Adam C Richie-Halford¹, Jason Yeatman², Noah Simon³, and Ariel Rokem^4
1eScienceInstitute, University of Washington, Seattle, WA, United States, ²Graduate School of Education and Division of Developmental and Behavioral Pediatrics, Stanford University, Stanford, CA, United States, ³Department of Biostatistics, University of Washington, Seattle, WA, United States, ⁴Department of Psychology, University of Washington, Seattle, WA, United States

We present a novel method for the analysis of diffusion MRI tractometry data based on the sparse group lasso. It capitalizes on natural anatomical grouping of diffusion metrics, providing both accurate prediction of phenotypic information and results that are readily interpretable. We show the effectiveness of this approach in two settings. In a classification setting, patients with amyotrophic lateral sclerosis (ALS) are accurately distinguished from matched controls and SGL automatically identifies known anatomical correlates of ALS. In a regression setting, we accurately predict “brain age” in two previous dMRI studies. We demonstrate that our approach is both accurate and interpretable.

Madiha Arshad¹, Mahmood Qureshi¹, Omair Inam¹, and Hammad Omer^1
1Medical Image Processing Research Group (MIPRG), Department of Electrical and Computer Engineering, COMSATS University, Islamabad, Pakistan

Fast and accurate tissue extraction of human brain is an ongoing challenge. Two principal factors make this task difficult:(1) quality of the reconstructed images, (2) accuracy and availability of the segmentation masks. In this proposed method, firstly, a supervised deep learning framework is used for the reconstruction of solution image from the acquired uniformly under-sampled human brain data. Later, an unsupervised clustering approach i.e. k-means is used for the extraction of specific tissue from the reconstructed image. Experimental results show a successful extraction of cerebrospinal fluid (CSF), white matter and grey matter from the human brain image.

Soumick Chatterjee^1,2,3, Alessandro Sciarra^1,4, Max Dünnwald^3,4, Shubham Kumar Agrawal³, Pavan Tummala³, Disha Setlur³, Aman Kalra³, Aishwarya Jauhari³, Steffen Oeltze-Jafra^4,5,6, Oliver Speck^1,5,6,7, and Andreas Nürnberger^2,3,6
1Department of Biomedical Magnetic Resonance, Otto von Guericke University, Magdeburg, Germany, ²Data and Knowledge Engineering Group, Otto von Guericke University, Magdeburg, Germany, ³Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany, ⁴MedDigit, Department of Neurology, Medical Faculty, University Hopspital, Magdeburg, Germany, ⁵German Centre for Neurodegenerative Diseases, Magdeburg, Germany, ⁶Center for Behavioral Brain Sciences, Magdeburg, Germany, ⁷Leibniz Institute for Neurobiology, Magdeburg, Germany

While commonly used approach for disease localization, we propose an approach to detect anomalies by differentiating them from reliable models of anatomies without pathologies. The method is based on a Variational Auto Encoder to learn the anomaly free distribution of the anatomy and a novel image subtraction approach to obtain pixel-precise segmentation of the anomalous regions. The proposed model has been trained with the MOOD dataset. Evaluation is done on BraTS 2019 dataset and a subset of the MOOD, which contain anomalies to be detected by the model.

Soumick Chatterjee^1,2,3, Arnab Das³, Chirag Mandal³, Budhaditya Mukhopadhyay³, Manish Vipinraj³, Aniruddh Shukla³, Oliver Speck^1,4,5,6, and Andreas Nürnberger^2,3,6
1Department of Biomedical Magnetic Resonance, Otto von Guericke University, Magdeburg, Germany, ²Data and Knowledge Engineering Group, Otto von Guericke University, Magdeburg, Germany, ³Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany, ⁴German Centre for Neurodegenerative Diseases, Magdeburg, Germany, ⁵Leibniz Institute for Neurobiology, Magdeburg, Germany, ⁶Center for Behavioral Brain Sciences, Magdeburg, Germany

In medical image analysis, it is desirable to decipher the black-box nature of Deep Learning models in order to build confidence in clinicians while using such methods. Interpretability techniques can help understand the model’s reasonings, e.g. by showcasing the anatomical areas the network focuses on. While most of the available interpretability techniques work with classification models, this work presents various interpretability techniques for segmentation models and shows experiments on a vessel segmentation model. In particular,  we focus on input attributions and layer attribution methods which give insights on the critical features of the image identified by the model.

