基于3D U-Net网络的肿瘤分割方法设计

Design of Tumor Segmentation Method Based on 3D U-Net

脑肿瘤MRI图像形态各异,类别严重不平衡,采用传统机器学习的半自动分割或深度学习的全自动分割方法,分割精度都不高。针对此问题,文章将3D U-Net[1]模型改进成一个层数更深的网络模型,此结构可以提取更多图像特征,但同时会导致网络难以训练,收敛过慢。为应对这种情况设计了一个叠加式残差块,在保留更多图像特征的同时,避免了深层网络无法收敛的问题。另外以混合损失函数代替传统Dice损失函数,可以增加脑肿瘤像素区域对总损失的贡献,提高稀疏分类错误对模型的惩罚,缓解数据类别不平衡问题。实验结果表明,在全肿瘤区域、肿瘤核心区域和肿瘤增强区域,提出的深层网络和混合损失函数的方法在分割精度上分别达到了0.88、0.82、0.66,在算法准确率上有了一定提升。

Brain tumor MRI images have different morphologies and serious imbalances. Using traditional machine learning semi-automatic segmentation or deep learning automatic segmentation methods,the segmentation accuracy is not high. In response to this problem,the 3 D U-Net[1]network is improved into a network with deeper layers. This structure can extract more image features,but it will make the network difficult to train and converge too slowly. In order to deal with this situation,a superimposed residual block is designed to avoid the problem that the deep network cannot converge while retaining more image features. In addition,the traditional Dice loss function is replaced by a mixed loss function,which increases the contribution of the brain tumor pixel area to the total loss,improves the punishment of the model by sparse classification errors,and alleviates the imbalance of data categories. The experimental results show that the proposed deep network and hybrid loss function method have achieved segmentation accuracy of 0.88,0.82 and 0.66 respectively in the whole tumor area,tumor core area,and tumor enhancement area,and the algorithm accuracy has been improved to some extent.