Marcia Sahaya Louis^1,2, Huijun Vicky Liao², Rohit Singh³, Ajay Joshi¹, Jong Woo Lee⁴, and Alexander Lin^2
1ECE, Boston University, Boston, MA, United States, ²Radiology, Brigham and Women's Hospital, Boston, MA, United States, ³Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology, Cambridge, MA, United States, ⁴Brigham and Women's Hospital, Boston, MA, United States

More than half of patients who undergo targeted temperature management (TTM) after cardiac arrest do not survive hospitalization and 50% of those survivors suffer from long-term cognitive deficits.  The goal of this study is to use machine learning methods to characterize the pattern of metabolic changes in patients with good and poor outcomes after cardiac arrest.  A machine learning pipeline that incorporates z-scores, decision-tree modeling, principal component analysis, and linear support vector machine was applied to MR spectroscopy data acquired after cardiac arrest.  Results confirm that N-acetylaspartate and lactate are important markers but other unexpected findings emerged as well.

Hoai Nam Dang¹, Jonathan Endres¹, Felix Glang², Simon Weinmüller¹, Alexander Loktyushin², Klaus Scheffler², Arnd Doerfler¹, Andreas Maier³, and Moritz Zaiss^1,2
1Department of Neuroradiology, University Clinic Erlangen, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany, ²Magnetic Resonance Center, Max-Planck-Institute for Biological Cybernetics, Tübingen, Germany, ³Department of Computer Science, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany

We propose an end-to-end optimization approach to reduce T2 blurring in single-shot TSE sequences, by performing a joint optimization of Flip Angle Design and DenseNet parameters using he original transverse magnetization image at a certain TE as target. Our approach generalizes well to in vivo measurements at 3T, revealing enhanced white and grey matter contrast and fine structures in human brain. 

Iulian Emil Tampu^1,2, Ida Blystad^2,3,4, Neda Haj-Hosseini¹, and Anders Eklund^1,5
1Biomedical Engineering, Linkoping University, Linkoping, Sweden, ²Center for Medical Image Science and Visualization (CMIV), Linkoping University, Linkoping, Sweden, ³Department of Radiology in Linköping, Region Östergötland, Center for Diagnostics, Linkoping, Sweden, ⁴Department of Health, Medicine and Caring Sciences, Division of Diagnostics and Specialist Medicine, Linkoping University, Linkoping, Sweden, ⁵Department of Computer and Information Science, Linkoping University, Linkoping, Sweden

Manual annotation of gliomas in magnetic resonance (MR) images is a laborious task, and it is impossible to identify active tumor regions not enhanced in the conventionally acquired MR modalities. Recently, quantitative MRI (qMRI) has shown capability in capturing tumor-like values beyond the visible tumor structure. Aiming at addressing the challenges of manual annotation, qMRI data was used to train a 2D U-Net deep-learning model for brain tumor segmentation. Results on the available data show that a 7% higher Dice score is obtained when training the model on qMRI post-contrast images compared to when the conventional MR images are used.

Edward J Peake¹, Andy N Priest^1,2, and Martin J Graves^1,2
1Imaging, Cambridge University Hospitals NHS Foundation Trust, Cambridge, United Kingdom, ²Department of Radiology, University of Cambridge, Cambridge, United Kingdom

Radiomic features are sensitive to changes in imaging parameters in MRI. This makes it challenging to develop robust machine learning models using imaging features.  We explore the effect of clinically available deep learning image reconstruction on the performance of radiomic features. Correlation coefficient values varied (0.56 - 1.00) when comparing radiomic features of deep learning reconstructed images and ‘conventional’ MRI scans. The noise reduction level had a large impact on correlation coefficients, but variations were also significant between different types of imaging feature. Identification of highly correlated features may help identify more stable sets of radiomic features for machine learning.

