基于改进UNet的CT图像分割方法

CT Image Segmentation Method Based on Improved UNet

摘要: 针对UNet以及UNet变体网络在CT图像分割任务中的输入是CT图像序列中的单张切片，或者单张切片的多视图图像，且没有考虑CT图像序列之间的特征关系，提出了一种新型的CSI-UNet肺结节分割网络。该网络以多张连续CT图像切片作为输入，充分考虑了分割目标切片与相邻切片之间的特征。CSI-UNet模型对UNet的编码器进行改进，使之能接受连续的多张CT图像切片，在跳跃连接结构中，引入新提出的独立通用的连续切片层融合模块CSF-Block来完成特征提取和概率融合，从而得到更精准的分割结果。实验中利用LUNA16肺结节分割公开数据集，对该网络和UNet模型进行训练和验证。最后，使用Dice系数和MIoU对分割结果进行评估。在测试集上，UNet模型中Dice系数和均交并比分别能达到79.59%、78.81%，CSI-UNet模型分别能达到85.56%、84.72%。实验结果表明，CSI-UNet可以准确分割出胸部CT图像上的结节区域。

Abstract: The input of UNet and UNet variant network in CT image segmentation task is a single slice in CT image sequence, or a multi-view image of a single slice, and the feature relationship between CT image sequences is not considered, a new CSI-UNet lung nodule segmentation method is proposed. The network takes multiple continuous CT slices as input and fully considers the characteristics between the segmented target slice and the adjacent slices. The CSI-UNet model improves the encoder of UNet, so that it can accept multiple continuous CT image slices. In the skip connection structure, the newly proposed independent and general continuous slices layer fusion module CSF-Block is introduced to complete feature extraction and probabilistic fusion, so as to obtain more accurate segmentation results. In the experiment, LUNA16 lung nodule segmentation public data set was used to train and verify the network and UNet model. Finally, Dice coefficient and MIoU are used to evaluate the segmentation results. On the test set, Dice coefficient and MIoU in UNet model can reach 79.59% and 78.81% respectively, and CSI-UNet model can reach 85.56% and 84.72% respectively. The experimental results show that CSI-UNet can accurately segment the nodule region on the chest CT image.