View Presentation Video

View Presentation Video

Keywords: Image acquisition: Machine learning, Image acquisition: Image processing, Image acquisition: Reconstruction

Artificial intelligence/machine learning (AI/ML) based techniques have gathered interest as a possible means to improve MRI processing pipelines, with applications ranging from image reconstruction from raw data to extraction of quantitative biomarkers from imaging data in post-processing. Our purpose is to review existing and emerging AI/ML methods for various MRI processing applications. 

View Presentation Video

Arjun Desai^1,2, Beliz Gunel¹, Batu Ozturkler¹, Brian A Hargreaves^1,2, Garry E Gold², Shreyas Vasanawala², John Pauly¹, Christopher Ré³, and Akshay S Chaudhari^2,4
1Electrical Engineering, Stanford University, Stanford, CA, United States, ²Radiology, Stanford University, Stanford, CA, United States, ³Computer Science, Stanford University, Stanford, CA, United States, ⁴Biomedical Data Science, Stanford University, Stanford, CA, United States

Keywords: Image Reconstruction, Machine Learning/Artificial Intelligence, Artifacts

Deep learning (DL) has demonstrated promise for fast, high quality accelerated MRI reconstruction. However, current supervised methods require access to fully-sampled training data, and self-supervised methods are sensitive to out-of-distribution data (e.g. low-SNR, anatomy shifts, motion artifacts). In this work, we propose a self-supervised, consistency-based method for robust accelerated MRI reconstruction using physics-driven data priors (termed VORTEX-SS). We demonstrate that without any fully-sampled training data, VORTEX-SS 1) achieves high performance on in-distribution, artifact-free scans, 2) improves reconstructions for scans with physics-driven perturbations (e.g. noise, motion artifacts), and 3) generalizes to distribution shifts not modeled during training.

View Presentation Video

Sipei Li^1,2, Jianzhong He^2,3, Tengfei Xue^2,4, Guoqiang Xie^2,5, Shun Yao^2,6, Yuqian Chen^2,4, Erickson F. Torio², Yuanjing Feng³, Dhiego CA Bastos², Yogesh Rathi², Nikos Makris^2,7, Ron Kikinis², Wenya Linda Bi², Alexandra J Golby², Lauren J O’Donnell², and Fan Zhang^2
1University of Electronic Science and Technology of China, Chengdu, China, ²Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, United States, ³Zhejiang University of Technology, Hangzhou, China, ⁴University of Sydney, Sydney, Australia, ⁵Nuclear Industry 215 Hospital of Shaanxi Province, Xianyang, China, ⁶The First Affiliated Hospital, Sun Yat-sen University, Guangzhou, China, ⁷Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States

Keywords: Nerves, Brain

We present a novel deep learning framework, DeepRGVP, for the retinogeniculate pathway (RGVP) identification from dMRI tractography data. We propose a novel microstructure-supervised contrastive learning method (MicroSCL) that leverages both streamline labels and tissue microstructure (fractional anisotropy) for RGVP and non-RGVP. We propose a simple and effective streamline-level data augmentation method (StreamDA) to address highly imbalanced training data. We perform comparisons with three state-of-the-art methods on an RGVP dataset. Experimental results show that DeepRGVP has superior RGVP identification performance.

View Presentation Video

Felix Frederik Zimmermann¹, Andreas Kofler¹, Christoph Kolbitsch¹, and Patrick Schuenke^1
1Physikalisch-Technische Bundesanstalt (PTB), Berlin and Braunschweig, Germany

Keywords: Quantitative Imaging, Machine Learning/Artificial Intelligence

Typically, in quantitative MRI, an inverse problem of finding parameter maps from magnitude images has to be solved.  Neural networks can be applied to replace non-linear regression models and implicitly learn a suitable spatial regularization. However, labeled training data is often limited. Thus, we propose a combination of training on synthetic data and on unlabeled in-vivo data utilizing pseudo-labels and a Noise2Self-inspired technique. We present a convolutional neural network trained to predict T₁, B₀, and B₁ maps and their estimated aleatoric uncertainties from a single WASABITI scan.