Alfonso Mastropietro¹, Daniele Procissi², Elisa Scalco¹, Giovanna Rizzo¹, and Nicola Bertolino^2
1Istituto di Tecnologie Biomediche, Consiglio Nazionale delle Ricerche, Segrate, Italy, ²Radiology, Northwestern University, Chicago, IL, United States

Fitting the IVIM bi-exponential model is challenging especially at low SNRs and time consuming. In this work we propose a supervised artificial neural network approach to obtain reliable parameters estimation as demonstrated in both simulated data and real acquisition. The proposed approach is promising and can outperform, in specific conditions, other state-of-the-art fitting methods.

Rachel E Roca¹, Joshua D Herman¹, Alexandra G O'Neill¹, Sajan G Lingala², and Angel R Pineda^1
1Mathematics Department, Manhattan College, Riverdale, NY, United States, ²Roy J. Carver Department of Biomedical Engineering, University of Iowa, Iowa City, IA, United States

The changes in image quality caused by varying the parameters and architecture of neural networks are difficult to predict.  It is important to have an objective way to measure the image quality of these images.  We propose using a task-based method based on detection of a signal by human and ideal observers.  We found that choosing the number of channels and amount of dropout of a U-Net based on the simple task we considered might lead to images with artifacts which are not acceptable.  Task-based optimization may not align with artifact minimization.

Yida Wang¹, Naying He², Yan Li², Yi Duan¹, Ewart Mark Haacke^2,3, Fuhua Yan², and Guang Yang^1
1East China Normal University, Shanghai Key Laboratory of Magnetic Resonance, Shanghai, China, ²Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ³Department of Biomedical Engineering, Wayne State University, Detroit, MI, United States

We proposed a deep learning (DL) method to automatically segment brain nuclei including caudate nucleus, globus pallidus, putamen, red nucleus, and substantia nigra on Quantitative Susceptibility Mapping (QSM) data. Due to the large differences of shape and size of brain nuclei, the output branches at different semantic levels in U-net++ model were designed to simultaneously output different brain nuclei. Deep supervision was applied for improving segmentation performance. The segmentation results showed the mean Dice coefficients for the five nuclei achieved a value above 0.8 in validation dataset and the trained network could accurately segment brain nuclei regions on QSM images.

Joshua D Herman¹, Rachel E Roca¹, Alexandra G O'Neill¹, Sajan G Lingala², and Angel R Pineda^1
1Mathematics Department, Manhattan College, Riverdale, NY, United States, ²Roy J. Carver Department of Biomedical Engineering, University of Iowa, Iowa City, IA, United States

Artifacts from neural network reconstructions are difficult to characterize. It is important to assess the image quality in terms of the task for which the images will be used. In this work, we evaluated the effect of undersampling on detection of signals in images reconstructed with a neural network by both human and ideal observers. We compared these results to standard metrics (SSIM and NRMSE). Our results suggest that the undersampling level chosen by SSIM, NRMSE and ideal observer would likely be different than that of a human observer on a detection task for a small signal.

Sairam Geethanath¹, Pavan Poojar¹, Keerthi Sravan Ravi¹, and Godwin Ogbole^2
1Columbia University, New York, NY, United States, ²Department of Radiology, University College Hospital(UCH) Ibadan, Ibadan, Nigeria

The benefits of deep learning (DL) based denoising of MR images include reduced acquisition time and improved image quality at low field strength. However, simulating noisy images require biophysical models that are field and acquisition dependent. Scaling these simulations is complex and computationally intensive. In this work, we instead leverage the native noise of the data, dubbed “native noise denoising network” (NNDnet). We applied NNDnet to three different MR data types and computed the peak signal-to-noise ratio (> 38dB)  for training performance and image entropy (> 4.25) for testing performance in the absence of a reference image.

Jiaqi Dou¹, Hao Liu¹, Qiang Zhang¹, Dongye Li², Yuze Li¹, Dongxiang Xu³, and Huijun Chen^1
1Center for Biomedical Imaging Research, School of Medicine, Tsinghua University, Beijing, China, ²Department of Radiology, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China, ³Department of Radiology, University of Washington, Seattle, WA, United States

Intracranial atherosclerosis is a major cause of stroke worldwide. Vessel wall quantitative measurement is an essential tool for plaque analysis, while manual vessel wall segmentation is time-consuming and costly. In this study, we proposed a fully automated vessel wall segmentation framework for intracranial arteries using only 3D black-blood MRI, in which 3D lumen segmentation and skeletonization were applied to locate the arteries of interest for further 2D vessel wall segmentation. It achieved high segmentation performance for both normal (DICE=0.941) and stenotic (DICE=0.922) vessel wall and provided a promising tool for quantitative intracranial atherosclerosis analysis in large population studies.

