“Select and retrieve via direct upsampling” network (SARDU-Net): a data-driven, model-free, deep learning approach for quantitative MRI protocol design

Abstract

PurposeWe introduce “Select and retrieve via direct upsampling” network (SARDU-Net), a data-driven deep learning framework for model-free quantitative MRI (qMRI) experiment design. Here we provide a practical demonstration of its utility on in vivo joint diffusion-relaxation imaging (DRI) of the prostate.MethodsSARDU-Net selects subsets of informative qMRI measurements within lengthy pilot scans. The algorithm consists of two deep neural networks (DNNs) that are trained jointly end-to-end: a selector, identifying a subset of input qMRI measurements, and a predictor, estimating fully-sampled signals from such a subset. We studied 3T prostate DRI scans performed on 3 healthy volunteers with 16 unique (b,TE) values (diffusion-/T2-weighting), and used SARDU-Net to identify sub-protocols of 12 and 9 measurements. The reproducibility of the sub-protocol selection was evaluated, and sub-protocols were assessed for their potential of informing multi-contrast analysis, as for example Hybrid Multi-dimensional MRI (HM-MRI).ResultsSARDU-Net identifies informative sub-protocols of specified size from a small number of pilot scans. The procedure is reproducible across training folds and random initialisations. Moreover, SARDU-Net sub-protocols corresponding to up to ~50% scan time reduction support downstream HM-MRI modelling for which they were not optimised explicitly, providing quality of fit in the top 5% of all tested sub-protocols.ConclusionsSARDU-Net gives new opportunity to identify economical but informative qMRI data sets for clinical applications under high time pressure in a fully data-driven way from a few pilot scans. The simple SARDU-Net architecture makes the algorithm easy to train and appealing when extensive sub-protocol searches are intractable.

Neural Networks for parameter estimation in microstructural MRI: a study with a high-dimensional diffusion-relaxation model of white matter microstructure

“…In this work, we tested the learning-based approach on protocols designed via Cramer-Rao lower-bound optimization (Alexander, 2008; Coelho et al, 2019; Lampinen et al, 2020). While not yet tested in conjunction with learning-based fitting pipelines, alternative optimization methods based on either efficient signal decomposition schemes (Bates et al, 2020; Song and Xiao, 2020) or deep learning algorithms for feature selection (Grussu et al, 2020b; Pizzolato et al, 2020) are also expected to have a positive impact on the performance of the DNN fitting approach.…”

On the generalizability of diffusion MRI signal representations across acquisition parameters, sequences and tissue types: chronicles of the MEMENTO challenge

“…Next to these two main families, there are also hybrid approaches that, for instance, aim to characterize the fibre orientation distribution without explicitly modelling the fibre composition (J. D. Tournier et al 2008;, or use a statistical model for different compartments (Scherrer et al 2016;Pasternak et al 2009;De Luca, Bertoldo, and Froeling 2017), which can be defined apriori or driven from the data (Keil et al 2017;De Luca et al 2018). Besides these "classical" approaches to model dMRI, the last couple of years have witnessed a vast increase in the number of machine learning techniques applied to dMRI to predict signal decay (Golkov et al 2016;Grussu et al 2020), fibre orientations (Poulin et al 2019; or the underlying tissue parameters (Nedjati-Gilani et al 2017).…”

Combined diffusion‐relaxometry microstructure imaging: Current status and future prospects

“…Similarly, a framework based on PCA‐like optimization has been proposed 63 and aims to retrieve the sparse basis from the data. Finally, recent frameworks using machine learning techniques such as SARDU‐net 64 have shown excellent ability to retrieve information based on fewer samples. 44 …”