Santiago Coelho¹, Els Fieremans¹, and Dmitry S. Novikov^1
1Center for Biomedical Imaging and Center for Advanced Imaging Innovation and Research (CAI$^2$R), Department of Radiology, New York University School of Medicine, New York, NY, United States

In the machine-learning (ML) era, we are transitioning from max-likelihood parameter estimation to learning the mapping from measurements to model parameters. While such maps look smooth, there is danger of them becoming too smooth: At low SNR, ML estimates become the mean of the training set. Here we derive fit quality (MSE) as function of SNR, and show that MSE for various ML methods (regression, neural-nets, random forest) approaches a universal curve interpolating between Cramér-Rao bound at high SNR, and variance of the prior at low SNR. Theory is validated numerically and on white matter Standard Model in vivo.

Filip Szczepankiewicz¹ and Jens Sjölund^2
1Clinical Sciences Lund, Lund University, Lund, Sweden, ²Department of Information Technology, Uppsala University, Uppsala, Sweden

Diffusion weighted imaging is perturbed by the presence of background gradients, or so-called 'cross-terms,' causing bias in estimated parameters and fiber orientations. In this work, we present a novel gradient waveform design that removes the cross-term sensitivity entirely. This design is valuable for diffusion MRI methods that are otherwise corrupted by background gradients, and it also facilitates arbitrary sequence timing, b-tensor shapes and suppression of concomitant gradient effects.

Misha P. T. Kaandorp^1,2,3, Sebastiano Barbieri⁴, Remy Klaassen⁵, Hanneke W.M. van Laarhoven⁵, Hans Crezee⁶, Peter T. While^2,3, Aart J. Nederveen¹, and Oliver J. Gurney-Champion^1
1Department of Radiology and Nuclear Medicine, Amsterdam UMC, Amsterdam, Netherlands, ²Department of Radiology and Nuclear Medicine, St. Olav's University Hospital, Trondheim, Norway, ³Department of Circulation and Medical Imaging, NTNU: Norwegian University of Science and Technology, Trondheim, Norway, ⁴Centre for Big Data Research in Health, UNSW, Sydney, Australia, ⁵Department of Medical Oncology, Amsterdam UMC, Amsterdam, Netherlands, ⁶Department of Radiation Oncology, Amsterdam UMC, Amsterdam, Netherlands

We implemented an improved unsupervised physics-informed deep neural network approach for intravoxel-incoherent motion modeling to DWI data by exploring several hyperparameters. Whereas the original IVIM-NET_orig showed high dependency between the predicted IVIM parameters, our optimized approach resolved this high dependency, produced better accuracy and was more consistent. In simulations, IVIM-NET_optim outperformed least-squares and Bayesian fitting approaches. In patients with pancreatic ductal adenocarcinoma, IVIM-NET_optim produced substantially less noisy parameter maps and lower intersession within-subject standard deviations than the alternatives. IVIM-NET_optim also detected the most individual patients with significant parameter changes in the group of patients who received chemoradiotherapy.

Noemi G. Gyori^1,2, Marco Palombo¹, Christopher A. Clark², Hui Zhang¹, and Daniel C. Alexander^1
1Centre for Medical Image Computing, University College London, London, United Kingdom, ²Great Ormond Street Institute of Child Health, University College London, London, United Kingdom

The performance of supervised machine learning tools is only as good as the data used to train them. In this work, we investigate the impact of training data distribution on tissue microstructure estimates in the human brain. We focus on two strategies: uniform sampling from the entire parameter space and sampling from parameter combinations observed using traditional model fitting. We demonstrate that training on previously observed combinations may be advantageous for detecting small variations in healthy tissue. However, for detecting atypical tissue abnormalities, our results favour uniform training data sampling in which all plausible parameter combinations are represented. 

João Pedro de Almeida Martins^1,2, Markus Nilsson¹, Björn Lampinen³, Marco Palombo⁴, Carl-Fredrik Westin^5,6, and Filip Szczepankiewicz^1,5,6
1Department of Clinical Sciences, Radiology, Lund University, Lund, Sweden, ²Department of Radiology and Nuclear Medicine, St. Olav's University Hospital, Trondheim, Norway, ³Department of Clinical Sciences, Medical Radiation Physics, Lund University, Lund, Sweden, ⁴Centre for Medical Image Computing and Dept of Computer Science, University College London, London, United Kingdom, ⁵Radiology, Brigham and Women’s Hospital, Boston, MA, United States, ⁶Harvard Medical School, Boston, MA, United States

The application of function fitting neural networks in microstructural MRI has so far been restricted to lower-dimensional biophysical models. Moreover, the data sufficiency requirements of learning-based approaches remain unclear. Here, we use supervised learning to vastly accelerate the fitting of a high-dimensional relaxation-diffusion model of tissue microstructure and develop analysis tools for assessing the accuracy and sensitivity of model fitting networks. The developed learning-based fitting pipelines were tested on relaxation-diffusion data acquired with optimal and sub-optimal protocols. We found no evidence that machine-learning algorithms can correct for a degenerate fitting landscape or replace a careful design of the acquisition protocol.

