Li Junyuan, Pan Shaoyan, Zhang Xiaoxuan, Lin Cheng Ting, Stayman J Webster, Gang Grace J
IEEE Trans Biomed Eng. 2025 Jan;72(1):286-296. doi: 10.1109/TBME.2024.3451409. Epub 2025 Jan 15.
Data-driven methods for lesion generation are quickly emerging due to the need for realistic imaging targets for image quality assessment and virtual clinical trials. We proposed a generative adversarial network (GAN) architecture for conditional generation of lung lesions based on user-specified classes of lesion size and solidity. The network consists of two discriminators, one for volumetric lesion data, and one for radiomics features derived from the lesion volume. A Wasserstein loss with gradient penalty was adopted for each discriminator. Training data were drawn from contoured and annotated lesions from a public lung CT database. Four quantitative evaluation methods were devised to assess the network performance: 1) overfitting (similarity between generated and real lesions), 2) diversity (similarity among generated lesions), 3) conditional consistency (capability of generating lesions according to user-specified classes), and 4) similarity in distributions of various lesion properties between the generated and real lesions. Ablation studies were also performed to investigate the importance of individual network component. The proposed network was found to generate lesions that resemble real lesions by visual inspection. Solid lesions are distinct from non-solid ones, and lesion sizes largely correspond to their specified classes. With a classifier trained on real lesions, the classification accuracies of generated and real lesions in both solid and non-solid classes are similar. Radiomics features of generated and real lesions were found to have similar distributions, indicated by the relatively low Kullback-Leibler (KL) divergence values. Furthermore, the correlations between pairwise radiomics features in generated lesions were comparable to those of real lesions. The proposed network presents a promising approach for generating realistic lesions with clinically relevant features crucial for the comprehensive assessment of medical imaging systems.
由于图像质量评估和虚拟临床试验需要逼真的成像目标，基于数据驱动的病变生成方法正在迅速兴起。我们提出了一种生成对抗网络（GAN）架构，用于根据用户指定的病变大小和实性类别有条件地生成肺部病变。该网络由两个判别器组成，一个用于体积病变数据，另一个用于从病变体积中提取的放射组学特征。每个判别器都采用了带有梯度惩罚的 Wasserstein 损失。训练数据取自一个公共肺部 CT 数据库中经过轮廓勾勒和注释的病变。设计了四种定量评估方法来评估网络性能：1）过拟合（生成病变与真实病变之间的相似性），2）多样性（生成病变之间的相似性），3）条件一致性（根据用户指定类别生成病变的能力），以及 4）生成病变与真实病变之间各种病变属性分布的相似性。还进行了消融研究以探究各个网络组件的重要性。通过目视检查发现，所提出的网络生成的病变与真实病变相似。实性病变与非实性病变不同，病变大小在很大程度上与其指定类别相对应。使用在真实病变上训练的分类器，生成病变和真实病变在实性和非实性类别中的分类准确率相似。生成病变和真实病变的放射组学特征具有相似的分布，这由相对较低的 Kullback-Leibler（KL）散度值表明。此外，生成病变中两两放射组学特征之间的相关性与真实病变相当。所提出的网络为生成具有对医学成像系统综合评估至关重要的临床相关特征的逼真病变提供了一种有前景的方法。
