Pan Shaoyan, Flores Jessica, Lin Cheng Ting, Stayman J Webster, Gang Grace J
Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore MD, 21205, USA.
Department of Biomedical Engineering, Johns Hopkins University, Baltimore MD, 21205, USA.
Proc SPIE Int Soc Opt Eng. 2021 Feb;11595. doi: 10.1117/12.2582151. Epub 2021 Feb 15.
Realistic lesion generation is a useful tool for system evaluation and optimization. In this work, we investigate a data-driven approach for categorical lung lesion generation using public lung CT databases. We propose a generative adversarial network with a Wasserstein discrimination and gradient penalty to stabilize training. We further included conditional inputs such that the network can generate user-specified lesion categories. Novel to our network, we directly incorporated radiomic features in an intermediate supervision step to encourage similar textures between generated and real lesions. We evaluated the network using lung lesions from the Lung Image Database Consortium (LIDC) database. The lesions are divided into two categories: solid vs. non-solid. We performed quantitative evaluation of network performance base on four criteria: 1) overfitting in terms of structural and morphological similarity to the training data, 2) diversity of generated lesions in terms of similarity to other generated data, 3) similarity to real lesions in terms of distribution of example radiomics features, and 4) conditional consistency in terms of classification accuracy using a classifier trained on the training lesions. We imposed a quantitative threshold for similarity based on visual inspection. The percentage of non-solid and solid lesions that satisfy low overfitting and high diversity is 96.9% and 88.6% of non-solid and solid lesions respectively. The distribution of example radiomics features are similar in the generated and real lesions indicated by a low Kullback-Leibler divergence score. Classification accuracy for the generated lesions are comparable with that for the real lesions. The proposed network is a promising approach for data-driven generation of realistic lung lesions.
逼真的病变生成是系统评估和优化的有用工具。在这项工作中，我们研究了一种使用公共肺部CT数据库进行分类肺部病变生成的数据驱动方法。我们提出了一种带有Wasserstein判别和梯度惩罚的生成对抗网络，以稳定训练。我们进一步纳入了条件输入，以便网络可以生成用户指定的病变类别。我们的网络新颖之处在于，在中间监督步骤中直接纳入了放射组学特征，以促进生成的病变与真实病变之间的纹理相似性。我们使用来自肺部影像数据库联盟（LIDC）数据库的肺部病变对网络进行了评估。病变分为两类：实性与非实性。我们基于四个标准对网络性能进行了定量评估：1）在与训练数据的结构和形态相似性方面的过拟合；2）在与其他生成数据的相似性方面生成病变的多样性；3）在示例放射组学特征分布方面与真实病变的相似性；4）使用在训练病变上训练的分类器在分类准确性方面的条件一致性。我们基于视觉检查设定了相似性的定量阈值。满足低过拟合和高多样性的非实性和实性病变的百分比分别为非实性病变的96.9%和实性病变的88.6%。通过低Kullback-Leibler散度得分表明，生成病变和真实病变中示例放射组学特征的分布相似。生成病变的分类准确性与真实病变相当。所提出的网络是一种有前途的数据驱动生成逼真肺部病变的方法。
