Graduate School of Health Sciences, Fujita Health University, 1-98 Dengakugakubo, Kutsukake, Toyoake, Aichi, 470-1192, Japan.
Fujita Health University Hospital, 1-98 Dengakugakubo, Kutsukake, Toyoake, Aichi, 470-1192, Japan.
Radiol Phys Technol. 2020 Jun;13(2):160-169. doi: 10.1007/s12194-020-00564-5. Epub 2020 May 1.
It is often difficult to distinguish between benign and malignant pulmonary nodules using only image diagnosis. A biopsy is performed when malignancy is suspected based on CT examination. However, biopsies are highly invasive, and patients with benign nodules may undergo unnecessary procedures. In this study, we performed automated classification of pulmonary nodules using a three-dimensional convolutional neural network (3DCNN). In addition, to increase the number of training data, we utilized generative adversarial networks (GANs), a deep learning technique used as a data augmentation method. In this approach, three-dimensional regions of different sizes centered on pulmonary nodules were extracted from CT images, and a large number of pseudo-pulmonary nodules were synthesized using 3DGAN. The 3DCNN has a multi-scale structure in which multiple nodules in each region are inputted and integrated into the final layer. During the training of multi-scale 3DCNN, pre-training was first performed using 3DGAN-synthesized nodules, and the pulmonary nodules were then comprehensively classified by fine-tuning the pre-trained model using real nodules. Using an evaluation process that involved 60 confirmed cases of pathological diagnosis based on biopsies, the sensitivity was determined to be 90.9% and specificity was 74.1%. The classification accuracy was improved compared to the case of training with only real nodules without pre-training. The 2DCNN results of our previous study were slightly better than the 3DCNN results. However, it was shown that even though 3DCNN is difficult to train with limited data such as in the case of medical images, classification accuracy can be improved by GAN.
仅凭影像诊断通常难以区分肺部良性和恶性结节。当 CT 检查怀疑为恶性时，会进行活检。然而，活检具有高度侵袭性，良性结节的患者可能会接受不必要的操作。在这项研究中，我们使用三维卷积神经网络（3DCNN）对肺结节进行自动分类。此外，为了增加训练数据的数量，我们利用了生成对抗网络（GAN），这是一种深度学习技术，可用作数据增强方法。在这种方法中，从 CT 图像中提取以肺部结节为中心的不同大小的三维区域，并使用 3DGAN 合成大量伪肺结节。3DCNN 具有多尺度结构，每个区域中的多个结节输入并集成到最后一层。在多尺度 3DCNN 的训练过程中，首先使用 3DGAN 合成的结节进行预训练，然后使用真实结节微调预训练模型对肺结节进行综合分类。通过涉及 60 个基于活检的病理诊断确诊病例的评估过程，确定灵敏度为 90.9%，特异性为 74.1%。与仅使用真实结节而不进行预训练的情况相比，分类准确性得到了提高。我们之前研究的 2DCNN 结果略优于 3DCNN 结果。然而，事实证明，即使在像医学图像这样的数据有限的情况下，3DCNN 也很难进行训练，但通过 GAN 可以提高分类准确性。
