医学科学院医学与健康科技创新工程项目(2022-I2MC&T-B-075)
国家自然科学基金(11905295,81901115)
中国癌症基金会“北京希望马拉松”专项基金(LC2021B01)。