Gregory Simchick^1,2, Ruiqi Geng^1,2, Yuxin Zhang^1,2, and Diego Hernando^1,2,3
1Radiology, University of Wisconsin-Madison, Madison, WI, United States, ²Medical Physics, University of Wisconsin-Madison, Madison, WI, United States, ³Biomedical Engineering, University of Wisconsin-Madison, Madison, WI, United States

Multiple studies have demonstrated the potential utility of intravoxel incoherent motion (IVIM) DWI in the evaluation of liver disease. However, IVIM estimates obtained using conventional monopolar diffusion gradient waveforms often suffer from high variability and instability. In this work, Cramer-Rao lower bound (CRLB) optimization was performed to determine the optimal sampling of b-M₁-space based on noise performance. Then, based on this optimized sampling, b-M₁-optimized waveforms were designed and employed in order to obtain IVIM estimates with improved stability in the right lobe of the liver in volunteers in comparison to conventional monopolar waveforms.

Rita Alves¹, Rafael Neto Henriques¹, Sune Nørhøj Jespersen^2,3, and Noam Shemesh^1
1Champalimaud Research, Champalimaud Centre for the Unknown, Lisbon, Portugal, ²Center of Functionally Integrative Neuroscience (CFIN) and MINDLab, Department of Clinical Medicine, Aarhus University, Aarhus, Denmark, ³Department of Physics and Astronomy, Aarhus University, Aarhus, Denmark

Correlation Tensor MRI (CTI) has been recently introduced for resolving the underlying sources of diffusional kurtosis. Here, we aimed to resolve the underlying kurtosis sources in ischemic tissue. Ex and in vivo CTI experiments in a mouse model of ischemia revealed enhanced sensitivity and specificity compared to their conventional counterparts. Our results suggest that microscopic kurtosis – associated with restricted diffusion and structural disorder – substantially contributes to the total kurtosis excess likely reflecting excitotoxic properties. Kurtosis associated with diffusion magnitude variance better reflected edema and free water. These first results are promising for elucidating biological factors in ischemia.

Deneb Boito^1,2, Magnus Herberthson³, Tom Dela Haije⁴, and Evren Özarslan^1,2
1Department of Biomedical Engineering, Linköping University, Linköping, Sweden, ²Center for Medical Image Science and Visualization, Linköping University, Linköping, Sweden, ³Department of Mathematics, Linköping University, Linköping, Sweden, ⁴Department of Computer Science, University of Copenhagen, Copenhagen, Denmark

We explore the feasibility of reducing scan time while retaining interpretable results when analysing multidimensional diffusion MRI data. The methods commonly employed for the estimation of the QTI model parameters might suffer severely from data under-sampling, making their use limited in clinical practice where acquisition time is limited. We show that by imposing relevant positivity constraints in the estimation of the mean, covariance, and moment tensors, the QTI+ framework strongly improves the quality of the derived maps, particularly so for data acquisition protocols featuring few number of volumes.

Benjamin Ades-Aron¹, Santiago Coelho¹, Jelle Veraart¹, Gregory Lemberskiy¹, Genevieve Barroll¹, Steven Baete¹, Timothy Shepherd¹, Dmitry S. Novikov¹, and Els Fieremans^1
1Radiology, NYU School of Medicine, New York, NY, United States

Translation of diffusion MRI-derived quantitative biomarkers into clinical decision making has been hampered by within-scanner and cross-scanner variability. We compare intra-scan, cross-scan, and cross-protocol variability of multi-shell diffusion MRI for three subjects and evaluated two Random Matrix Theory (RMT)-based denoising techniques to enhance repeatability. Without denoising, best scan-rescan repeatability was found for intra-scanner measurements with the highest SNR (shortest TE), RMT-based denoising greatly reduced variability across scanners and TE, resulting in coefficients of variation about 5% for all comparisons. Increased precision across scanners and protocols should increase statistical power and further enable clinical trials of quantitative higher-order diffusion MRI.

Eric Seth Michael¹, Franciszek Hennel¹, and Klaas Paul Pruessmann^1
1Institute for Biomedical Engineering, ETH Zurich and University of Zurich, Zurich, Switzerland

Oscillating gradient spin-echo (OGSE) sequences typically suffer low diffusion sensitization and long echo times, thereby compromising the diffusion-to-noise ratio (DNR). To address these issues, recently proposed OGSE shapes providing increased diffusion sensitization were combined with single-shot spiral readout trajectories using a high-performance gradient system. This implementation was used to study the frequency dependence of diffusivity in the in-vivo human brain for OGSE frequencies up to 125 Hz and b-values up to 1000 s/mm², yielding an improved depiction of diffusion dispersion. The results indicate the significance of higher b-values in characterizing this relationship.