Hung Do¹, Mo Kadbi¹, Dawn Berkeley¹, Brian Tymkiw¹, and Erin Kelly^1
1Canon Medical Systems USA, Inc., Tustin, CA, United States

We prospectively evaluate the generalized performance of the Deep Learning Reconstruction (DLR) method on 55 datasets acquired from 16 different anatomies. For each pulse sequence in each of the 16 anatomies, DLR and 3 predicate methods were reconstructed for randomized blinded review by 3 radiologists based on 8 scoring criteria plus a force-ranking. DLR was scored statistically higher than all 3 predicate methods in 92% of the pairwise comparisons in terms of overall image quality, clinically relevant anatomical/pathological features, and force-ranking. This work demonstrates that DLR generalizes to various anatomies and is frequently preferred over existing methods by experienced readers.

Shenjun Zhong¹, Adam Morris², Zhaolin Chen¹, and Gary Egan^1
1Monash Biomedical Imaging, Monash University, Australia, Melbourne, Australia, ²Monash eResearch Center, Monash University, Australia, Melbourne, Australia

There is a lack of pre-trained deep learning model weights on large scale medical image dataset, due to privacy concerns. Federated learning enables training deep networks while preserving privacy. This work explored co-training multi-task models on multiple heterogeneous datasets, and validated the usage of federated learning could serve the purpose of pre-trained weights for downstream tasks.

Katharina Viktoria Hoebel^1,2, Christopher P Bridge^1,3, Jay Biren Patel^1,2, Ken Chang^1,2, Marco C Pinho¹, Xiaoyue Ma⁴, Bruce R Rosen¹, Tracy T Batchelor⁵, Elizabeth R Gerstner^1,5, and Jayashree Kalpathy-Cramer^1
1Athinoula A. Martinos Center for Biomedical Imaging, Boston, MA, United States, ²Harvard-MIT Division of Health Sciences and Technology, Cambridge, MA, United States, ³MGH and BWH Center for Clinical Data Science, Boston, MA, United States, ⁴Department of Magnetic Resonance, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, China, ⁵Stephen E. and Catherine Pappas Center for Neuro-Oncology, Massachusetts General Hospital, Boston, MA, United States

The outlines of target structures on medical imaging can be highly ambiguous. The uncertainty about the “true” outline is evident in high inter-rater variability of manual segmentations. So far, no method is available to identify cases likely to exhibit a high inter-rater variability. Here, we demonstrate that ground truth independent uncertainty metrics extracted from a MC dropout segmentation model developed on labels of only one rater correlate with inter-rater variability. This relationship can be used to identify ambiguous cases and flag them for more detailed review supporting consistent and reliable patient evaluation in research and clinical settings.

Rutwik Shah¹, Bruno Astuto Arouche Nunes¹, Tyler Gleason¹, Justin Banaga¹, Kevin Sweetwood¹, Allen Ye¹, Will Fletcher¹, Rina Patel¹, Kevin McGill¹, Thomas Link¹, Valentina Pedoia¹, Sharmila Majumdar¹, and Jason Crane^1
1Department of Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States

Radiologists play a central role in image annotation for training Machine Learning models. Key challenges in this regard include low inter-reader agreement for challenging cases and concerns of interpersonal bias amongst trainers. Inspired by biological swarm intelligence, we explored the use of real time consensus labeling by three sub-specialty (MSK) trained radiologists and five radiology residents in improving training data. A second swarm session with three residents was conducted to explore the effect of swarm size. These results were validated against clinical ground truth and also compared with results from a state-of-the-art AI model tested on the same dataset. 

Suheyla Cetin-Karayumak¹, Evdokiya Knyazhanskaya², Brynn Vessey², Sylvain Bouix¹, Benjamin Wade³, David Tate⁴, Paul Sherman⁵, and Yogesh Rathi^1
1Brigham and Women's Hospital and Harvard Medical School, Boston, MA, United States, ²Brigham and Women's Hospital, Boston, MA, United States, ³Ahmanson-Lovelace Brain Mapping Center, UCLA, Los Angeles, CA, United States, ⁴University of Utah, Salt Lake City, UT, United States, ⁵U.S. Air Force School of Aerospace Medicine, San Antonio, TX, United States

This study aims to tackle the structural MRI (T1) data harmonization problem by presenting a novel multi-site T1 data harmonization, which uses the CycleGAN network with segmentation loss (CycleGANs). CycleGANs aims to learn an efficient mapping of T1 data across scanners from the same set of subjects while simultaneously learning the mapping of free surfer parcellations. We demonstrated the efficacy of the method with the Dice overlap scores between FreeSurfer parcellations across two datasets before and after harmonization.

