基于多模态深度学习的脑肿瘤分割实验研究

Experimental research on brain tumor segmentation based on multimodal deep learning

针对现有3DU-Net网络在脑肿瘤分割中存在的训练过程中损失函数值难以降低,对增强瘤、肿瘤核分割精度较差等问题,该文提出了某模型网络的优化方案。首先使用残差网络结构降低训练难度;进一步引入注意力机制对多模态MRI的融合权值进行自适应学习,充分利用不同模态特征信息;最后在网络解码器部分采用双路卷积结构,提高了网络的特征提取能力。实验结果表明,改进后的网络训练损失函数更容易收敛到较小值,且对3种肿瘤的平均分割Dice系数提高了0.018 9,平均Hausdorff距离缩短了1.197 1,在整体分割性能上优于改进前的网络。

In view of the existing problems of 3D U-Net network in brain tumor segmentation,such as difficult to reduce the value of loss function in the training process,poor segmentation accuracy of enhanced tumor and tumor core,an optimization scheme for the model network is proposed in this paper.First,the residual network structure is used to decrease the difficulty of training.Furthermore,the attention mechanism is adopted for fusion weights adaptive learning of multimodal MRI to make full use of different modal characteristic information.Finally,the two-path convolution structure is used in the network decoder part to improve the capability of feature extraction of the network.The experimental results show that the training loss function of the improved network is easier to converge to a smaller value,the average segmentation Dice coefficient of the three kinds of tumors is increased by 0.018 9,and the average Hausdorff distance is shortened by 1.197 1,which is better than the network before improvement in the overall segmentation performance.