【摘要】 提出一种融合上下文注意力的两段式生成对抗网络用于肺结节生成和分类。上下文注意力采用一种通道增强的多头上下文注意力机制，将通道注意力和多头上下文注意力结合，更好地处理特征图中的复杂语义关系，有效增强了模型的特征提取能力；两段式生成对抗网络框架用于实现肺结节在指定肺部区域的注入，该框架将生成任务分为两个阶段：第一阶段生成肺结节感兴趣区域图像，然后通过泊松融合模块与指定的肺实质进行融合，生成初始样本；第二阶段使用改进的CycleGAN模型对初始样本进行微调。同时，在判别器中引入跨层激励模块和辅助分类器实现对特征通道的再校正以及对肺结节的分类。在LIDC-IDRI数据集上进行实验验证，实验结果表明，所提方法在肺结节生成上的FID、IS和KID评分分别为115.153、2.619±0.095和0.062；在肺结节恶性度分类上准确率为70.23%，灵敏度、F1值和AUC分别为68.66%、68.92%和87.59%，表现出优于ADGAN等基于GAN的分类模型，以及VGG16等基准网络的性能。更多还原

【Abstract】 A two-stage-based generative adversarial network incorporating contextual transformer is proposed for synthesis and multiclass classification of pulmonary nodules. Contextual transformer adopts a channel-enhanced multi-head contextual transformer mechanism which combines channel attention and multi-head contextual transformer to better deal with the complex semantic relationship in the feature space, thereby effectively enhancing the feature extraction capability of the model. A two-stagebased generative adversarial network framework is used to achieve the injection of pulmonary nodules in the designated lung area,and divide the synthesis task into two stages. In the first stage, pulmonary nodule regions of interest images are generated and then fused with designated lung parenchyma through a Poisson blending module to generate preliminary samples; in the second stage,an improved CycleGAN model is used to fine-tune the preliminary samples. Meanwhile, the skip layer excitation module and auxiliary classifier are introduced into the discriminator for realizing the re-correction of the feature channel and the classification of pulmonary nodules. Experiments on LIDC-IDRI dataset reveal that the proposed method has a FID, IS and KID of 115.153,2.619±0.095 and 0.062 on pulmonary nodule synthesis, and achieves an accuracy, sensitivity, F1 value and AUC of 70.23%, 68.66%,68.92% and 87.59% on pulmonary nodule malignancy classification, respectively, outperforming GAN-based classification models such as ADGAN, as well as benchmark networks such as VGG16.更多还原