View Presentation Video

Aya Ghoul¹, Kerstin Hammernik^2,3, Patrick Krumm⁴, Sergios Gatidis^1,5, and Thomas Küstner^1
1Medical Image And Data Analysis (MIDAS.lab), Department of Diagnostic and Interventional Radiology, University Hospital of Tuebingen, Tuebingen, Germany, ²Lab for AI in Medicine, Technical University of Munich, Munich, Germany, ³Department of Computing, Imperial College London, London, United Kingdom, ⁴Department of Radiology, University Hospital of Tuebingen, Tuebingen, Germany, ⁵Max Planck Institute for Intelligent Systems, Tuebingen, Germany

Keywords: Machine Learning/Artificial Intelligence, Motion Correction, Image registration, Image Reconstruction

Motion-resolved reconstruction methods permit for considerable acceleration for cardiac CINE acquisition. Solving for the non-rigid cardiac motion is computationally demanding, and even more challenging in highly accelerated acquisitions, due to the undersampling artifacts in image domain. Here, we introduce a novel deep learning-based image registration network, GMA-RAFT, for estimating cardiac motion from accelerated imaging. A transformer-based module enhances the iterative recurrent refinement of the estimated motion by introducing structural self-similarities into the decoded features. Experiments on Cartesian and radial trajectories demonstrate superior results compared to other deep learning and state-of-the-art baselines in terms of motion estimation and motion-compensated reconstruction.

View Presentation Video

Gustavo Chau Loo Kung*^1,2, Emmanuelle M. M. Weber*², Juliet Knowles³, Ankita Batra³, Lijun Ni³, and Jennifer McNab^2
1Bioengineering Department, Stanford University, Stanford, CA, United States, ²Radiology Department, Stanford University, Stanford, CA, United States, ³Neurology Department, Stanford University, Stanford, CA, United States

Keywords: Machine Learning/Artificial Intelligence, Microstructure, Histology, Diffusion Imaging, g-ratio, axon diameter

To better establish the influence of histological features on the MRI signal, we present a multi-task neural network trained to predict parametrized microstructural distributions (axon diameters and g-ratios) from diffusion and magnetization transfer MRI data. To begin, we trained the model using histologically-derived synthetic MRI data  before applying transfer learning by fine tuning on empirical data. Our initial results on both synthetic and empirical ex vivo mouse brain MRI data demonstrate the feasibility of this approach.

View Presentation Video

Jiaxin Xiao¹, Zihan Li², Berkin Bilgic^3,4, Jonathan R. Polimeni^3,4, Susie Huang^3,4, and Qiyuan Tian^3,4
1Department of Electronic Engineering, Tsinghua University, Beijing, China, ²Department of Biomedical Engineering, Tsinghua University, Beijing, China, ³Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Charlestown, MA, United States, ⁴Harvard Medical School, Boston, MA, United States

Keywords: Machine Learning/Artificial Intelligence, Data Analysis

Neural network (NN) based approaches for super-resolution MRI typically require high-SNR high-resolution reference data acquired in many subjects, which is time consuming and a barrier to feasible and accessible implementation. We propose to train NNs for Super-Resolution using Noisy Reference data (SRNR), leveraging the mechanism of the classic NN-based denoising method Noise2Noise. We systematically demonstrate that results from NNs trained using noisy and high-SNR references are similar for both simulated and empirical data. SRNR suggests a smaller number of repetitions of high-resolution reference data can be used to simplify the training data preparation for super-resolution MRI.

View Presentation Video

Nalini M. Singh^1,2, Malte Hoffmann^3,4, Elfar Adalsteinsson^2,5,6, Bruce Fischl^2,3,4, Polina Golland*^1,5,6, Adrian V. Dalca*^1,3,4, and Robert Frost*^3,4
1Computer Science and Artificial Intelligence Laboratory (CSAIL), MIT, Cambridge, MA, United States, ²Harvard-MIT Health Sciences and Technology, MIT, Cambridge, MA, United States, ³Department of Radiology, Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ⁴Department of Radiology, Harvard Medical School, Boston, MA, United States, ⁵Department of Electrical Engineering and Computer Science, MIT, Cambridge, MA, United States, ⁶Institute for Medical Engineering and Science, MIT, Cambridge, MA, United States