Tobias Goodwin-Allcock¹, Robert Gray², Parashkev Nachev², Jason McEwan³, and Hui Zhang^1
1Department of Computer Science and Centre for Medical Image Computing, UCL, London, United Kingdom, ²Department of Brain Repair & Rehabilitation, Institute of Neurology, UCL, London, United Kingdom, ³Kagenova Limited, Guildford, United Kingdom

We demonstrate the advantages of spherical convolutional neural networks over conventional fully connected networks at estimating rotationally invariant microstructure indices. Fully-connected networks (FCN) have outperformed conventional model fitting for estimating microstructure indices, such as FA. However, these methods are not robust to changes diffusion weighted image sampling scheme nor are they rotationally equivariant. Recently spherical-CNN have been supposed as a solution to this problem. However, the advantages of spherical-CNNs have not been leveraged. We demonstrate both spherical-CNNs robust to new gradient schemes as well as the rotational equivariance. This has potential to decrease the number of training datapoints required.

Ethan J Ulrich¹, Jasser Dhaouadi¹, Robben Schat², Benjamin Spilseth², and Randall Jones^1
1Bot Image, Omaha, NE, United States, ²Radiology, University of Minnesota, Minneapolis, MN, United States

Multiple prostate cancer detection AI models—including random forest, neural network, XGBoost, and a novel boosted parallel random forest (bpRF)—are trained and tested using radiomics features from 958 bi-parametric MRI (bpMRI) studies from 5 different MRI platforms. After data preprocessing—consisting of prostate segmentation, registration, and intensity normalization—radiomic features are extracted from the images at the pixel level. The AI models are evaluated using 5-fold cross-validation for their ability to detect and classify cancerous prostate lesions. The free-response ROC (FROC) analysis demonstrates the superior performance of the bpRF model at detecting prostate cancer and reducing false positives.

Jake Valsamis¹, Nicholas Luciw¹, Nandinee Haq¹, Sarah Atwi ¹, Simon Duchesne ^2,3, William Cameron¹, and Bradley J MacIntosh^1,4
1Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada, ²Radiology Department, Faculty of Medicine, Laval University, Québec, QC, Canada, ³Quebec City Mental Health Institute, Québec, QC, Canada, ⁴Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada

Persistent exposure to highly pulsatile blood can damage the brain’s microvasculature. A convenient method for measuring cerebral pulsatility would allow investigation into its relationship with vascular dysfunction and cognitive decline. In this work, we propose a convolutional neural network (CNN) based deep learning solution to estimate cerebral pulsatility using only the frequency content from BOLD MRI scans. Various frequency component inputs were assessed, and echo time dependence was evaluated with a 5-fold cross-validation. Pulsatility was estimated from BOLD MRI data acquired on a different scanner to assess generalizability. The CNN reliably estimated pulsatility and was robust to various scan parameters. 

Ian Robert Storey¹, Catherine Anne Spilling^1,2, Xujiong Ye³, Thomas Richard Barrick¹, and Franklyn Arron Howe^1
1St George's, University of London, London, United Kingdom, ²King's College London, London, United Kingdom, ³University of Lincoln, Lincoln, United Kingdom

A fully connected neural network (FCN) was trained to map a short diffusion weighted image acquisition to high quality Quasi-Diffusion Imaging (QDI) parameter maps. The FCN produced denoised and enhanced QDI parameter maps compared to weighted least squares fitting of data to the QDI model. The FCN shows generalisation to unseen pathology such as grade IV glioma dMRI data and demonstrates the FCN can produce high quality QDI tensor maps from clinically feasible 2 minute data acquisitions. An FCN further enhances the ability of QDI to provide  non-Gaussian diffusion imaging within clinically feasible acquisition times.  

Shihan Qiu^1,2, Anthony G. Christodoulou^1,2, Yibin Xie¹, and Debiao Li^1,2
1Biomedical Imaging Research Institute, Cedars-Sinai Medical Center, Los Angeles, CA, United States, ²Department of Bioengineering, UCLA, Los Angeles, CA, United States

Deep learning methods have been developed to estimate quantitative maps from conventional weighted images, which has the potential to improve the availability and clinical impact of quantitative MRI. However, high-resolution labels required for network training are not commonly available in practice. In this work, a hybrid supervised and physics-informed self-supervised loss function was proposed to train parameter estimation networks when only limited low-resolution labels are accessible. By taking advantage of high-resolution information from the input weighted images, the proposed method generated sharp quantitative maps and had improved performance over the supervised training method purely relying on the low-resolution labels.