Lavanya Umapathy^1,2, Zhiyang Fu^1,2, Rohit Philip², Diego Martin³, Maria Altbach^2,4, and Ali Bilgin^1,2,4,5
1Department of Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ²Department of Medical Imaging, University of Arizona, Tucson, AZ, United States, ³Department of Radiology, Houston Methodist Hospital, Houston, TX, United States, ⁴Department of Biomedical Engineering, University of Arizona, Tucson, AZ, United States, ⁵Program in Applied Mathematics, University of Arizona, Tucson, AZ, United States

The availability of large unlabeled datasets compared to labeled ones motivate the use of self-supervised pretraining to initialize deep learning models for subsequent segmentation tasks. We consider two pre-training approaches for driving a CNN to learn different representations using: a) a reconstruction loss that exploits spatial dependencies and b) a contrastive loss that exploits semantic similarity. The techniques are evaluated in two MR segmentation applications: a) liver and b) prostate segmentation in T2-weighted images. We observed that CNNs pretrained using self-supervision can be finetuned for comparable performance with fewer labeled datasets.

Sara L Saunders^1,2, Mitchell J Gross², Gregory J Metzger¹, and Patrick J Bolan^1
1Center for MR Research / Radiology, University of Minnesota, MINNEAPOLIS, MN, United States, ²Biomedical Engineering, University of Minnesota, Minneapolis, MN, United States

This work compares T2 maps calculated from multi-echo spin-echo MR images using a conventional non-linear least squares (NLLS) fitting method to those constructed with a U-Net, a type of convolutional neural network. The performance of the U-Net and NLLS methods was compared in two retrospectively simulated experiments with a) reduced echo train lengths and b) decreased SNR to emulate accelerated acquisitions. The U-Net generally gave higher accuracy than NLLS fitting, with the trade-off of a modest increase of blurring of the resultant T2 maps.

Ye Tian¹, Junhao Zhang¹, Xuemin Zhu¹, Pin-Yu Lee¹, Vishwanatha Mitnala Rao¹, Zhuoyao Xin¹, Andrew F Laine², Scott A Small³, and Jia Guo^4
1Columbia University, New York, NY, United States, ²Biomedical Engineering, Columbia University, New York, NY, United States, ³Neurology, Columbia University, New York, NY, United States, ⁴Psychiatry, Columbia University, New York, NY, United States

To detect distinctive structural abnormalities of schizophrenia early in magnetic resonance imaging (MRI) data, we propose a 3D Medical Image Global-Regional (3D-MIGR) Transformer with a VGG11BN backbone followed by a global-regional transformer encoder, which outperforms state-of-the-art models. We trained and tested our model on 887 pre-processed structural whole-head (WH) T1W 3D images from 3 datasets with similar acquisition parameters. 3D-MIGR Transformer improves AUROC to 0.990 and demonstrates strong generality. We found that combining volume-level contextual information with patch-level features enhances performance and allows us to identify ventricle areas as the most informative regions in regional schizophrenia likelihood visualization.

Nicholas John Sisco¹, Francesca Bagnato², Ashley M Stokes¹, and Richard D Dortch^1
1Division of Neuroimaging Researc, Barrow Neurological Institute, Phoenix, AZ, United States, ²Neurology Department, Vanderbilt University Medical Center, Nashville, TN, United States

Using a simple linear model, we predicted myelin-specific PSR maps from clinically available images. We trained this linear model on various combinations of routine images, including gray-matter standardized T₁- and T₂-weighted images and diffusion maps. The model reasonably predicted PSR values from structural images alone, although model performance was improved by the addition of diffusion parameters. In the future, this model may enable researchers and clinicians to assess myelin status without the need for specialized myelin imaging sequences. In addition, myelin maps could be obtained retrospectively in large imaging databases without myelin specific methods.

Yanchen Guo¹, Poorya Chavoshnejad², Mir Jalil Razavi², and Weiying Dai^1
1Computer Science, State University of New York at Binghamton, Binghamton, NY, United States, ²Mechanical Engineering, State University of New York at Binghamton, Binghamton, NY, United States

Finite Element (FE)-based mechanical models can simulate the brain growth and folding process. But they are time consuming due to the large number of nodes in a real human brain and the reverse process to the initial smooth brain surfaces is difficult because it is not invertible problem. Here, we demonstrate a proof-of-concept that deep-learning neural networks (DNN) can learn the growth and folding process of human brain in forward and reverse directions and can predict/retrieve the developed/primary folding patterns in a very fast speed.

