联合多MR序列迁移学习网络用于自动分级胶质瘤

目的 提出一种联合多MR序列迁移学习网络算法用于自动分级低级别(LGG)与高级别(HGG)胶质瘤，并评估其效能。方法 于开源数据库中提取76例LGG和259例HGG患者头部MR轴位T1WI、T2WI及液体衰减反转恢复(FLAIR)序列图像，均包含与轴位图像所见肿瘤最大层面相邻的20个层面图像，共6 700幅图像；采用相同的随机数列按7∶1.5∶1.5比例将各序列图像分为训练集(n=4 690)、验证集(n=1 005)及测试集(n=1 005)。以GoogLeNet预训练网络为胶质瘤分级模型的参数迁移源，重新设计输出模块，分别训练T1WI、T2WI及FLAIR单一模型，根据训练过程中的准确率和损失值曲线判断其收敛性；引入联合多序列模型投票机制，以降低单一序列模型对误分类的影响，利用测试集数据评价单一序列模型及联合多序列模型的效能。结果 各单一序列模型对训练集和验证集胶质瘤分级的准确率曲线均呈稳步上升趋势，损失值曲线均呈稳步下降趋势，之后均逐渐收敛。单一T1WI、T2WI、FLAIR模型及联合多序列模型对测试集胶质瘤分级的曲线下面积(AUC)分别为0.951 3、0.934 2、0.96...

Combining multiple MR sequences transfer learning networks for automatic grading of glioma

Objective To propose an automatic grading algorithm with combining multiple MR sequences transfer learning networks for low grade glioma (LGG) and high grade glioma (HGG), and to evaluate its efficacy. Methods Axial MR T1WI, T2WI and fluid attenuated inversion recovery (FLAIR) images of 76 LGG and 259 HGG patients were extracted from public database, each consisted 20 adjacent images of the largest tumor layer on the axial image, including 6 700 images which were divided into training set (n=4 690), validation set (n=1 005) or test set (n=1 005) at the ratio of 7:1.5:1.5 with the same random sequence. GoogLeNet pretrained network was used as transferred source of the parameters of glioma grading model, and the output module was redesigned. The single sequence models, i.e. T1WI, T2WI and FLAIR model, were trained respectively, and the convergence of the models were judged according to the accuracy curve and loss value curve during training process. The combined multi-sequence model voting mechanism was introduced to reduce the impact of single sequence model on misclassification, then the performances of single sequence model and combined multi-sequence model were evaluated based on the test set. Results The accuracy curves of glioma grading of each single sequence model in both training set and validation set showed steady upward trend, and the loss value curves showed steady downward trend, and then gradually converged. The area under the curve (AUC) of T1WI, T2WI, FLAIR model and combined multi-sequence model for grading of glioma in test set was 0.951 3, 0.934 2, 0.961 4 and 0.995 0, and the accuracy was 97.01%, 97.01%, 98.11% and 99.00%, respectively. Conclusion Combining multiple MR sequences transfer learning networks for automatic grading of glioma was concise and highly efficient.