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 generated diagnostic quality amyloid positron emission tomography (PET) images with deep learning enhancement of actual ultra-low-dose (~2% of the original) PET images and simultaneously acquired structural magnetic resonance imaging (MRI) inputs. Here, we will investigate whether simultaneity is a requirement for such structural MRI inputs. If simultaneity is not required, this will increase the utility of MRI-assisted ultra-low-dose PET imaging by including data acquired on separate PET/ computed tomography (CT) and standalone MRI machines.

Mario Serrano-Sosa¹, Jared Van Snellenberg², Jiayan Meng², Jacob Luceno², Karl Spuhler³, Jodi Weinstein², Anissa Abi-Dargham², Mark Slifstein², and Chuan Huang^2,4
1Biomedical Engineering, Stony Brook University, Stony Brook, NY, United States, ²Psychiatry, Renaissance School of Medicine at Stony Brook University, Stony Brook, NY, United States, ³Radiation Oncology, NYU Langone, New York, NY, United States, ⁴Radiology, Renaissance School of Medicine at Stony Brook University, Stony Brook, NY, United States

Segmenting striatal subregions can be difficult; wherein atlas-based approaches have been shown to be less reliable in patient populations and have problems segmenting smaller striatal ROI’s. We developed a Multi-Task Learning model to segment multiple 3D striatal subregions using a Convolutional Neural Network and compared it to the Clinical Imaging Center atlas (CIC). Dice Score Coefficient and multi-modal objective assessment (PET and fMRI) were conducted to evaluate the reliability of MTL-generated segmentations compared to atlas-based. Overall, MTL-generated segmentations were more comparable to manual than CIC across all ROI’s and analyses. Thus, we show MTL method provides reliable striatal subregion segmentations.

Jiahong Ouyang¹, kevin Chen², and Greg Zaharchuk^2
1Electrical Engineering, Stanford University, Stanford, CA, United States, ²Department of Radiology, Stanford University, Stanford, CA, United States

PET is a widely used imaging technique but it requires exposing subjects to radiation and is not offered in the majority of medical centers in the world. Here, we proposed to synthesize FDG-PET images from multi-contrast MR images by a U-Net based network with symmetry-aware spatial-wise attention, channel-wise attention, split-input modules, and random dropout training strategy. The experiments on a brain tumor dataset of 70 patients demonstrated that the proposed method was able to generate high-quality PET from MR images without the need for radiotracer injection.  We also demonstrate methods to handle potential missing or corrupted sequences.

Bendik Skarre Abrahamsen¹, Tone Frost Bathen^1,2, Live Eikenes¹, and Mattijs Elschot^1,2
1Department of Circulation and Medical Imaging, NTNU, Trondheim, Norway, ²Department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway

Attenuation correction is a challenge in PET/MRI. In this study we propose a novel attenuation correction method based on estimating the bias image between PET reconstructed using a 4-class attenuation correction map and PET reconstructed with an attenuation correction map where bone information is added from a co-registered CT image. A generative adversarial network was trained to estimate the bias between the PET images. The proposed method  has comparable performance to other Deep Learning based attenuation correction methods where no additional MRI sequences are acquired. Bias estimation thus constitutes a viable alternative to pseudo-CT generation for PET/MR attenuation correction.

Wenli Li¹, Ziyu Meng¹, Ruihao Liu¹, 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

Multimodal brain imaging acquires complementary information of the brain. However, due to the high dimensionality of the data, it is challenging to capture the underlying joint spatial and cross-modal dependence required for statistical inference in various brain image processing tasks. In this work, we proposed a new multimodal image fusion method that synergistically integrates tensor modeling and deep learning. The tensor model was used to capture the joint spatial-intensity-modality dependence and deep learning was used to fuse spatial-intensity-modality information. Our method has been applied to multimodal brain image segmentation, producing significantly improved results.

