SOM-GAN：一种用于非配对医学图像合成的结构保持一对多生成对抗网络

摘要(中文译文)

基于生成对抗网络（GAN）的图像翻译已广泛应用于跨域医学图像合成。然而，现有大多数方法遵循一对一的映射范式，需要为每个目标域单独训练模型，在多目标任务（如多序列MRI合成）中增加了训练成本。尽管近期的一对多方法提高了翻译效率，但由于缺乏输入图像与合成图像之间显式的几何约束，它们往往难以保持结构一致性。分割基础模型的最新进展使医学成像中的解剖标注更加易于获取。然而，现有的图像翻译方法尚未充分利用这些标注所提供的丰富解剖与结构先验信息。因此，生成的图像可能出现解剖扭曲和语义错位，可能限制其在对解剖结构敏感的下游任务中的可靠性，例如多序列MRI中的器官和肿瘤分割。为解决上述局限，我们提出SOM-GAN，一种结构保持的一对多生成对抗网络，用于从单一源域向多个目标域进行高效且结构保持的图像合成。SOM-GAN将翻译过程建模为针对每个目标域的一对三方博弈，并采用双路径对抗设计。具体而言，多域外观判别器促使生成逼真的目标域纹理渲染，而多域结构判别器通过对图像与语义标注的联合分布进行对抗对齐，引入显式的结构约束。通过交替优化这些对抗目标，SOM-GAN使多域条件生成器能够在多个目标域上合成高质量图像。在四项典型且具有挑战性的多目标医学图像翻译任务上的大量实验表明，SOM-GAN在合成图像保真度以及下游临床分割和诊断分类任务中均优于最先进的方法。代码已发布于https://github.com/JinhaoLi98/SOM-GAN。

Abstract

GAN-based image translation has been widely used for cross-domain medical image synthesis. However, most existing methods follow a one-to-one mapping paradigm, requiring a separate model for each target domain and increasing the training cost in multi-target tasks such as multi-sequence MRI synthesis. Although recent one-to-multiple methods improve translation efficiency, they often struggle to preserve structural consistency because they lack explicit geometric constraints between the input and the synthesized images. Recent advances in segmentation foundation models have made anatomical annotations more accessible in medical imaging. However, existing image translation methods have not fully exploited the rich anatomical and structural priors provided by these annotations. As a result, the generated images may exhibit anatomical distortions and semantic misalignments, potentially limiting their reliability for anatomy-sensitive downstream tasks, such as organ and tumor segmentation in multi-sequence MRI. To address these limitations, we propose SOM-GAN, a Structure-preserving One-to-Multiple Generative Adversarial Network for efficient and structure-preserving image synthesis from a single source domain to multiple target domains. SOM-GAN formulates the translation process as a pair of three-player games for each target domain with a dual-path adversarial design. Specifically, a multi-domain appearance discriminator encourages realistic target-domain texture rendering, while a multi-domain structure discriminator introduces explicit structural constraints by adversarially aligning the joint distributions of images and semantic annotations. Through alternating optimization of these adversarial objectives, SOM-GAN enables the multi-domain conditional generator to synthesize high-quality images across multiple target domains. Extensive experiments on four typical yet challenging multi-target medical image translation tasks demonstrate that SOM-GAN outperforms state-of-the-art methods in both synthesized-image fidelity and downstream clinical segmentation and diagnostic classification tasks. The code is released at https://github.com/JinhaoLi98/SOM-GAN.

如何引用

Computerized medical imaging and graphics : the official journal of the Computerized Medical Imaging Society 的其他文献

• 在跨域医学图像分割中，通过特征增强的加权平均实现连续的测试时间自适应