Xinzeng Wang¹, Ali Ersoz², Daniel Litwiller³, Jingfei Ma⁴, Jason Stafford⁴, and Ersin Bayram^1
1GE Healthcare, Houston, TX, United States, ²GE Healthcare, Waukesha, WI, United States, ³GE Healthcare, Denver, CO, United States, ⁴MD Anderson Cancer Center, Houston, TX, United States

Compared to DW-EPI, DW PROPELLER is less sensitive to susceptibility, chemical shift, and motion, and thus shows better image quality in areas, such as skull base, head-neck and pelvis. However, the SNR and in-plane resolution of DW PROPELLER images are often inferior to DW-EPI and often requires a longer scan time to compensate for this. In this work, we evaluated a deep-learning reconstruction method to improve the SNR of in-plane resolution of DW PROPELLER images without increasing acquisition time.

Daniel V Litwiller¹, Xinzeng Wang², R Marc Lebel³, Baolian Yang⁴, Jeffrey McGovern⁴, Brian Burns⁵, and Suchandrima Banerjee^6
1GE Healthcare, Denver, CO, United States, ²GE Healthcare, Houston, TX, United States, ³GE Healthcare, Calgary, AB, Canada, ⁴GE Healthcare, Waukesha, WI, United States, ⁵GE Healthcare, Olympia, WA, United States, ⁶GE Healthcare, Menlo Park, CA, United States

The high sensitivity of MRI at 7T enables brain imaging with unprecedented spatial resolution, which can be important to the assessment of a variety of neurological disorders, such as multiple sclerosis, epilepsy, and neurodegenerative disease. With sub-millimeter voxel dimensions, and prolonged acquisition times, however, sensitivity to motion and pulsatility is increased dramatically. This increased sensitivity to motion can be managed with techniques like PROPELLER. Here, we present an initial assessment of a deep learning-based image reconstruction for high-resolution, 7T PROPELLER, and evaluate its ability to improve signal-to-noise ratio, and anatomical conspicuity, without increasing scan time.

Weijie Chen¹, Seyed Iman Zare Estakhraji², and Alan B McMillan^3
1Department of Electrical and Computer Engineering, University of Wisconsin-Madison, Madison, WI, United States, ²Department of Radiology, University of Wisconsin-Madison, Madison, WI, United States, ³Department of Biomedical Engineering, Department of Radiology and Medical Physics, Madison, WI, United States

Applications such as PET/MR and MR-only Radiotherapy Planning need the capability to derive a CT-like image from MRI inputs to enable accurate attenuation correction and dose estimation. More recently, transformer models have been proposed for computer vision applications. Models in the transformer family discard traditional convolution-based network structures and emphasize the importance of non-local information yielding potentially more realistic outputs. To evaluate the performance of SwinIR, TransUet, and Unet. After comparing results visually and quantitatively, the SwinIR models and TransUnet models appear to provide higher-quality synthetic CT scans compared to the conventional Unet.  

Eze Ahanonu¹, Zhiyang Fu^1,2, Kevin Johnson², Rohit Philip¹, Diego R Martin³, Maria Altbach^2,4, and Ali Bilgin^1,2,4
1Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ²Medical Imaging, University of Arizona, Tucson, AZ, United States, ³Radiology, Houston Methodist Hospital, Houston, TX, United States, ⁴Biomedical Engineering, University of Arizona, Tucson, AZ, United States

This study presents a deep learning based technique for slice super resolution in RADTSE T2 mapping. The proposed method may be used to accelerate full liver imaging, while maintaining sufficient through-plane resolution for detection of small pathologies.

Juan Liu^1
1Yale University, NEW HAVEN, CT, United States

Recently, deep learning approaches have been proposed for QSM processing - background field removal, field-to-source inversion, and single-step QSM reconstruction. In these tasks, the networks usually take local fields or total fields as inputs, which have valid voxels within volume of interests (VOIs) and invalid voxels outside of VOIs. CNNs fail to consider this spatial information when using spatial invariant filters and conventional padding mechanism, which could introduce spatial artifacts in the QSM results. Here, we propose an improved padding technique utilizing neighboring valid voxels of invalid voxels to estimate the invalid voxels in feature maps of CNNs.