Keywords: Machine Learning/Artificial Intelligence, Motion Correction, Image Reconstruction, Deep Learning

We demonstrate a deep learning approach for fast retrospective intraslice rigid motion correction in segmented multislice MRI. A hypernetwork uses auxiliary rigid motion parameter estimates to produce a reconstruction network based on the motion parameters that are specific to the input  image. This strategy produces higher quality reconstructions than those produced by model-based techniques or by networks that do not use motion estimates. Further, this approach mitigates sensitivity to misestimation of the motion parameters.

View Presentation Video

Malte Hoffmann^1,2, Nalini M Singh^3,4, Adrian V Dalca^1,2,3, Bruce Fischl^1,2,3,4, and Robert Frost^1,2
1Department of Radiology, Harvard Medical School, Boston, MA, United States, ²Department of Radiology, Massachusetts General Hospital, Boston, MA, United States, ³Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA, United States, ⁴Harvard-MIT Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, United States

Keywords: Machine Learning/Artificial Intelligence, Artifacts, deep learning, AI-guided radiology, neuroimaging, computer vision

Subject motion remains the major source of artifacts in magnetic resonance imaging (MRI). Motion correction approaches have been successfully applied in research, but clinical MRI typically involves repeating corrupted acquisitions. To alleviate this inefficiency, we propose a deep-learning strategy for training networks that predict a quality rating from the first few shots of accelerated multi-shot multi-slice acquisitions, scans frequently used for neuroradiological screening. We demonstrate accurate prediction of the scan outcome from partial acquisitions, assuming no further motion. This technology has the potential to inform the operator's decision on aborting corrupted scans early instead of waiting until the acquisition completes.

View Presentation Video

Adam Lim^1,2, Justin Lo^1,2, Matthias Wagner³, Birgit Ertl-Wagner^3,4, and Dafna Sussman^1,2,5
1Department of Electrical, Computer and Biomedical Engineering, Toronto Metropolitan University, Toronto, ON, Canada, ²Institute for Biomedical Engineering, Science and Technology (iBEST), Toronto Metropolitan University and St. Michael’s Hospital, Toronto, ON, Canada, ³Division of Neuroradiology, The Hospital for Sick Children, Toronto, ON, Canada, ⁴Department of Obstetrics and Gynecology, University of Toronto, Toronto, ON, Canada, ⁵Department of Medical Imaging, University of Toronto, Toronto, ON, Canada

Keywords: Machine Learning/Artificial Intelligence, Artifacts, Deep Learning, Generative Adversarial Network, Image Denoising

Motion artifacts are a common issue in fetal MR imaging that limit the visibility of essential fetal anatomy. In such cases, the sequence acquisition must be repeated in order for an accurate diagnosis. This study introduces a deep learning approach utilizing a Generative Adversarial Network (GAN) framework for removing motion artifacts in fetal MRIs. Results exceeded current state-of-the-art methods by achieving an average SSIM of 93.7%, and PSNR of 33.5dB. The presented network demonstrates rapid and accurate results that can be advantageous in clinical use.

View Presentation Video

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

The diagnostic confidence in the interpretation of cardiac MR scans can be improved by enhancing the signal-to-noise ratio (SNR). Traditional image denoising has been studied extensively to improve SNR in cardiac MRI, but with limited success due to the resulting blurring. In this study, we sought to develop and evaluate cardiac MR denoising inline neural network (CaDIN) for improving SNR in cardiac MRI.

View Presentation Video

Yunpeng Zhang¹, Huixiang Zhuang¹, Ziyu Meng¹, Ruihao Liu^1,2, Wen Jin^2,3, Wenli Li¹, Zhi-Pei Liang^2,3, and Yao Li^1
1School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China, ²Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, IL, United States, ³Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, United States

Keywords: Machine Learning/Artificial Intelligence, Segmentation

Accurate segmentation of brain tissues is important for brain imaging applications. Learning the high-dimensional spatial-intensity distributions of brain tissues is challenging for classical Bayesian classification and deep learning-based methods. This paper presents a new method that synergistically integrate a tissue spatial prior in the form of a mixture-of-eigenmodes with deep learning-based classification. Leveraging the spatial prior, a Bayesian classifier and a cluster of patch-based position-dependent neural networks were built to capture global and local spatial-intensity distributions, respectively. By combining the spatial prior, Bayesian classifier, and the proposed networks, our method significantly improved the segmentation performance compared with the state-of-the-art methods.