Tomoki Miyasaka¹, Satoshi Funayama², Daiki Tamada², Utaroh Motosugi³, Hiroyuki Morisaka², Hiroshi Onishi², and Yasuhiko Terada^1
1Institute of Applied Physics, University of Tsukuba, Tsukuba, Japan, ²Department of Radiology, University of Yamanashi, Chuo, Japan, ³Department of Radiology, Kofu-Kyoritsu Hospital, Kofu, Japan

The use of deep learning (DL) for compressed sensing (CS) have recently received increased attention. Generally, DL-CS uses single-contrast CS reconstruction (SCCS) where the single-contrast image is used as the network input. However, in clinical routine examinations, different contrast images are acquired in the same session, and CS reconstruction using multi-contrast images as the input (MCCS) has the potential to show better performance. Here, we applied DL-MCCS to brain MRI images acquired during routine examinations. We trained data-driven and model-based networks, and showed that for both cases, MCCS outperformed SCCS.

Yu Zhang¹, Weijian Huang¹, Fei Li¹, Qiang He², Haoyun Liang¹, Xin Liu¹, Hairong Zheng¹, and Shanshan Wang^1
1Paul C Lauterbur Research Center, Shenzhen Inst. of Advanced Technology, shenzhen, China, ²United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China

Multi-modal magnetic resonance (MR) image registration is essential in the clinic to achieve accurate imaging-based disease diagnosis and treatment planning. Although the existing registration methods have achieved good performance and attracted widespread attention, the image details may be lost after registration. In this study, we propose a multi-modal MR image registration with topology-preserving dual consistency constraint, which achieves the best registration performance with a Dice score of 0.813 in identifying stroke lesions.

Shihan Qiu^1,2, Yuhua Chen^1,2, Sen Ma¹, Zhaoyang Fan^1,2, Anthony G. Christodoulou^1,2, Yibin Xie¹, and Debiao Li^1,2
1Biomedical Imaging Research Institute, Cedars-Sinai Medical Center, Los Angeles, CA, United States, ²Department of Bioengineering, UCLA, Los Angeles, CA, United States

MR Multitasking is an efficient approach for quantification of multiple parametric maps in a single scan. The Bloch equations can be used to derive conventional contrast-weighted images, which are still preferred by clinicians for diagnosis, from quantitative maps. However, due to imperfect modeling and acquisition, these synthetic images often exhibit artifacts. In this study, we developed a deep learning-based method to synthesize contrast-weighted images from Multitasking spatial factors without explicit Bloch modeling. We demonstrated that our method provided synthetic images with higher quality and fidelity than the model-based approach or a similar deep learning method using quantitative maps as input.

Henitsoa RASOANANDRIANINA¹, Anais BERNARD¹, Guillaume GAUTIER¹, Julien ROUYER¹, Yves HAXAIRE², Christophe AVARE³, and Lucile BRUN^1
1Department of Research & Innovation, Olea Medical, La Ciotat, France, ²Clinical Program Department, Olea Medical, La Ciotat, France, ³Avicenna.ai, La Ciotat, France

In this study, we present an automatic, multi-regional and multi-sequence deep-learning-based algorithm for background segmentation on both DSC and DCE images which consisted of a 2D U-net trained with a large multi-centric and multi-vendor database including DSC brain, DCE brain, DCE breast, DCE abdomen and DCE pelvis data. Cross-validation-based training results showed an overall good performance of the proposed algorithm with a median Dice score of 0.974 in test set and 0.979 over all datasets, and a median inference duration of 0.15s per volume on GPU. This is the first reported deep-learning-based multi-sequence and multi-regional background segmentation on MRI data.

Christopher Man^1,2, Linfang Xiao^1,2, Yilong Liu^1,2, Vick Lau^1,2, Zheyuan Yi^1,2,3, 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 SAR, China, ²Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR, China, ³Department of Electrical and Electronic Engineering, Southern University of Science and Technology, Shenzhen, China

This study presents a deep learning based reconstruction for multi-contrast MR data with orthogonal undersampling directions across different contrasts. It enables exploiting the rich structural similarities from multiple contrasts as well as the incoherency arose from complementary sampling. The results show that the proposed method can achieve robust reconstruction for single-channel multi-contrast MR data at R=4.

