Philip M. Adamson¹, Jeffrey Dominic¹, Arjun Desai¹, Christian Bluethgen², Jeff P. Wood³, Ali B. Syed², Robert Boutin², Kathryn J. Stevens², Daniel Spielman², Shreyas Vasanawala², John M. Pauly¹, Akshay S. Chaudhari², and Beliz Gunel^1
1Department of Electrical Engineering, Stanford University, Palo Alto, CA, United States, ²Department of Radiology, Stanford University, Palo Alto, CA, United States, ³Austin Radiological Association, Austin, TX, United States

Evaluation of accelerated magnetic resonance imaging (MRI) reconstruction methods is imperfect due to the discordance between quantitative image quality metrics (IQMs) and radiologist-perceived image quality. Self-supervised learning (SSL) is a deep learning (DL) method that has become a popular pre-training tool due to its ability to capture generalizable and domain-specific feature representations of the underlying data without the need for labels. In this study, we derive a data-driven self-supervised feature distance (SSFD) IQM to assess MR image reconstruction quality. We demonstrate that SSFD is more highly correlated to three radiologist’s perceived image quality on DL-based sparse reconstructions than conventional IQMs.

Patrick Salome^1,2,3,4, Francesco Sforazzini^1,2,3,4, Andreas Kudak^3,5,6, Matthias Dostal^3,5,6, Nina Bougatf^3,5,6, Jürgen Debus^3,4,5,7, Maximilian Knoll^1,3,4,5, and Amir Abdollahi^1,3,4,5
1CCU Translational Radiation Oncology, German Cancer Research Center (DKFZ), Heidelberg, Germany, ²Medical Faculty, Heidelberg University Hospital, Heidelberg, Germany, ³Heidelberg Ion-Beam Therapy Center (HIT), Heidelberg, Germany, ⁴German Cancer Consortium (DKTK) Core Center, Heidelberg, Germany, ⁵Radiation Oncology, Heidelberg University Hospital, Heidelberg, Germany, ⁶CCU Radiation Therapy, German Cancer Research Center (DKFZ), Heidelberg, Germany, ⁷National Center for Tumor Diseases (NCT), Heidelberg, Germany

MR-based radiomics prognostic signatures have great clinical potential but are currently hindered due to a lack of standardisation in the radiomics workflow. This study focuses on a crucial step of these workflows, i.e. intensity normalisation, while presenting a methodology that allows for determining the most suitable MR intensity normalization method for a specific entity.

Raviteja Kotikalapudi^1,2, Balint Kincses^1,2,3, Kevin Hoffschlag¹, Matthias Zunhammer², Tobias Schmidt-Wilcke^4,5, Zsigmond T Kincses³, Ulrike Bingel², and Tamas Spisak^1,2
1Laboratory of Predictive NeuroImaging, ​​Institute for Diagnostic and Interventional Radiology and Neuroradiology, University Hospital Essen, Essen, Germany, ²The Bingel Laboratory, Translational Pain Research Unit, University Hospital Essen, Essen, Germany, ³Department of Neurology, University of Szeged, Szeged, Hungary, ⁴Institute for Clinical Neurosciences and Medical Psychology, Heinrich Heine University, Dusseldorf, Germany, ⁵Neurocentre, District Hospital Mainkofen, Mainkofen, Deggendorf, Germany

Individual sensitivity to pain is both a precursor and a symptom of many clinical pain conditions. A pain predictive model would have potential applications in objectively characterizing pain in acute and chronic pain individuals. Here, we developed a cortical thickness-based predictive model of pain sensitivity using a machine learning approach and multi-centre T₁-weighted MRI and quantitative pain threshold data. We found that our model significantly predicts pain sensitivity, that was measured through heat, cold and mechanical stimuli. Furthermore, the predictions were exclusively driven by cortical thickness and not confounded by variables of demographic and psychological value.

Eric K. Gibbons¹, Zhongnan Fang², Arjun D. Desai³, Christopher M. Sandino³, Garry E. Gold⁴, Brian A. Hargreaves⁴, and Akshay S. Chaudhari^4
1Department of Electrical and Computer Engineering, Weber State University, Ogden, UT, United States, ²Lvis Corporation, Palo Alto, CA, United States, ³Department of Electrical Engineering, Stanford University, Stanford, CA, United States, ⁴Department of Radiology, Stanford University, Stanford, CA, United States

Deep-learning (DL) can be used to extend compressed sensing (CS) to learn the regularization function in a data-driven manner. In contrast, super resolution (SR) algorithms have been used to transform rapidly-acquired low-resolution images into higher-resolution images. This work compares DL-CS with DL-SR for accelerated MRI on a test dataset of 50 patients with conventional image quality metrics and clinically-relevant quantitative T₂ relaxation measurements. We demonstrate that DLCS approaches outperform DLSR approaches for accelerated MRI. 