View Presentation Video

Zihao Chen^1,2, Hsu-Lei Lee¹, Yibin Xie¹, Debiao Li^1,2, and Anthony Christodoulou^1,2
1Biomedical Imaging Research Institute, Cedars-Sinai Medical Center, Los Angeles, CA, United States, ²Department of Bioengineering, University of California, Los Angeles, Los Angeles, CA, United States

Keywords: Image Reconstruction, Machine Learning/Artificial Intelligence

CMR Multitasking is a promising approach for quantitative imaging without breath-holds or ECG monitoring but standard iterative reconstruction is too long for clinical use. Supervised artificial intelligence (AI) can accelerate reconstruction but lacks generalizability and transparency, and T1 mapping precision has not been sufficient. Here we propose an AI-Assisted Iterative (AAI) reconstruction which takes an AI reconstruction output as a “warm start” to a well-characterized iterative reconstruction algorithm with only 2 iterations. The proposed method produces better image fidelity and more precise T1 maps than other accelerated reconstruction methods, in less than 15 seconds (16x faster than conventional iterative reconstruction).

View Presentation Video

Maarten Terpstra^1,2, Sjors Verschuren^1,2, Tom Bruijnen^1,2, Matteo Maspero^1,2, and Cornelis van den Berg^1,2
1Department of Radiotherapy, University Medical Center Utrecht, Utrecht, Netherlands, ²Computational Imaging Group for MR Diagnostics & Therapy, University Medical Center Utrecht, Utrecht, Netherlands

Keywords: Pulse Sequence Design, Machine Learning/Artificial Intelligence, Autonomous sequence optimization

For MRI-guided interventions, tumor contrast and visibility are crucial. However, the tumor tissue parameters can significantly vary among subjects, with a range of T1, T2, and proton density values that may cause sub-optimal image quality when scanning with population-optimized protocols. Patient-specific sequence optimization could significantly increase image quality, but manual parameter optimization is infeasible due to the high number of parameters. Here, we propose to perform automatic patient-specific sequence optimization by applying deep reinforcement learning and reaching near-optimal SNR and CNR with minimal additional acquisitions.

View Presentation Video

Marta Gaviraghi¹, Antonio Ricciardi², Fulvia Palesi¹, Wallace Brownlee², Paolo Vitali^3,4, Ferran Prados^2,5,6, Baris Kanber^2,5, and Claudia A. M. Gandini Wheeler-Kingshott^1,2,7
1Department of Brain and Behavioural Sciences, University of Pavia, Pavia, Italy, ²NMR Research Unit, Department of Neuroinflammation, Queen Square Multiple Sclerosis Centre, UCL Queen Square Institute of Neurology, University College London (UCL), London, United Kingdom, ³Department of Radiology, IRCCS Policlinico San Donato, Milan, Italy, ⁴Department of Biomedical Sciences for Health, Università degli Studi di Milano, Milan, Italy, ⁵Department of Medical Physics and Biomedical Engineering, Centre for Medical Image Computing (CMIC), University College London, London, United Kingdom, ⁶Universitat Oberta de Catalunya, E-Health Center, Barcelona, Spain, ⁷Brain Connectivity Centre, IRCCS Mondino Foundation, Pavia, Italy

Keywords: Image Reconstruction, Machine Learning/Artificial Intelligence, fractional anisotropy

Quantitative maps obtained with diffusion weighted (DW) imaging such as fractional anisotropy (FA) are useful in pathologies. Often, to speed up acquisition time, the number of DW volumes acquired is reduced. We investigated the performance and clinical sensitivity of deep learning (DL) networks to calculate FA starting from different numbers of DW volumes. Using 4 or 7 volumes, clinical sensitivity was affected because no consistent differences between groups were found, contrary to our “one-minute FA” that uses 10 DW volumes. When developing DL for reduced acquisition data, the ability to generalize and biomarker sensitivity must be assessed.

