Camila Munoz¹, Sam Ellis¹, Stephan G Nekolla², Karl P Kunze^1,3, Teresa Vitadello⁴, Radhouene Neji^1,3, René M. Botnar¹, Julia A. Schnabel¹, Andrew J. Reader¹, and Claudia Prieto^1
1School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, ²Nuklearmedizinische Klinik und Poliklinik, Technische Universitat Munchen, Munich, Germany, ³MR Research Collaborations, Siemens Healthcare Limited, Frimley, United Kingdom, ⁴Department of Internal Medicine I, University hospital rechts der Isar, Technical University of Munich, Munich, Germany

Simultaneous PET-MR has shown promise for addressing several of the technical challenges that may degrade image quality in PET imaging, such as high noise levels, attenuation artefacts, and motion artefacts. While state-of-the-art PET image reconstruction techniques have addressed these issues separately, their combined effect has not been demonstrated. Here we introduce a single framework that integrates MR-based motion correction and anatomical guidance for improved simultaneous diagnostic cardiac PET-MR imaging. We evaluated the proposed framework on cardiac [¹⁸F]FDG PET-MR datasets and results show that, compared to conventional reconstruction algorithms, our framework results in sharper images, with increased contrast and reduced noise. 

Heeseung Lim¹, Benjamin Wilk ^1,2, Jane Sykes ¹, John Butler ¹, Gerald Moran³, Jonathan Thiessen^1,2, Gerald Wisenberg^1,4, and Frank S Prato^1,2
1Lawson Health Research Institute, London, ON, Canada, ²Medical Biophysics, Western University, London, ON, Canada, ³Siemens Healthcare Limited, Oakville, ON, Canada, ⁴MyHealth Centre, Arva, ON, Canada

This study investigates the impact of motion in myocardial PET/MRI data by registering pre-mortem images to post-mortem images from a simultaneously acquired PET/MRI scan. After a registration in MRI, PET data is transformed and analyzed using the Patlak model. There are significant differences and correlation between pre- and post-mortem PET images. There are discrepancies in different segments of heart, but none were found to be significantly different. These results suggest that the level of inflammation in the heart might be misrepresented. Further comparison with motion corrected images will provide more concrete discrepancy due to motion for myocardial PET/MRI imaging.   

Mehdi Khalighi¹, Timothy Deller², Kevin Chen¹, Tyler Toueg³, Dawn Holley¹, Kim Halbert¹, Floris Jansen², Elizabeth Mormino³, Michael Zeineh¹, Farshad Moradi¹, Greg Zaharchuk¹, and Andrei Iagaru^1
1Radiology, Stanford University, Stanford, CA, United States, ²Engineering Dept., GE Healthcare, Waukesha, WI, United States, ³Neurology, Stanford University, Stanford, CA, United States

After PET images are reconstructed by OSEM, a spatial filter which exploits the correlation between neighboring voxels, is applied to remove noise. A new filtering method is proposed that also exploits the correlation between voxels from the same tissue. Anatomical priors are processed for bias correction and registered to PET images. For each voxel, similar voxels within the PET volume are identified using KNN method in the feature-space built by anatomical priors. These similar voxels and also the neighboring voxels are then used to remove the high frequency noise on PET images using a Gaussian and a weighted averaging filter.

Patrick Korf¹, Wolfgang Thaiss², Ambros J. Beer², Meinrad Beer³, Dominik Nickel¹, and Thomas Vahle^1
1Siemens Healthcare GmbH, Erlangen, Germany, ²Department of Nuclear Medicine, University Hospital Ulm, Ulm, Germany, ³Department of Diagnostic and Interventional Radiology, University Hospital Ulm, Ulm, Germany

In whole-body PET/MR exams, MR-based attenuation correction is usually performed with a Dixon protocol of an MR VIBE sequence acquired in breath-hold followed by a segmentation into different tissue classes. As an extension we present a free-breathing approach for attenuation correction that can be used for patients that have problems or are even unable to perform the required breath-holds. The presented approach relies on a self-gated, compressed sensing accelerated gradient-echo sequence with Cartesian k-space sampling. We demonstrate the generation of free-breathing attenuation maps in 2 human volunteers and 10 patients.