Zhuo Chen¹, Juan Gao¹, Xin Tang¹, and Chenxi Hu^1
1The Institute of Medical Imaging Technology, School of Biomedical Engineering, Shanghai Jiao Tong University (SJTU), Shanghai, China

Keywords: Machine Learning/Artificial Intelligence, Artifacts, Banding artifacts, Flow artifacts

bSSFP cine imaging suffers from banding and flow artifacts in the region of off-resonance. Suppressing one kind of artifacts may evoke the other kind. For example, phase cycling suppresses banding artifacts, yet its acquisition at multiple frequency offsets often evokes flow artifacts. Here, we develop a partially interpretable neural network for jointly suppressing banding and flow artifacts without phase cycling. Based on a single cine image, the method generates an artifact-corrected image and a voxel-identity map, which guides the artifact suppression and improves its interpretability. Preliminary investigation shows that the method reduces banding and flow artifacts without introducing new artifacts.

View Presentation Video

Srivathsa Pasumarthi Venkata¹ and Ryan Chamberlain^1
1Research & Development, Subtle Medical Inc, Menlo Park, CA, United States

Keywords: Machine Learning/Artificial Intelligence, Brain, Image Registration

Deep learning based affine registration is a fast and computationally efficient alternative to conventional iterative methods. However, existing solutions are not sensitive to local misalignments. We propose a non-rigid guided affine registration network with stochastic depth which was designed with an affine branch and an optional non-rigid branch. The probability of dropping the non-rigid branch was gradually increased over training epochs. During inference, the non-rigid branch was fully removed, thus making it a pure affine network guided by non-rigid transformations. Model training and quantitative evaluation was performed using a pre-registered multi-contrast brain MRI public dataset.

View Presentation Video

Nataliia Molchanova^1,2,3, Vatsal Raina^3,4, Francesco La Rosa⁵, Andrey Malinin⁶, Henning Müller³, Mark Gales⁴, Cristina Granziera⁷, Mara Graziani^3,8, and Merirxell Bach Cuadra^1,9
1Radiology department, Lausanne University Hospital (CHUV), Lausanne, Switzerland, ²Doctoral School of the Faculty of Biology and Medicine, University of Lausanne (UNIL), Lausanne, Switzerland, ³University of Applied Sciences of Western Switzerland, Sierre, Switzerland, ⁴University of Cambridge, Cambridge, United Kingdom, ⁵Icahn School of Medicine at Mount Sinai, New York, NY, United States, ⁶Shifts Project, Helsinki, Finland, ⁷University Hospital Basel, Basel, Switzerland, ⁸IBM Research Europe, Zurich, Switzerland, ⁹Center for Biomedical Imaging (CIBM), University of Lausanne, Lausanne, Switzerland

Keywords: Machine Learning/Artificial Intelligence, Multiple Sclerosis, Machine learning/Artificial intelligence, Brain, Uncertainty estimation, Reliable AI

We approach the problem of quantifying the degree of reliability of supervised deep learning models used by clinicians for automatic multiple sclerosis lesion segmentation on MRI. In particular, we quantify the correspondence of various uncertainty measures to the errors that a deep learning model makes in overall segmentation or lesion detection. The evaluation is done both on in- and out-of- domain datasets (40 and 99 patients respectively), and provides insights about the measures that can point clinicians to potential errors of an automatic algorithm regardless of the distributional shift.

