超网络摊销MeanFlow实现3D MRI单步重建
原标题：One Evaluation, Any Operating Point: Hypernetwork-Amortized MeanFlow for 3D MRI Reconstruction
AI 摘要
该论文提出一种用于加速3D MRI重建的MeanFlow方法，将数据一致性权重、加速倍数和笛卡尔采样模式整体视为可学习操作点，通过超网络映射到网络的逐通道调制，并端到端经过五步可微共轭梯度投影微调。研究发现学习采集模式带来的增益远大于其他操作点，且单步评估即可超越20步patch-diffusion先验，同时完全采样档案并非必需。
正文节选
One Evaluation, Any Operating Point: Hypernetwork-Amortized MeanFlow for 3D MRI Reconstruction Abstract Generative priors reconstruct accelerated 3D MRI well but pay twice at deployment: tens of network evaluations per volume, and hyperparameters re-tuned per protocol. A third cost is never counted because it is fixed before the model exists: the sampling pattern the scanner uses. We treat the whole operating point as an input. A 3D MeanFlow patch network (a one-step flow model, first trained as