Xiaofeng Liu¹, Fangxu Xing¹, Chao Yang², C.-C. Jay Kuo³, Suma Babu⁴, Georges El Fakhri¹, Thomas Jenkins⁵, and Jonghye Woo^1
1Gordon Center for Medical Imaging, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States, ²Facebook AI, Boston, MA, United States, ³Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, United States, ⁴Sean M Healey & AMG Center for ALS, Department of Neurology, Massachusetts General Hospital, Harvard Medical School, BOSTON, MA, United States, ⁵Sheffield Institute for Translational Neuroscience, University of Sheffield, Sheffield, United Kingdom

A challenge in Amyotrophic Lateral Sclerosis (ALS) research and clinical practice is to detect the disease early to ensure patients have access to therapeutic trials in a timely manner. To this end, we present a successive subspace learning model for accurate classification of ALS from T2-weighted MRI. Compared with popular CNNs, our method has modular structures with fewer parameters, so is well-suited to small dataset size and 3D data. Our approach, using 20 controls and 26 patients, achieved an accuracy of 93.48% in differentiating patients from controls, which has a potential to help aid clinicians in the decision-making process. 

Hongyu Zhou¹, Chuanli Cheng¹, Xin Liu¹, Hairong Zheng¹, and Chao Zou^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China

This work proposed a robust MR phase unwrapping method based on a deep-learning method. Through comparisons of MR images over the entire body, the model showed promising performances in both unwrapping errors and computation times. Therefore, it has promise in applications that use MR phase information.

Renyang Gu¹, Michela Antonelli¹, Pritesh Mehta ², Ashik Amlani ³, Adrian Green³, Radhouene Neji ⁴, Sebastien Ourselin¹, Isabel Dregely¹, and Vicky Goh^1
1School of Biomedical Engineering & Imaging Sciences, King's College London, London, United Kingdom, ²Biomedical Engineering and Medical Physics, University College London, London, United Kingdom, ³Radiology, Guy’s and St Thomas’ Hospitals, London, United Kingdom, ⁴Siemens Healthcare Limited, Frimley, United Kingdom

Reliable skeletal segmentation of T1-weighted Dixon MRI is a first step towards measuring marrow fat-fraction as a surrogate metric for early marrow infiltration.  We proposed an uncertainty-aware 2D U-Net (uU-Net) to reduce the impact of noisy ground-truth labels on segmentation accuracy.  Five-fold cross-validation on a dataset of 30 myeloma patients provided a mean ± SD Dice coefficient of 0.74 ± 0.03 (vs. 0.73 ± 0.04, U-Net) and 0.63 ± 0.03 (vs 0.62 ± 0.04, U-Net) for pelvic and abdominal stations, respectively. Of clinical importance, improved segmentation of the ilium and vertebrae were achieved.

Linfang Xiao^1,2, Yilong Liu^1,2, Zheyuan Yi^1,2,3, Yujiao Zhao^1,2, Peiheng Zeng^1,2, Alex T.L. Leong^1,2, and Ed X. Wu^1,2
1Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong, China, ²Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, China, ³Department of Electrical and Electronic Engineering, Southern University of Science and Technology, Shenzhen, China

Traditional MRI diagnosis consists of image reconstruction from k-space data and pathology identification in the image domain. In this study, we propose a strategy of direct pathology detection from extremely sparse MR k-space data through deep learning. This approach bypasses the traditional MR image reconstruction procedure prior to pathology diagnosis and provides an extremely rapid and potentially powerful tool for automatic pathology screening. Our results demonstrate that this new approach can detect brain tumors and classify their sizes and locations directly from single spiral k-space data with high sensitivity and specificity.