Charles McGrath¹, Mohammed M Albannay¹, Alexander Jaffray¹, Christian Guenthner¹, and Sebastian Kozerke^1
1Institute for Biomedical Engineering, University and ETH Zurich, Zurich, Switzerland

The signal-to-noise ratio (SNR) is a key metric of imaging performance, however many formulations do not account for relaxation properties and sequence timing effects, which play a decisive role when studying the effect of static field strength $$$B_0$$$ on SNR. We have developed first-principle simulations that incorporate these effects in order to estimate SNR scaling relations, and show that SNR can deviate significantly from previous scaling laws, specifically for lower field strengths and when sequence restrictions apply.

John Adams^1,2, William Handler^1,2, and Blaine Chronik^1,2
1Department of Physics and Astronomy, Western University, London, ON, Canada, ²xMR Labs, London, ON, Canada

Renewed interest in clinical low-field MR systems has opened up a new design space for MR pulse sequence design. To explore these opportunities, and to better inform hardware design, we are developing a flexible simulation tool based on a numerical simulation of the Bloch equations. This tool will both be able to model pulse sequences under the influence of realistic applied fields, and account for changes in relaxation time due to changes in field strength. This abstract presents our work to date in developing this tool.

Bo Peng^1,2,3, Baohua Hu^1,2,3, Mao Sheng⁴, Yuqi Liu⁴, Zhongchang Miao⁵, Zijun Dong⁶, Jian Bao⁷, SiSeung Kim⁷, Bing Keong Li⁷, and Yakang Dai^1,2,3
1Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, China, ²Suzhou Key Laboratory of Medical and Health Information Technology, Suzhou, China, ³Jinan Guoke Medical Engineering Technology Development co., Ltd., Jinan, China, ⁴Department of Radiology, Children’s Hospital of Soochow University, Suzhou, China, ⁵Department of Radiology, The First People’s Hospital of Lianyungang, Jiangsu Province, China, ⁶Department of Medical Imaging, Lianyungang Women and Children Hospital and Health Institute, Jiangsu Province, China, ⁷Jiangsu LiCi Medical Device Co., Ltd., Lianyungang, China

Low-field MRI is foreseeable as a safer system for infants. However, low-field MR images have lower SNR and spatial resolution as compared to high-field images, thus processing of low-field infant brain MR image is challenging. In this study, an automated image processing method that can accurately perform brain extraction, tissue segmentation, and brain labeling on low-field infant brain MR images is developed. It is also capable to automatically construct the inner, middle, and outer surfaces of the cerebral cortex and provides automatic quantitative analysis of selected region of interest, which can be a helpful tool for researchers in neuroimaging studies.

Merel de Leeuw den Bouter¹, Martin van Gijzen¹, and Rob Remis^2
1Delft Institute of Applied Mathematics, Delft University of Technology, Delft, Netherlands, ²Circuits and Systems, Delft University of Technology, Delft, Netherlands

We present an image reconstruction approach that incorporates regularization by multiplying the data-fidelity term by a total variation functional. Usually, regularization is carried out in an additive manner, with a regularization parameter balancing out the two terms. Such a parameter often needs to be tuned for each dataset through extensive numerical experimentation. Our approach does not require such a parameter. We applied the method to in-vivo data acquired using a low-field MR scanner. Our results show that the method successfully denoises low-field MR images.

