通讯作者: 张志诚

DOI：10.12201/bmr.202511.00066

Survey on the Applications of Generative Models in Medical Image Analysis

• 摘要：目的/意义 医疗影像分析是精准医疗的核心，但高质量标注数据稀缺、跨设备域偏移等瓶颈长期制约其发展。深度生成模型以其强大的数据分布建模能力，为突破上述困境提供了关键技术路径。本文旨在全面系统地梳理生成模型在医疗影像分析领域的研究现状、前沿进展与核心挑战，为相关研究提供一份全面的技术图谱与前瞻性指引。方法/过程 本文采用文献综述的研究方法。首先，系统阐述了以生成对抗网络、变分自编码器和扩散模型为代表的主流生成模型的基本原理、技术演进及优缺点。其次，从影像跨模态合成、数据增强、重建去噪、超分辨率、分割检测等关键应用任务出发，对现有研究工作进行了详细归纳和分类。最后，对模型性能的评估框架进行梳理，整合了从技术指标到临床应用效能的多维度评测体系。结果/结论 研究表明，生成模型已在医疗影像分析的多个关键任务中展现出巨大潜力与应用价值。然而，其临床转化仍面临模型可控性与可解释性不足、泛化鲁棒性待提升、数据伦理与高计算开销等多重挑战。结论认为，未来研究的突破口在于构建可信、可控的医疗基础模型，深化多模态数据融合，并建立标准化的评测基准与协议。本综述的系统性梳理为推动生成模型在医疗领域的健康发展和临床落地提供了重要的理论参考和方向性指导。

  关键词： 医疗影像分析；生成对抗网络；变分自编码器；扩散模型；; ; ; 

   

  Abstract: Abstract Purpose/Significance Medical image analysis lies at the core of precision medicine. However, the scarcity of high-quality annotated data and domain shifts across imaging devices have long hindered its advancement. Deep generative models, with their powerful capacity for modeling complex data distributions, provide a key technical pathway to overcome these limitations. This paper aims to systematically review the current state, cutting-edge progress, and core challenges of generative models in medical image analysis, offering a comprehensive technological landscape and forward-looking guidance for future research.Method/Process This study adopts a literature review methodology. First, it systematically elaborates on the fundamental principles, technical evolution, and advantages and limitations of mainstream generative models, including Generative Adversarial Network (GAN), Variational Autoencoder (VAE), and Diffusion Models. Second, from the perspectives of key application tasks—such as cross-modality image synthesis, data augmentation, image reconstruction and denoising, super-resolution, segmentation, and detection—it provides a detailed summary and classification of existing research efforts. Finally, the paper reviews evaluation frameworks for model performance, integrating a multidimensional assessment system that spans from technical indicators to clinical utility. Result/Conclusion The findings indicate that generative models have demonstrated remarkable potential and application value across multiple critical tasks in medical image analysis. Nevertheless, clinical translation remains constrained by issues such as insufficient controllability and interpretability, limited generalization and robustness, data ethics concerns, and high computational costs. The study concludes that future breakthroughs will depend on developing trustworthy and controllable medical foundation models, deepening multimodal data integration, and establishing standardized evaluation benchmarks and protocols. This comprehensive review provides theoretical insight and directional guidance for advancing the healthy development and clinical implementation of generative models in the medical domain.

  Key words: Medical Image Analysis; GANs; VAEs; Diffusion Models；

  提交时间：2025-11-21

  版权声明：作者本人独立拥有该论文的版权，预印本系统仅拥有论文的永久保存权利。任何人未经允许不得重复使用。
• html

• 图表

• 史森中. 生成式人工智能（GAI）基层医疗应用风险及治理策略分析. 2024. doi: 10.12201/bmr.202408.00053

  李小鹏, 喻慧心, 常青, 赵玉虹. 医疗基础大模型的应用现状与挑战. 2025. doi: 10.12201/bmr.202511.00060

  马傲, 葛小玲. 人工智能大模型在医疗健康领域应用的研究. 2024. doi: 10.12201/bmr.202408.00039

  施呈昊, 屠馨怡, 史佳伟, 陈红爽, 王琴潞, 邹海欧. 大语言模型在临床实践中应用的范围综述. 2024. doi: 10.12201/bmr.202406.00001

  吴宏, 胡军, 陈尔真, 董晨杰, 李建华, 叶琪. 基于大模型的医疗质控应用系统构建. 2025. doi: 10.12201/bmr.202503.00004

  闾海荣. DeepSeek与医学大模型：技术创新与医学服务模式的重构. 2025. doi: 10.12201/bmr.202503.00023

  周建国. 大数据背景下结构式医学影像诊断报告的临床应用. 2024. doi: 10.12201/bmr.202409.00066

  郑琰莉, 韩福海, 李舒玉, 苏文星. 人工智能大模型在医疗领域的应用现状与前景展望. 2023. doi: 10.12201/bmr.202312.00027

  弓孟春, 李雨杭, 马永慧, 弓凯, 刘超, 欧阳自豪, 戴辉. 生成式医学人工智能(GMAI)的临床伦理治理:三维协同路径与中国实践. 2025. doi: 10.12201/bmr.202511.00057

  中国科学技术信息研究所, 北京市海淀区复兴路号. 基于生成式人工智能的医学视频创作路径探究. 2025. doi: 10.12201/bmr.202508.00005

• 序号	提交日期	编号	操作
  1	2025-09-12	10.12201/bmr.202511.00066V1	下载

• 公开评论  匿名评论  仅发给作者 提交评论