Matteo Figini^1,2, Hongxiang Lin^1,2,3, Felice D'Arco⁴, Godwin Inalegwu Ogbole⁵, Maria Camilla Rossi Espagnet⁶, Olalekan Ibukun Oyinloye⁷, Joseph O Yaria⁸, Donald Amasike Nzeh⁷, Mojisola Omolola Atalabi⁹, Lisa Ronan^1,2, David W Carmichael^10,11, Judith Helen Cross^4,11, Ikeoluwa A Lagunju¹², Delmiro Fernandez-Reyes^2,12, and Daniel C Alexander^1,2
1Centre for Medical Image Computing, University College London, London, United Kingdom, ²Computer Science, University College London, London, United Kingdom, ³Research Center for Healthcare Data Science, Zhejiang Lab, Hangzhou, China, ⁴Great Ormond Street Hospital for Children, London, United Kingdom, ⁵Radiology, College of Medicine, University of Ibadan, Ibadan, Nigeria, ⁶Neuroradiology, Sapienza University, Rome, Italy, ⁷Radiology, University of Ilorin Teaching Hospital, Ilorin, Nigeria, ⁸Neurology, University College Hospital Ibadan, Ibadan, Nigeria, ⁹Radiology, University College Hospital Ibadan, Ibadan, Nigeria, ¹⁰School of Biomedical Engineering & Imaging Sciences, King's College London, London, United Kingdom, ¹¹UCL Great Ormond Street Institute of Child Health, London, United Kingdom, ¹²Paediatrics, College of Medicine, University of Ibadan, Ibadan, Nigeria

We applied Image Quality Transfer to enhance the contrast and resolution in the slice direction of low-field structural MRI, using a deep learning model trained on simulated images. Six radiologists blindly reviewed the enhanced images compared to low- and high-field images from 12 paediatric patients with epilepsy. Results demonstrated significant improvement of the visualisation of brain structures in sagittal and coronal orientations, and marginal improvement of the contrast between grey and white matter. If these promising results are confirmed in a larger study and in lesions, IQT could be an important tool to enhance the diagnostic power of low-field MRI.

Soumick Chatterjee^1,2,3, Alessandro Sciarra^1,4, Max Dünnwald^3,4, Anitha Bhat Talagini Ashoka³, Mayura Gurjar Cheepinahalli Vasudeva³, Shudarsan Saravanan³, Venkatesh Thirugnana Sambandham³, Steffen Oeltze-Jafra^4,5, Oliver Speck^1,5,6,7, and Andreas Nürnberger^2,3,5
1Department of Biomedical Magnetic Resonance, Otto von Guericke University Magdeburg, Magdeburg, Germany, ²Data and Knowledge Engineering Group, Otto von Guericke University Magdeburg, Magdeburg, Germany, ³Faculty of Computer Science, Otto von Guericke University Magdeburg, Magdeburg, Germany, ⁴MedDigit, Department of Neurology, Medical Faculty, University Hospital, Magdeburg, Germany, ⁵Center for Behavioral Brain Sciences, Magdeburg, Germany, ⁶German Centre for NeurodegenerativeDiseases, Magdeburg, Germany, ⁷Leibniz Institute for Neurobiology, Magdeburg, Germany

Many deep learning-based techniques have been proposed in recent years to reconstruct undersampled MRI – showing their potential for shortening the acquisition time. Before using them in actual practice, they are usually evaluated by comparing their results against the available ground-truth – which is not available during real applications. This research shows the potential of using uncertainty estimation to evaluate the reconstructions without using any ground-truth images. The method has been evaluated for the task of super-resolution MRI, for acceleration factors ranging from two to four – in all three dimensions.

Priscille de Dumast^1,2, Hamza Kebiri^1,2, Meritxell Bach Cuadra^1,2, and Hélène Lajous^1,2
1CIBM Center for Biomedical Imaging, Lausanne, Switzerland, ²Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland

Manual fetal brain tissue segmentation is needed for training machine learning methods but is a tedious and error-prone task. The generation of synthetic magnetic resonance images can overcome the lack of clinical annotations by supplementing scarce clinical fetal datasets. However, we highlight that the choice of the numerical model from which additional data are derived is key to maximize the segmentation accuracy of clinical data via domain adaptation strategies. We demonstrate that the resort to high-resolution segmented images from real neurotypical and pathological cases enhances the morphological variability compared to an atlas, resulting in improved fetal brain tissue delineation overall.

