作者：赵俊强1,2,3,4, 李骥5, 王昌1,2,3,4, 刘华楠1,2,3,4, 李南艾1,2,3,4, 吴星阳1,2,3,4
作者单位：1. 新乡医学院，河南 新乡 453003;
2. 新乡市智能影像诊断工程技术研究中心，河南 新乡 453003;
3. 新乡市医学VR（AR）与智能反馈重点实验室，河南 新乡 453003;
4. 临床与生物医学大数据融合技术河南省工程实验室，河南 新乡 453003;
5. 郑州大学第一附属医院，河南 郑州 450052
关键词：肺结节分割;自注意力机制;生成对抗网络;上下文信息
摘要：
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(double-attention 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。
Research of DA-GAN pulmonary nodule segmentation
ZHAO Junqiang1,2,3,4, LI Ji5, WANG Chang1,2,3,4, LIU Huanan1,2,3,4, LI Nan’ai1,2,3,4, WU Xingyang1,2,3,4
1. Xinxiang Medical University, Xinxiang 453003, China;
2. Xinxiang Intelligent Image Diagnosis Engineering Technology Research Center, Xinxiang 453003, China;
3. Xinxiang Key Laboratory of Medical Virtual Reality (Augmented Reality) and Intelligent Feedback, Xinxiang 453003, China;
4. Henan Engineering Laboratory of Combinatorial Technique for Clinical and Biomedical Big Data, Xinxiang 453003, China;
5. The First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, China
Abstract: 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 self-attention 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%.
Keywords: pulmonary nodule segmentation;self-attention mechanism;generative adversarial network;context dependence
2022, 48(9):118-124 收稿日期: 2021-04-08;收到修改稿日期: 2021-06-01
基金项目: 河南省教育教学改革重点课题（2021SJGLX217）；河南省教育科学“十四五”规划重点课题（2021JKZD09）；河南省科技攻关项目（222102310615）；河南省省部共建重点课题（SBGJ202102189）；新乡医学院第一附属医院开放课题（XZZX2022011）
作者简介: 赵俊强（1979-），男，河南洛阳市人，副教授，博士，研究方向为医学人工智能与VR、医疗大数据
参考文献
[1] HAN F, WANG H, ZHANG G, et al. Texture feature analysis for computer-aided diagnosis on pulmonary nodules[J]. Journal of Digital Imaging, 2015, 28: 99-115
[2] HIRSCHAUER T J, ADELI H, BUFORD J A. Computer-aided diagnosis of Parkinson’s disease using enhanced probabilistic neural network[J]. Journal of medical systems, 2015, 39(11): 179
[3] KOMINAMI Y, YOSHIDA S, TANAKA S, et al. Computer-aided diagnosis of colorectal polyp histology by using a real-time image recognition system and narrow-band imaging magnifying colonoscopy[J]. Gastrointestinal Endoscopy, 2016, 83(3): 643-649
[4] MUKHERJEE S, HUANG X, BHAGALIA R R. Lung nodule segmentation using deep learned prior based graph cut[C]//The 2017 IEEE International Symposium on Biomedical Imaging. IEEE, 2017.
[5] WANG S, ZHOU M, GEVAERT O, et al. A multi-view deep convolutional neural networks for lung nodule segmentation[C]//2017 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2017.
[6] KESHANI M, AZIMIFAR Z, TAJERIPOUR F, et al. Lung nodule segmentation and recognition using SVM classifier and active contour modeling: A complete intelligent system[J]. Computers in Biology & Medicine, 2013, 43(4): 287-300
[7] NITHILA E E, KUMAR S S. Segmentation of lung nodule in CT data using active contour model and Fuzzy C-mean clustering[J]. Alexandria Engineering Journal, 2016, 55(3): 2583-2588
[8] SAVIC M, MA Y, RAMPONI G, et al. Lung nodule segmentation with a region-based fast marching method[J]. Sensors, 2021, 21(5): 1908
[9] CHA J, FARHANGI M M, DUNLAP N, et al. Segmentation and tracking of lung nodules via graph‐cuts incorporating shape prior and motion from 4D CT[J]. Medical Physics, 2018, 45(1): 297-306
[10] KEETHA N V, BABU P S A, ANNAVARAPU C S R. U-Det: A modified U-Net architecture with bidirectional feature network for lung nodule segmentation[J/OL]. Image and Video Processing:1-14[2022-04-01]. https://doi.org/10.48550/arXiv.2003.09293
[11] LIU W, LIU X, LI H, et al. Integrating lung parenchyma segmentation and nodule detection with deep multi-task learning[J]. IEEE Journal of Biomedical and Health Informatics, 2021, 25(99): 3073-3081
[12] 张宜劼. 连续时间Dynkin博弈的纳什均衡点存在性的一个新证明[J]. 宁夏大学学报(自然科学版), 2019, 40(3): 197-203
[13] 王昕, 赵飞, 蒋佐富, 等. 迁移学习和卷积神经网络电力设备图像识别方法[J]. 中国测试, 2020, 46(5): 108-113
[14] RONNEBERGER O , FISCHER P , BROX T . U-Net: convolutional networks for biomedical image segmentation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer International Publishing, 2015.
[15] 王磐, 强彦, 杨晓棠. 基于双注意力3D-UNet的肺结节分割网络模型[J]. 计算机工程, 2021, 47(2): 307-313
[16] LI P, WU X, HU Y, et al. M2FPA: A multi-yaw multi-pitch high-quality database and benchmark for facial pose analysis[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019.
[17] 张丽. 基于薄扫CT序列图像的肺结节分割及分类方法研究[D]. 太原: 太原理工大学, 2019.
[18] ARMATO S, MCLENNAN G, MCNITT-RAY M, et al. WE-B-201B-02: The lung image database consortium (LIDC) and image database resource initiative (IDRI): A completed public database of ct scans for lung nodule analysis[J]. Medical Physics, 2010, 37(6): 3416-3417
[19] HE K M, ZHANG X Y, REN S Q. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification[C]//CVPR. IEEE Computer Society, 2015.
[20] NOWOZIN S. Optimal decisions from probabilistic models: The intersection-over-union case[C]//IEEE Conference on Computer Vision & Pattern Recognition. IEEE Computer Society, 2014: 548-555.
