针对医学图像分割中网络深度过深和上下文信息欠缺导致的分割精度降低等问题,提出了一种基于改进U-Net的磁共振成像(MRI)脑肿瘤图像分割算法。该算法通过嵌套残差模块和密集跳跃连接组成一种深度监督网络模型。为了减小编码路径和解码路径特征图之间的语义差距,将U-Net中的跳跃连接改为多类型的密集跳跃连接;为了解决网络过深导致的退化问题,加入残差模块,以防止网络梯度消失。实验结果表明,本算法分割肿瘤整体、肿瘤核心、增强肿瘤的Dice系数分别为0.88、0.84、0.80,满足临床应用的需求。

In view of the problems of deep network depth and lack of context information in medical image segmentation, which leads to the reduction of segmentation accuracy, an improved U-Net-based magnetic resonance imaging (MRI) brain tumor image segmentation algorithm is proposed in this paper. The algorithm forms a deep supervised network model by nesting residual block and dense skip connections. Change the skip connection in U-Net to multiple types of dense skip connection to reduce the semantic gap between the encoding path and the decoding path feature map; add a residual block to solve the degradation problem caused by too deep network to prevent the network gradient from disappearing. Experimental results show that the Dice coefficients of the algorithm for segmenting the whole tumor, tumor core, and enhanced tumor are 0.88, 0.84, and 0.80, respectively, which meets the needs of clinical applications.