Mohammed R. S. Sunoqrot¹, Kirsten M. Selnæs^1,2, Bendik S. Abrahamsen¹, Alexandros Patsanis¹, Gabriel A. Nketiah^1,2, Tone F. Bathen^1,2, and Mattijs Elschot^1,2
1Department of Circulation and Medical Imaging, Norwegian University of Science and Technology (NTNU), Trondheim, Norway, ²Department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway

Multiparametric MRI (mpMRI) is a valuable tool for the diagnosis of prostate cancer. Computer-aided detection and diagnosis (CAD) systems have the potential to improve robustness and efficiency compared to traditional radiological reading of mpMRI in prostate cancer. Co-registration of diffusion-weighted and T2-weighted images is a crucial step of CAD but is not always flawless. Automated detection of poorly co-registered cases would therefore be a useful supplement. This work shows that a fully automated quality control system for co-registration of prostate MR images based on deep learning has potential for this purpose.

Kevin Blansit¹, Zhehao Hu¹, Greg Zaharchuk¹, Enhao Gong¹, and Keshav Datta^1
1Subtle Medical, Menlo Park, CA, United States

Faster MRI scans can be achieved by acquiring low resolution images or low SNR images and enhancing them to standard-of-care using deep learning techniques. However, to achieve clinical diagnostic quality images, this requires a large number of paired clinical datasets to train the model. Here we show that a multi-task deep convolutional neural network (DCNN) trained using only simulated motion artifact, low SNR, and low resolution images is capable of improving the quality of clinically acquired images from motion corrupted and accelerated sequences.

Robert Wujek¹, Melissa Prah², Mona Al-Gizawiy², and Kathleen Schmainda^2
1Graduate School, Medical College of Wisconsin, Wauwatosa, WI, United States, ²Biophysics, Medical College of Wisconsin, Wauwatosa, WI, United States

Delineation of invasive tumor from peritumoral edematous tissue remains a major obstacle to glioma treatment. To address this problem, a neural network was trained to distinguish between these regions using biopsies paired with colocalized MRI inputs. In addition to histologically confirmed biopsies, virtual biopsies sampled from non-contrast enhancing, FLAIR enhancing regions of non-invasive tumors (meningioma, metastasis) were used with an assumed classification of “non-tumor”. The current work is a preliminary assessment of this assumption and it's impact on model performance.
 

Yiheng Li^1
1Subtle Medical, Santa Clara, CA, United States

To perform robust quality control of medical imaging registration, we propose a method that can QC co-registration for multiple organs, without the restriction of the modalities. A rule-based image synthesis pipeline is used to generate random contrasts and shapes as training images. ResNet34 is trained to predict image alignment. Two MRI datasets with the spine or brain as subjects are used as external validation sets. The proposed model trained with synthetic images is validated on either one of the real MRI datasets and outperforms the same model trained on the other MRI dataset, which shows better generalizability.

Joonhyeok Yoon¹, Sooyeon Ji¹, Eun-Jung Choi¹, Hwihun Jeong¹, and Jongho Lee^1
1Seoul National University, Seoul, Korea, Republic of

CycleGAN shows good performance with harmonization task. However, generative models have risk of structure modification. we proposed a cross-vendor harmonization model with paired CycleGAN based architecture for both high performance and structural consistency. We acquired 4 in-vivo dataset from Siemens and Philips scanner. For algorithm, we adapted CycleGAN for generation model, and we utilized L1 loss from pix2pix and patchGAN discriminator for structural consistency. Evaluations were performed both quantitatively and qualitatively. To quantitative evaluation, we assessed means of structural similarity index measure (SSIM). Proposed model shows better results compared to CycleGAN architecture.

Vincent Roca¹, Grégory Kuchcinski^1,2, Morgan Gautherot¹, Xavier Leclerc^1,2, Jean-Pierre Pruvo^1,2, and Renaud Lopes^1,2
159000, Univ Lille, UMS 2014 – US 41 – PLBS – Plateformes Lilloises en Biologie & Santé, Lille, France, ²59000, Univ Lille, Inserm, Lille Neuroscience & Cognition, Lille, France

In MRI multicentric studies, inter-scanner harmonization is necessary to avoid taking into account variations due to technical differences in the analysis. In this study, we focused on CycleGAN models for 3D T1 weighted brain images harmonization. More precisely, we didn't follow the classical 2D CycleGAN architecure and developped a 3D cycleGAN model. We compared harmonization quality of these two kinds of models using 20 imaging features quantifying T1 signal, contrast between brain structures and segmentation quality. Results illustrate the potential of 3D CycleGAN for better synthesize images in inter-scanner MRI harmonization tasks.