Rita Kharboush¹, Anita Karsa¹, Barbara Dymerska¹, and Karin Shmueli^1
1Medical Physics and Biomedical Engineering, University College London, London, United Kingdom

No existing phase unwrapping technique achieves completely accurate unwrapping. Therefore, we trained a convolutional neural network for phase unwrapping on (flipped and scaled) brain images from 12 healthy volunteers. An exact model of phase unwrapping was used: ground-truth (label) phase images (unwrapped with an iterative Laplacian Preconditioned Conjugate Gradient technique) were rewrapped (projected into the 2π range) to provide input images. This novel model can be used to train any neural network. Networks trained using masked (and unmasked) images showed unwrapping performance similar to state-of-the-art SEGUE phase unwrapping on test brain images and showed some generalisation to pelvic images.

Danfeng Xie¹, Yiran Li¹, and Ze Wang^1
1Department of Diagnostic Radiology and Nuclear Medicine, University of Maryland School of Medicine, Baltimore, MD, United States

This study represents the first effort to apply transfer learning of Deep learning-based ASL denoising (DLASL) method on clinical ASL data. Pre-trained with young healthy subjects’ data, DLASL method showed improved Contrast-to-Noise Ratio (CNR) and Signal-to-Noise Ratio (SNR) and higher sensitivity for detecting the AD related hypoperfusion patterns compared with the conventional method. Experimental results demonstrated the high transfer capability of DLASL for clinical studies. 

Sophie You¹, Evan M. Masutani¹, Joy Liau², Marcus T. Alley³, Shreyas S. Vasanawala³, and Albert Hsiao^2
1School of Medicine, University of California, San Diego, La Jolla, CA, United States, ²Department of Radiology, University of California, San Diego, La Jolla, CA, United States, ³Department of Radiology, Stanford University School of Medicine, Stanford, CA, United States

4D Flow MRI is valuable for the evaluation of cardiovascular disease, but abdominal applications are currently limited by the need for background phase error correction. We propose an automated deep learning-based method that utilizes a multichannel 3D convolutional neural network (CNN) to produce corrected velocity fields. Comparisons of arterial and venous flow, as well as flow before and after bifurcation of major abdominal vessels, show improved flow continuity with greater agreement after automated correction. Results of automated corrections are comparable to manual corrections. CNN-based corrections may improve reliability of flow measurements from 4D Flow MRI.

Marina Manso Jimeno^1,2, John Thomas Vaughan Jr.^1,2, and Sairam Geethanath^2
1Columbia University, New York, NY, United States, ²Columbia Magnetic Resonance Research Center (CMRRC), New York, NY, United States

fMRI acquisitions benefit from spiral trajectories; however, their use is commonly restricted due to off-resonance blurring artifacts. This work presents a deep-learning-based model for spiral deblurring in inhomogeneous fields. Training of the model utilized blurred simulated images from interleaved EPI data with various degrees of off-resonance. We investigated the effect of using the field map during training and compared correction performance with the MFI technique. Quantitative validation results demonstrated that the proposed method outperforms MFI for all inhomogeneity scenarios with SSIM>0.97, pSNR>35 dB, and HFEN<0.17. Filter visualization suggests blur learning and mitigation as expected.

Kehan Qi¹, Yu Gong^1,2, Haoyun Liang¹, Xin Liu¹, Hairong Zheng¹, and Shanshan Wang^1
1Paul C Lauterbur Research Center, Shenzhen Inst. of Advanced Technology, shenzhen, China, ²Northeastern University, Shenyang, China

Noises, artifacts, and loss of information caused by the MR reconstruction may compromise the final performance of the downstream applications such as image segmentation. In this study, we develop a re-weighted multi-task deep learning method to learn prior knowledge from the existing big dataset and then utilize them to assist simultaneous MR reconstruction and segmentation from under-sampled k-space data. It integrates the reconstruction with segmentation and produces both promising reconstructed images and accurate segmentation results. This work shows a new way for the direct image analysis from k-space data with deep learning.

Marcia Sahaya Louis^1,2, Eduardo Coello², Huijun Liao², Ajay Joshi¹, and Alexander P Lin^2
1Boston University, Boston, MA, United States, ²Brigham and Women's hospital, Boston, MA, United States

Recent years have witnessed novel applications of machine learning in radiology. Developing robust machine learning based methods for removing spectral artifacts and reconstructing the intact metabolite spectrum is an open challenge in MR spectroscopy (MRS). We had shown autoencoder models reconstruct metabolite spectrum from unsuppressed water spectrum for short TE with relatively high SNR. In this work we presents an autoencoder model with feature fusion method to extract the shallow and deep features from a water unsuppressed 1H MR spectrum. The model learns to map the extracted feature to a latent code and reconstruct the intact metabolite spectrum