基于改进U-Net网络的肺部CT图像结节分割方法

Segmentation of lung nodules in CT images based on improved U-Net

肺结节的精准分割对于肺肿瘤的良恶性诊断具有重要意义。针对肺结节在肺部CT图像中占比较小且形态各异,为肺结节的识别带来障碍等问题,提出一种改进的U-Net肺结节分割算法:加入双注意力模块,强化重要特征;使用残差模块简化网络训练,避免梯度消失;加入空洞空间金字塔池化模块,得到上下文不同尺度的特征;融合Focal loss函数与Dice损失函数,解决样本分布不均衡和难分样本挖掘问题。对Luna16数据集的100例患者的CT图像进行了测试。实验结果表明,该方法的交并比和F1分数分别达到了0.7888和0.8959。与Seg‐Net、U-Net、U-Net++网络和其他改进策略相比,本文方法可以准确地分割出肺结节,具有更好的分割性能。

The precise segmentation of lung nodules is important for the diagnosis of benign and malignant lung cancer.An improved U-Net lung nodule segmentation algorithm is proposed to solve the problems of small proportion in lung CT images and their various shapes which has brought obstacle to the recognition of lung nodules.Firstly,a dual attention module is added to strengthen important features.Secondly,the residual module is used to simplify network training and avoid the disappearance of the gradient.Thirdly,the ASPP module is added to the network to obtain the features of different scales in the context.Finally,fusion of Focal loss function and Dice loss function are used to solve the problem of unbalanced sample distribution and difficult-to-separate sample mining.Then,this paper uses CT images of 100 patients in the Luna16 data set to test the proposed algorithm.The results of the test show that the IOU and F1 of this method have reached 0.7888 and 0.8959 respectively.Compared with the SegNet,U-Net,U-Net++network and other improvement strategies,the network in this paper can accurately segment lung nodules and has better segmentation performance.