Valentin Demeusy¹, Florent Roche¹, Fabrice Vincent¹, Jean-Pierre Guichard², Jessica Lebenberg^3,4, Eric Jouvent^3,5, and Hugues Chabriat^3,5
1Imaging Core Lab, Medpace, Lyon, France, ²Department of Neuroradiology, Hôpital Lariboisière, APHP, Paris, France, ³FHU NeuroVasc, INSERM U1141, Paris, France, ⁴Université de Paris, Paris, France, ⁵Departement of Neurology, Hôpital Lariboisière, APHP, Paris, France

We propose a novel automatic WMH segmentation method based on a convolutional neural network to study the longitudinal WMH changes among a cohort of 101 CADASIL patients. We demonstrate that this method is able to produce consistent quantitative measures of WMH volume by the strong correlation between the computed baseline WMH volume and the clinically assessed Fazekas score. Our main results show that the progression of WMH is correlated to the baseline volume and that this progression largely vary at individual level although a rapid extension is mainly detected between 40 and 60 years in the whole population.

Daniel ZL Kor¹, Saad Jbabdi¹, Jeroen Mollink¹, Istvan N Huszar¹, Menuka Pallebage- Gamarallage², Adele Smart², Connor Scott², Olaf Ansorge², Amy FD Howard¹, and Karla L Miller^1
1Wellcome Centre for Integrative Neuroimaging, University of Oxford, Oxford, United Kingdom, ²Neuropathology, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom

Immunohistochemistry (IHC) images are often used as a microscopic validation tool for MRI. Acquisition of MRI and IHC in the same ex-vivo tissue sample can enable direct correlation between MRI measures and purported sources of image contrast derived from IHC, ideally at the voxel level. However, most IHC analyses still involve manual intervention (e.g. setting of thresholds). Here, we describe an end-to-end pipeline for automatically extracting stained area fraction maps to quantify the IHC stain for a given microstructural feature. The pipeline has improved reproducibility and robustness to histology artefacts, compared to manual MRI-histology analyses that suffer from inter-operator bias.

Qing Xu¹, Weiqiang Dou², and Jing Ye^1
1Northern Jiangsu People's Hospital, Yangzhou, China, ²GE Healthcare, MR Research China, Beijing, China

The aim of this study was to evaluate the feasibility of magnetic resonance imaging (MRI) based deep learning (DL) model in differentiating benign from malignant renal tumors. The performance of the applied DL model was further compared with that from a random forest radiomics model. More robust performance was achieved using MRI based DL model than the radiomics model (AUC = 0.925 vs 0.854, p<0.05). Therefore, the applied MRI based deep transfer learning model might be considered a convenient and reliable approach for differentiating benign from malignant renal tumors in clinic.

Morteza Esmaeili¹, Vegard Antun², Riyas Vettukattil³, Hassan Banitalebi¹, Nina Krogh¹, and Jonn Terje Geitung^1,3
1Akershus University Hospital, Lørenskog, Norway, ²Department of Mathematics, University of Oslo, Oslo, Norway, ³Faculty of Medicine, Institute of Clinical Medicine, University of Oslo, Oslo, Norway

Machine learning approaches provide convenient autonomous object classification in medical imaging domains. This study uses an explainable method to evaluate the high-level features of deep learning methods in tumor localization.

Soumick Chatterjee^1,2,3, Himanshi Bajaj³, Suraj Bangalore Shashidhar³, Sanjeeth Busnur Indushekar³, Steve Simon³, Istiyak Hossain Siddiquee³, Nandish Bandi Subbarayappa³, 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

Deep learning algorithms have been used extensively in tackling medical image registration issues. However, these methods have not thoroughly evaluated on datasets representing real clinic scenarios. Hence in this survey, three state-of-the-art methods were compared against the gold standards ANTs and FSL, for performing deformable image registrations on publicly available IXI dataset, which resembles clinical data. The comparisons were performed for intermodality and intramodality registration tasks; though in all the respective papers, only the intermodality registrations were exhibited. The experiments have shown that for intramodality tasks, all the methods performed reasonably well and for intermodality tasks the methods faced difficulties.