基于深层多尺度聚合3D U-Net的肾脏与肾肿瘤分割方法

Segmentation method of kidney and kidney tumors based on deep multi-scale aggregation 3D U-Net

针对电子计算机断层扫描(CT)图像中肾肿瘤形态复杂多变、肿瘤目标小、肿瘤边缘复杂等问题,提出了深层多尺度聚合3D U-Net网络分割模型。该模型在U-Net++基础上新增了3个下采样操作,利用密集嵌套的3D U-Net和解码器层的跳跃连接以及各层级3D U-Net子网络之间的跳跃连接,促进各个层级和各个尺度的特征信息融合,增强了对细节特征的提取能力,从而提升了对小尺度肾肿瘤和肿瘤边缘的分割精度。实验结果表明:该模型能够准确分割边缘复杂以及尺度较小的肾肿瘤,在KiTS19公开数据集上进行评估,本文模型对肾脏分割的Dise系数为0.968 2,对肿瘤分割的Dise系数为0.790 8,分割性能良好。

Aiming at the problems of complex and changeable kidney tumors morphology,small tumor targets,and complex tumor edges in CT images,a DMSA 3D U-Net network segmentation model of deep multi-scale aggregation(DMSA)is proposed.Based on U-Net++,the model introduces three new down-sampling operations.Using the densely nested 3D U-Net and decoder end jump connections,as well as the jump connections between 3D U-Net sub-networks at each level,the model promotes the fusion of feature information at each level and at each scale,enhances the ability to extract detailed features.The above steps improve the segmentation accuracy of small-scale kidney tumors and tumor edges.The experimental results show that the proposed model can accurately segment kidney tumors with small scale and complex margins.When evaluated on the KiTS19 public dataset,the proposed model achieved an accuracy of 0.9682 for kidney segmentation,0.7908 for tumor segmentation,indicating its segmentation performance is good.