Purpose: We introduce "Select and retrieve via direct upsampling" network (SARDU-Net), a data-driven framework for model-free quantitative MRI (qMRI) protocol design, and demonstrate it on in vivo brain and prostate diffusion-relaxation imaging (DRI). Methods: SARDU-Net selects subsets of informative measurements within lengthy pilot scans, without the requirement to identify tissue parameters for which to optimise for. The algorithm consists of a selector, identifying measurement subsets, and a predictor, estimating fully-sampled signals from the subsets. We implement both using deep neural networks, which are trained jointly end-to-end. We demonstrate the algorithm on brain (32 diffusion-/T1-weightings) and prostate (16 diffusion-/T2-weightings) DRI scans acquired on 3 healthy volunteers on two separate 3T Philips systems each. We used SARDU-Net to identify sub-protocols of fixed size, assessing the reproducibility of the procedure and testing sub-protocols for their potential to inform multi-contrast analyses via T1-weighted spherical mean diffusion tensor (T1-SMDT, brain) and hybrid multi-dimensional MRI (HM-MRI, prostate) modelling. Results: In both brain and prostate, SARDU-Net identifies sub-protocols that maximise information content in a reproducible manner across training instantiations. The sub-protocols enable multi-contrast modelling for which they were not optimised explicitly, providing robust T1- SMDT and HM-MRI maps and goodness-of-fit in the top 5% against extensive sub-protocol comparisons. Conclusions: SARDU-Net gives new opportunities to identify economical but informative qMRI protocols from a subset of the pilot scans that can be used for acquisition-time-sensitive applications. The simple architecture makes the algorithm easy to train when exhaustive searches are intractable, and applicable to a variety of anatomical contexts.

中文翻译：

---

“通过直接上采样选择和检索”网络（SARDU-Net）：一种用于定量MRI协议设计的数据驱动，无模型的深度学习方法

目的：我们引入“通过直接上采样选择和检索”网络（SARDU-Net），这是一种数据驱动的无模型定量MRI（qMRI）协议设计框架，并在体内大脑和前列腺扩散松弛成像中进行了演示（ DRI）。方法：SARDU-Net在冗长的先导扫描中选择信息量测量的子集，而无需识别要为其进行优化的组织参数。该算法包括一个选择器（用于识别测量子集）和一个预测器（用于从子集估计完全采样的信号）。我们使用端到端联合训练的深度神经网络来实现这两种方法。我们演示了在三个独立的3T Philips系统上分别对3位健康志愿者进行的大脑（32个扩散/ T1加权）和前列腺（16个扩散/ T2加权）DRI扫描的算法。我们使用SARDU-Net来识别固定大小的子协议，评估程序的可重复性，并测试子协议通过T1加权球均弥散张量（T1-SMDT，大脑）为多对比度分析提供信息的潜力，以及混合多维MRI（HM-MRI，前列腺）建模。结果：在大脑和前列腺中，SARDU-Net都可以识别子协议，这些子协议可以在训练实例之间以可重现的方式最大化信息内容。子协议可实现未明确优化的多对比度建模，可提供强大的T1-SMDT和HM-MRI映射，以及与广泛的子协议进行比较时前5％的拟合优度。结论：SARDU-Net提供了新的机会，可以从可用于采集时间敏感型应用程序的一部分先导扫描中识别经济却实用的qMRI协议。当穷举搜索难以解决时，简单的体系结构使该算法易于训练，并且适用于各种解剖环境。