MeDUET：用于 3D 医学图像合成与分析的解耦统一预训练框架
摘要
自监督学习（SSL）和扩散模型已推动表征学习和图像合成，但在 3D 医学成像领域仍 largely used separately for analysis and synthesis。将其统一具有诱人性，但困难在于，多源数据 exhibits pronounced style shifts，而下游任务主要依赖解剖结构，导致解剖内容与获取风格被耦合。在本文中，我们提出 MeDUET，一个 3D 医学图像 Disentangled UnifiEd PreTraining 框架，基于变分自编码器的潜在空间实现。我们的核心思想是将异构多中心数据下的统一预训练视为一个因子可识别问题，其中内容应持续捕获解剖结构，风格应持续捕获外观特征。MeDUET 通过三个组件解决此问题：Token demixing 提供因子分离的可控监督，Mixed Factor Token Distillation 减少混合区域中的因子泄漏，Swap-invariance Quadruplet Contrast 促进因子级不变性和判别性。通过学习到的这些因子，MeDUET 能有效迁移至合成与分析任务，yielding higher fidelity, faster convergence, and better controllability for synthesis, while achieving competitive or superior domain generalization and label efficiency on diverse medical benchmarks. Overall, MeDUET shows that multi-source heterogeneity can serve as useful supervision, with disentanglement providing an effective interface for unifying 3D medical image synthesis and analysis. Our code is available at https://github.com/JK-Liu7/MeDUET.
引用
@article{arxiv.2602.17901,
title = {MeDUET: Disentangled Unified Pretraining for 3D Medical Image Synthesis and Analysis},
author = {Junkai Liu and Ling Shao and Le Zhang},
journal= {arXiv preprint arXiv:2602.17901},
year = {2026}
}