View Presentation Video

Diego Fajardo-Rojas¹, Riine Heinsalu¹, Megan Hall², Mary Rutherford³, Joseph Hajnal⁴, Emma Robinson⁴, Lisa Story², and Jana Hutter^3
1School of Biomedical Engineering & Imaging Sciences, King's College London, London, United Kingdom, ²Department of Women & Children's Health, Guy's and St Thomas' NHS Foundation Trust, London, United Kingdom, ³Department of Perinatal Imaging & Health, King's College London, London, United Kingdom, ⁴Department of Biomedical Engineering, King's College London, London, United Kingdom

Keywords: Machine Learning/Artificial Intelligence, Machine Learning/Artificial Intelligence

The accurate prediction of preterm birth is a clinically crucial but challenging problem due to its complex aetiology. In this work, data from fetal anatomical and functional multi-organ MRI acquisitions are used to train Random Forests and Support Vector Machines to predict gestational age at delivery. These predictions are classified as 'term' or 'preterm'. The model with highest sensitivity, a Random Forest, achieved 0.85 sensitivity, 0.81 accuracy, 0.8 specificity, 1.99 weeks Mean Absolute Error, and 0.58 R2 score. This work proves the potential of Machine Learning models trained on anatomical and functional MRI data to predict gestational age at delivery.

View Presentation Video

Antonio Candito¹, Martina Torcè², Richard Holbrey³, Alina Dragan¹, Christina Messiou¹, Nina Tunariu¹, Dow-Mu Koh¹, and Matthew D Blackledge^1
1The Institute of Cancer Research, London, United Kingdom, ²Imperial College London, London, United Kingdom, ³Mint Medical, Heidelberg, Germany

Keywords: Machine Learning/Artificial Intelligence, Segmentation

Whole-Body Diffusion Weighted Imaging (WBDWI) requires automated tools that delineate malignant bone disease based on high b-value signal intensity, leading to state-of-the-art imaging biomarkers of response. As an initial step, we have developed an automated deep-learning pipeline that automatically delineates the skeleton from WBDWI. Our approach is trained on paired examples, where ground truth is defined through a set of weak labels (non-binary segmentations) derived from a computationally expensive atlas-based segmentation approach. The model showed on average a dice score, precision and recall between the manual and derived skeleton segmentations on test datasets of 0.74, 0.78, and 0.7, respectively.

View Presentation Video

Natalia Konovalova¹, Aniket Tolpadi^1,2, Felix Liu¹, Rupsa Bhattacharjee¹, Felix Gassert¹, Paula Giesler¹, Johanna Luitjens¹, Sharmila Majumdar¹, and Valentina Pedoia^1
1Department of Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, ²Department of Bioengineering, University of California, Berkeley, Berkeley, CA, United States

Keywords: Joints, Machine Learning/Artificial Intelligence

Meniscal lesions are a common knee pathology, but pathology detection from MRI is usually evaluated on full-length acquisitions. We trained UNet and KIKI I-Net reconstruction algorithms with several loss function configurations, showing k-space losses are not required to obtain robust reconstructions. We trained and evaluated Faster R-CNN to detect meniscal anomalies, showing similar performance on R=8 reconstructions and fully-sampled images, demonstrating its utility as an assessment tool for reconstruction performance and indicating reconstructed images are viable for downstream clinical postprocessing tasks.

View Presentation Video

Sandrine Bédard¹, Adrian El Baz¹, Uzay Macar¹, and Julien Cohen-Adad^1,2,3,4
1NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montreal, Montréal, QC, Canada, ²Functional Neuroimaging Unit, CRIUGM, University of Montreal, Montréal, QC, Canada, ³Mila - Quebec AI Institute, Montréal, QC, Canada, ⁴Centre de recherche du CHU Sainte-Justine, Université de Montréal, Montréal, QC, Canada

Keywords: Machine Learning/Artificial Intelligence, Segmentation

Several methods to segment the spinal cord have emerged over the past decade. However, they are dependent on the image contrast, resulting in differences of spinal cord cross-sectional area (CSA), a relevant biomarker in neurodegenerative diseases. We propose a novel  method using deep learning that produces the same segmentation regardless of the MRI contrast. Moreover, the segmentation is “soft” (non-binary) and can therefore encode partial volume information. CSA computed with this contrast-agnostic soft segmentation method has lower intra- and inter-subject variability, making it particularly relevant for multi-center studies.