Jo Schlemper¹, Neel Dey², Seyed Sadegh Mohseni Salehi¹, Carole Lazarus¹, Rafael O'Halloran¹, Prantik Kundu^1,3, and Michal Sofka^1
1Hyperfine Research Inc., Guilford, CT, United States, ²New York University, New York, NY, United States, ³Icahn School of Medicine at Mount Sinai, New York, NY, United States

An unsupervised deep learning framework is proposed for denoising low-field 64 mT diffusion-weighted MRI images (DWI). The denoised DWI (b-value = 890 s/mm²) and apparent diffusion coefficient (ADC) maps were evaluated in a user study by four expert graders in terms of sharpness, noise reduction, and overall utility. Our framework was found to enable low-field DWI restoration with strong noise while maintaining relevant image features. 62.50% and 64.28% of processed images were rated clearly/far better overall for DWI and ADC, respectively, with only 0.05% of processed DWI and 0% of processed ADC rated clearly/far worse.

Paul Wang¹, Michael Mullen², Lance DelaBarre², and Michael Garwood^2
1Center for Magnetic Resonance Research and Department of Biomedical Engineering, University of Minnesota, Minneapolis, MN, United States, ²Center for Magnetic Resonance Research and Department of Radiology, University of Minnesota, Minneapolis, MN, United States

In this work, we investigate whether an established method to correct distortions in conventional MRI can be repurposed to correct distortions arising from the nonlinearity of B₁ gradients when performing RF-encoded MRI at low field, where SAR constraints are reduced. Although several methods are capable of correcting image distortions arising from nonlinear B₀ gradients and/or large B₀ inhomogeneity, here we chose to adapt the method of Weis et al. Through theory and simulations, we demonstrate the ability to correct image distortions arising from nonlinear B₁ gradients in RF-encoded MRI.

Reina Ayde¹, Tobias Senft¹, Najat Salameh¹, and Mathieu Sarracanie^1
1Center for Adaptable MRI Technology (AMT Center), Department of Biomedical Engineering, University of Basel, Allschwil, Switzerland

Low magnetic field (LF) MRI is currently gaining momentum as a complementary, more flexible and cost-effective approach to MRI diagnosis. However, the impaired Signal-to-Noise Ratio, leading in turn to prolonged acquisition times, challenges its relevance at the clinical level. Recently, reconstructing an alias-free image using deep learning techniques has shown promising results. In this study, we leverage deep learning reconstruction to demonstrate the feasibility of highly undersampled (20% sampling) 3D LF MRI at 0.1 T. The model performance has been evaluated on both retrospective and acquired, prospective 3D LF data.

Zhuopin Sun¹, Steven Meikle^2,3, and Fernando Calamante^1,3,4
1School of Biomedical Engineering, The University of Sydney, Sydney, Australia, ²Faculty of Medicine and Health, The University of Sydney, Sydney, Australia, ³Brain and Mind Centre, The University of Sydney, Sydney, Australia, ⁴Sydney Imaging, The University of Sydney, Sydney, Australia

Recent advances in hybrid PET-MRI systems enable simultaneous acquisition of PET and MR data. PET is used to visualize and measure biochemically-specific metabolic processes, but has limited spatial resolution and signal-to-noise ratio. Combining diffusion MRI (dMRI) and PET data, which provide highly complementary information (e.g. structural connectivity and molecular information), has rarely been exploited previously in image postprocessing. The proposed CONNectome-based Non-Local Means (CONN-NLM) exploits synergies between dMRI-derived structural connectivity and PET intensity information to denoise PET data. This method is based on the rationale that structurally-connected voxels and voxels with similar intensity should be highly weighted when smoothing noise.

Kevin T. Chen¹, Olalekan Adeyeri², Tyler N Toueg³, Elizabeth Mormino³, Mehdi Khalighi¹, and Greg Zaharchuk^1
1Radiology, Stanford University, Stanford, CA, United States, ²Salem State University, Salem, MA, United States, ³Neurology and Neurological Sciences, Stanford University, Stanford, CA, United States

We have previously trained a deep learning network with simultaneous PET/MRI inputs to generate diagnostic quality images from ultra-low-dose amyloid PET acquisitions. With data bias being a known issue in deep learning-based applications, we aim to investigate whether this network could generalize to ultra-low-dose tau PET image enhancement. Results of this study show that data bias across radiotracers needs to be accounted for before applying an ultra-low-dose network trained on one tracer to another.

