摘要近年来，生成对抗网络(GAN)技术迅速发展，此法可通过学习CT与PET影像之间的映射关系实现跨模态生成，既能融合解剖与功能信息、提高图像质量，也在一定程度上减少患者辐射负担。该文系统梳理条件GAN、循环GAN等典型GAN架构的原理与应用，聚焦其在肿瘤初诊与分期、疗效评估与随访以及PET/CT辐射剂量降低等方面的研究进展，探讨小样本学习、模型可解释性和跨机构标准化等挑战，展望GAN跨模态生成技术的临床应用前景。

abstractsWith the rapid development of generative adversarial networks (GAN), learning the mapping between CT and PET images enables cross-modality generation. This not only integrates anatomical and functional information to improve image quality, but also helps reduce the radiation exposure to some extent. Based on a review of representative GAN architectures such as conditional GAN and CycleGAN, this paper discusses their research progress and limitations in various application scenarios, including initial tumor diagnosis and staging, treatment evaluation and follow-up, as well as methods for reducing PET/CT radiation dose. Additionally, challenges related to small-sample learning, model interpretability, and cross-institutional standardization are highlighted, and the clinical application prospects of GAN-based cross-modality generation technology are explored.