arXiv 2603.165512026-09-29cs.CVcs.AI

CompDiff：分层组合扩散用于公平和零样本交叉性医学图像生成

CompDiff enables fair and zero shot medical image generation across demographic intersections through compositional diffusion

Institute of Data Science, Faculty of Science and Engineering, Maastricht University, Maastricht, The Netherlands（数据科学研究所，科学与工程学院，马斯特里赫特大学，马斯特里赫特，荷兰）；Department of Advanced Computing Sciences, Faculty of Science and Engineering, Maastricht University, Maastricht, The Netherlands（先进计算科学系，科学与工程学院，马斯特里赫特大学，马斯特里赫特，荷兰）；VITO, Belgium（比利时VITO研究院）

Mahmoud K. Ibrahim, Bart Elen, Chang Sun, Ahmad Jiblawi, Mohammed M. Saleh, Maryam K. Ibrahim, Gökhan Ertaylan, Michel Dumontier

备注：: v4: substantially revised version (new title, reader study, additional co-authors). 38 pages main text + 25 pages supplement

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

CompDiff通过分层组合扩散框架解决生成模型中因数据不平衡导致的公平性和零样本交叉性生成问题，在图像质量、子组公平性和零样本交叉性生成方面表现优异。

摘要：

生成模型越来越多地用于增强医学影像数据集以实现更公平的AI。然而，一个关键假设常常被忽视：生成器本身在不同人口群体中生成等高质量图像。训练在不平衡数据上的模型可能会继承这些不平衡，导致罕见子组的合成质量下降，并在训练中未出现的人口交叉组合时遇到困难。我们称此为不平衡生成器问题。现有的解决方案如损失重新加权在优化层面操作，当某些组合的训练信号稀缺或缺失时效果有限。我们提出CompDiff，一种分层组合扩散框架，该框架在表示层面解决此问题。一个专门的分层条件网络（HCN）分解人口条件，生成一个与CLIP嵌入拼接的demographic token作为交叉注意力上下文。这种结构化因子分解鼓励跨子组的参数共享，并支持对罕见或未见过的人口交叉组合的组合泛化。在胸部X光（MIMIC-CXR）和视网膜图像（FairGenMed）上的实验表明，CompDiff在图像质量（FID：64.3 vs. 75.1）、子组公平性（ES-FID）和零样本交叉性泛化（在保留交叉组合上提高高达21%的FID）方面与标准微调和FairDiffusion相比表现良好。在CompDiff生成数据上训练的下游分类器也显示出改进的AUROC和减少的种族偏见，表明人口条件的架构设计是公平医学图像生成中一个重要且未被充分探索的因素。代码可在https://anonymous.4open.science/r/CompDiff-6FE6获取。

英文摘要：

Medical image generators trained on imbalanced data can fail at demographic intersections absent from training. We introduce CompDiff, which encodes age, sex and race separately and composes supervised demographic tokens alongside clinical text. Across chest radiographs and fundus images, CompDiff improves overall and subgroup fidelity relative to prompt conditioning (RoentGen-v2) and loss reweighting (FairDiffusion). It generalises in zero-shot generation to 16 chest X-ray intersections excluded from training, achieving the lowest mean FID-RadImageNet in every intersection. In a blinded reader study of these unseen intersections, two radiologists gave CompDiff the highest mean scores among generators for anatomical realism and agreement with the clinical impression, and selected its images most often as the most realistic. Pretraining with CompDiff images improved downstream classification, while CompDiff audit cohorts reduced estimation error on rare intersections. These findings support compositional demographic conditioning for extending medical image synthesis to underserved populations. Code: https://github.com/mahmoudibrahim98/CompDiff

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。