View Presentation Video

Gokberk Elmas^1,2, Salman Ul Hassan Dar^1,2, Yilmaz Korkmaz^1,2, Muzaffer Ozbey^1,2, and Tolga Cukur^1,2,3
1Department Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey, ²National Magnetic Resonance Research Center (UMRAM), Bilkent University, Ankara, Turkey, ³Aysel Sabuncu Brain Research Center, Bilkent University, Ankara, Turkey

Keywords: Image Reconstruction, Machine Learning/Artificial Intelligence

Generalization performance in learning-based MRI reconstruction relies on comprehensive model training on large, diverse datasets collected at multiple institutions. Yet, centralized training after cross-site transfer of imaging data introduces patient privacy risks. Federated learning (FL) is a promising framework that enables collaborative training without explicit data sharing across sites. Here, we introduce a novel FL method for MRI reconstruction based on a multi-site deep generative model. To improve performance and reliability against data heterogeneity across sites, the proposed method decentrally trains a generative image prior decoupled from the imaging operator, and adapts it to minimize data-consistency loss during inference.

View Presentation Video

Keywords: Multi-Contrast, Machine Learning/Artificial Intelligence, harmonization, normalization, reconstruction

Harmonization is necessary for large-scale multi-site neuroimaging studies to reduce the variations due to factors such as image acquisition, imaging devices, and acquisition protocols. This so-called scanner effect significantly impacts multivariate analysis and the development of computational predictive models using MRI. Our approach utilized an unsupervised learning based model to build a mapping between MR data acquired from two different scanners. Results illustrate the potential of unsupervised deep learning algorithms to harmonize MRI data, as well as to improve downstream tasks by applying the harmonization.

View Presentation Video

Keywords: Osteoarthritis, Machine Learning/Artificial Intelligence

Identifying patients with knee osteoarthritis (OA) whom the disease will progress is critical in clinical practice. Currently, the time-series information and interactions between the structures and sub-regions of the whole knee are underused for predicting. Therefore, we propose a temporal-structural graph convolutional network (TSGCN) using time-series data of 194 cases and 406 OA comparators. Each sub-region was regarded as a vertex and represented by the extracted radiomics features, the edges between vertexs were established by the clinical prior knowledge. The multiple-modality TSGCN (integrating information of MRIs, clinical and image-based semi-quantitative score) performed best comparing to the radiomics and CNN model.

View Presentation Video

Haoyang Pei^1,2, Yixuan Lyu^2,3, Sebastian Lambrecht^4,5,6, Doris Lin⁵, Li Feng¹, Fang Liu⁷, Paul Nyquist⁸, Peter van Zijl^5,9, Linda Knutsson^5,9,10, and Xiang Xu^1,5
1Biomedical Engineering and Imaging Institute and Department of Radiology, Icahn School of Medicine at Mount Sinai, New York City, NY, United States, ²Department of Electrical and Computer Engineering, NYU Tandon School of Engineering, New York City, NY, United States, ³Image Processing Center, School of Astronautics, Beihang University, Beijing, China, ⁴Department of Neurology, Technical University of Munich, Munich, Germany, ⁵Department of Radiology, Johns Hopkins University, Baltimore, MD, United States, ⁶Institute of Neuroradiology, Ludwig-Maximilians-Universität, Munich, Germany, ⁷Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, Boston, MA, United States, ⁸Department of Neurology, Johns Hopkins University, Baltimore, MD, United States, ⁹F.M Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, MD, United States, ¹⁰Department of Medical Radiation Physics, Lund University, Lund, Sweden

Keywords: Machine Learning/Artificial Intelligence, Perfusion

This study built a deep-learning-based method to directly extract DSC MRI perfusion and perfusion related parameters from DCE MRI. A conditional generative adversarial network was modified to solve the pixel-to-pixel perfusion map generation problem. We demonstrate that in both healthy and brain tumor patients, highly realistic perfusion and perfusion related parameter maps can be synthesized from the DCE MRI using this deep-learning method. In healthy controls, the synthesized parameters had distribution similar to the ground truth DSC MRI values. In tumor regions, the synthesized parameters correlated linearly with the ground truth values.