Ramy Hussein¹, Moss Zhao¹, Jia Guo², Kevin Chen¹, David Shin³, Michael Moseley¹, and Greg Zaharchuk^1
1Radiology, Stanford University, Stanford, CA, United States, ²Bioengineering, University of California, Riverside, Riverside, CA, United States, ³Neuro MR, GE Healthcare, Menlo Park, CA, United States

In this study, we demonstrate that an optimized 3D encoder-decoder structured convolutional neural network with attention gates can effectively integrate brain structural MRI and ASL perfusion images to produce high-quality synthetic PET CBF maps without using radiotracers. We performed experiments to evaluate different loss functions and the role of the attention mechanism. Our results showed that attention-based 3D encoder-decoder network with custom loss function produces the superior PET CBF prediction results, achieving SSIM of 0.94, MSE of 0.00025, and PSNR of 38dB.

Praveen Dassanayake^1,2, Lumeng Cui³, Elizabeth Finger^2,4, Andrea Soddu^5,6, Bjoern Jakoby⁷, Keith St. Lawrence^1,2, Gerald Moran⁸, and Udunna Anazodo^1,2
1Department of Medical Biophysics, University of Western Ontario, London, ON, Canada, ²Lawson Health Research Institute, London, ON, Canada, ³Division of Biomedical Engineering, University of Saskatchewan, Saskatoon, SK, Canada, ⁴Department of Clinical Neurological Sciences, University of Western Ontario, London, ON, Canada, ⁵Brain and Mind Institute, University of Western Ontario, London, ON, Canada, ⁶Department of Physics and Astronomy, University of Western Ontario, London, ON, Canada, ⁷Department of Physics, University of Surrey, Guildford, United Kingdom, ⁸Siemens Healthineers, Oakville, ON, Canada

Absolute quantification of tracer uptake in positron emission tomography (PET) requires the knowledge of an arterial input function (AIF) which involves invasive arterial blood sampling. Alternatively, input functions can be extracted by identifying feeding arteries from PET images. In this study we validated a software that uses magnetic resonance images to identify the feeding arteries in PET images to generate image derived input function (IDIF) for absolute quantification of PET. The ratio of area under curve between IDIFs and AIFs revealed that this tool can generate accurate IDIFs for non-invasive PET quantification.

Thibault Marin¹, Yanis Djebra^1,2, Paul Han¹, Vanessa Landes³, Yue Zhuo¹, Kuan-Hao Su⁴, Georges El Fakhri¹, and Chao Ma^1
1Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States, ²LTCI, Telecom Paris, Institut Polytechnique de Paris, Paris, France, ³GE Healthcare, Boston, MA, United States, ⁴GE Healthcare, Waukesha, WI, United States

Motion during acquisition of PET/MR data can severely degrade image quality of PET/MR studies. We have previously reported an MR-based motion correction technique capable of correcting for irregular motion patterns such as bulk motion and irregular respiratory motion. The method is based on a subspace MR model enabling reconstruction of real-time volumetric MR images (9 frames per second). In this work, we present improvements to the motion estimation method used to obtain motion fields from real-time MR images. We compare the performance of three packages for irregular motion patterns.

Xiaomeng Wu¹, Shuai Liu¹, Li Huo², Xihai Zhao³, and Fei Shang^1
1Department of Biomedical Engineering, School of Life Science, Beijing institute of technology, beijing, China, ²Department of Nuclear Medicine, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China, ³Department of Biomedical Engineering, Center for Biomedical Imaging Research, Tsinghua University School of Medicine, Beijing, China

Multi-modality images can provide comprehensive information for clinical diagnosis. In this study, a  registration method for cardiac PET and MR images was proposed by combining global registration and local registration. During global registration, axial PET and CT images were transformed to short axis with MR-survey as link. Then local registration was performed between PET/CT and MR-cine images, it was found that using the fusion of PET and CT images performed better compared with that only using a single modality image.