改进Unet++的肾脏肿瘤分割方法

SEGMENTATION OF KIDNEY TUMOR BASED ON IMPROVED UNET++

针对人工方式分割CT图像肾脏肿瘤区域耗时费力且存在主观因素影响等问题,提出一种基于卷积神经网络的肾脏肿瘤自动分割算法。算法以Unet++分割网络为基础框架,将预训练的ResNet-34网络中四个特征提取模块作为Unet++网络特征编码器,来提取图像特征信息;并将重新设计的空洞空间金字塔池化网络嵌入到Unet++每条解码路径中;不同的解码路径通过特征融合得到肾脏肿瘤分割结果。在KiTS19竞赛提供的数据集上进行验证,实验结果表明,该算法有效提高了CT图像肾脏肿瘤的分割精度。

Aimed at the problems that the time-consuming and subjective factors affecting the artificial segmentation of renal tumor region in CT images,an automatic segmentation method for renal tumor based on convolution neural network is proposed.Unet++segmentation network was used as the basic framework.Four feature extraction modules in the pre-trained ResNet-34 network were used as Unet++network feature encoders to extract image feature information.The redesigned atrous space pyramid pooling network was embedded in each decoding path of the Unet++.The renal tumor segmentation results were obtained by feature fusion of different decoding paths.Validation was performed on the data set provided by the KiTS19 competition.The experimental results show that the algorithm effectively improves the segmentation accuracy of kidney tumor CT images.