用于多模态MRI脑肿瘤图像分割的融合双重对抗学习CNN-Transformer模型

Fusing Dual Adversarial Learning CNN-Transformer Model forMulti-modal MRI Brain Tumor Image Segmentation

针对CNN(convolutional neural network)-Transformer分割模型在训练数据中提取特征信息不充分和鲁棒性差的问题,提出融合双重对抗学习的CNN-Transformer(CNN-Transformer model fusing dual adversarial learning,TransFDA)分割模型。首先,引入判别器模块,此模块不同于常用的对抗学习方法,先将原始数据与预测出的结果进行判别,然后将原始数据中未分割区域的信息传入分割模型,加强分割模型对此区域信息的学习。其次,为提升分割模型的鲁棒性,引入虚拟对抗训练,使用模型正常预测结果和添加扰动后得到的预测结果进行对抗学习,提升分割模型对不确定数据信息的学习能力。在Brats2020验证集中,肿瘤整体区域(whole tumor,WT)、肿瘤核心区域(tumor core,TC)和增强肿瘤区域(enhancing tumor,ET)的戴斯相似系数(Dice similarity coefficient,DSC)分别为0.8922、0.7909、0.7530。相较于同等实验条件下的TransBTS(brain tumor segmentation using Transformer)模型性能有所提升。定量和定性实验结果表明,所提TransFDA在不需要额外添加数据的情况下学习到了更多的特征信息,增强了模型的鲁棒性,显著提升了模型分割精度。

Aiming at the problems of inadequate feature information extraction and poor robustness of the CNN-Transformer segmentation model in training data,this paper proposes a CNN-Transformer model fusing dual adversarial learning(TransFDA).First,this paper introduces the discriminator module,which is different from the common adversarial learning methods,by discriminating the original data and the predicted results,and then passing the information of the unsegmented region in the original data into the segmentation model to enhance the learning of this region by the segmentation model.Second,to improve the robustness of the segmentation model,this paper introduces virtual adversarial training,which uses the normal prediction results of the model and the prediction results obtained after adding disturbance for adversarial learning to improve the learning ability of the segmentation model for information on uncertain data.In the Brats2020 validation dataset,the Dice similarity coefficient(DSC)of the whole tumor(WT)area,tumor core(TC)area and enhancing tumor(ET)area are 0.8922,0.7909,and 0.7530,respectively,increased compared with the TransBTS model under the same conditions.Quantitative and qualitative experiments show that TransFDA learns more feature information without adding additional data,which enhances the robustness of the model and significantly improves the model segmentation accuracy.