DA-GAN肺结节分割网络研究

Research of DA-GAN pulmonary nodule segmentation

为提高肺结节分割的精度,解决分割过程中的隐识别问题,该文提出双注意力生成对抗网络(doubleattention generative adversarial network,DA-GAN),通过在训练过程中捕捉信息的上下文依赖性和局部一致性,实现肺结节分割的高完整度以及高精度。该方法将自注意力机制引入生成器,使高分辨率特征图中的关键信息保留到低分辨率特征图中,实现上下文特征图之间长距离的依赖性,产生病灶区域的特征表示,使得分割图准确地保留肺结节区域,提高分割精度。同时,还强调肺结节区域的局部特征,增强分割结果图中结节位置的准确度。在语义分割的指导下,使用2个独立的判别器来区分肺结节的不同区域,即中心区域和边缘区域。通过在生成器和判别器中引入这两种互补的注意力机制,使网络可以学习到更丰富准确的特征表示,生成更精细的研磨图像,即准确的结节中心和边缘纹理。实验结果表明,对边缘多变且模糊的磨玻璃型结节的分割效果也较以往的算法得到很大程度的提升。通过定量和定性的实验结果证明该方法的有效性,获得约91.27%的像素准确率。

In order to improve the accuracy of pulmonary nodule segmentation and solve the problem of hidden recognition in the process of segmentation, this paper proposes double-attention generative adversarial network(DA-GAN). By capturing the context dependence and local consistency in the training process, the high integrity and high accuracy of pulmonary nodule segmentation can be achieved. In this method, the selfattention mechanism is introduced into the generator, so that the key information in the high-resolution feature map can be retained in the low resolution feature map, and the long-distance dependence between the context feature maps can be realized to produce more accurate feature representation. The segmentation map can accurately retain the pulmonary nodule region. At the same time, the local characteristics of the pulmonary nodule area were emphasized to enhance the accuracy of the nodule location in the segmentation results. Under the guidance of semantic segmentation, two independent discriminators were used to distinguish the different regions of pulmonary nodules, namely the central region and the marginal region. By introducing these two complementary mechanisms in generator and discriminator, the network can learn more abundant and accurate feature representation, and generate more precise grinding image, that is, more accurate nodule center and edge texture. The experimental results show that the segmentation effect of ground glass nodules with variable and fuzzy edges has been greatly improved compared with previous algorithms. The effectiveness of the method is proved by quantitative and qualitative experimental results, and the pixel accuracy of the method is 91.27%.