Xuemin Zhu¹, Pin-Yu Lee², Ye Tian², Andrew F. Laine³, Tal Nuriel⁴, and Jia Guo^5
1Zuckerman Institute, New York, NY, United States, ²Columbia University, New York, NY, United States, ³Biomedical Engineering, Columbia University, New York, NY, United States, ⁴Taub Institute, New York, NY, United States, ⁵Psychiatry, Columbia University, New York, NY, United States

Recent evidence suggests APOE4 is the most potent genetic risk factor for late-onset Alzheimer’s disease. However, researchers have not treated its impact on brain aging. We proposed a 3D Medical Image Global Regional (MIGR) Transformer to determine whether APOE4 affects brain age gap estimate using MRI. We achieved a mean absolute error of 4.71 on brain age estimate in a large healthy dataset (n = 2852) with an age range of 18-97 years from 13 public datasets. We investigated the brainAGE on three  stages of Alzheimer’s disease. We found an accelerating trend of brain aging for APOE4 carriers.

Mohamad Abdi¹, Yu Zhao¹, Sepehr Farhand¹, Ke Zeng¹, Mahesh Ranganath¹, Yoshihisa Shinagawa¹, and Gerardo Valadez Hermosillo^1
1Siemens Healthineers, Malvern, PA, United States

MRI requires careful design of imaging protocols and parameters to optimally assess a particular region of the body and/or pathological process. Selection of acquisition parameters is a challenging task because (a) the relationship between the acquisition parameters and the image features is typically non-trivial, and (b) not all users have the leverage to optimize their imaging protocols. To help users overcome these challenges and elevate the user experience, a deep metric learning tool was developed as a recommendation system for automatic candidate generation of imaging protocols. The feasibility of the model is evaluated using 3-dimensional brain MR images.

Yang Wen^1,2,3, Chuan Zhou², Leiting Chen², Yu Deng⁴, Martine Cleusix⁵, Raoul Jenni⁵, Philippe Conus⁶, Kim Q. Do⁵, and Lijing Xin^1,3
1Animal imaging and technology core (AIT), Center for Biomedical Imaging (CIBM), Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, ²University of Electronic Science and Technology of China, Chengdu, China, ³Laboratory of Functional and Metabolic Imaging, Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, ⁴Department of Biomedical Engineering, King's College London, London, United Kingdom, ⁵Center for Psychiatric Neuroscience, Department of Psychiatry, Centre Hospitalier Universitaire Vaudois (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland, ⁶Service of General Psychiatry, Department of Psychiatry, Centre Hospitalier Universitaire Vaudois (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland

The aims of this study were: (1) to test the feasibility of using a deep learning model with 7T sMRI as an input to predict cognition levels (CLs) at the single-subject level, and (2) to investigate whether the inclusion of CLs estimation could facilitate the classification for early psychosis (EP) patients and healthy controls (HCs). Promising accuracy was achieved in estimating CLs and the inclusion provides considerable classification improvement. Fivefold cross-validating experiments demonstrated higher classification AUC-ROC scores over published methods. Therefore, deep learning can be used to estimate CLs and CL estimation improves the classification performance of EP.

Badhan Kumar Das^1,2, Lorenz A. Kapsner¹, Sabine Ohlmeyer¹, Frederik B. Laun¹, Andreas Maier², Michael Uder¹, Evelyn Wenkel¹, Sebastian Bickelhaupt¹, and Andrzej Liebert^1
1Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany, ²Pattern Recognition Lab, Department of Computer Science, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany

The purpose of this work was to automatically classify BPE using T1-weighted subtraction volumes and diffusion-weighted imaging volumes in breast MRI. The dataset consisted of 621 routine breast MRI examination acquired at University Hospital Erlangen. 2D MIP and 3D T1-subtraction volumes were used for the automatic detection of BPE classes. Multi-b-value DWI (up to1500s/mm2) DWI images were used for automatic prediction. ResNet and DenseNet models were used for 2D and 3D data respectively. The study demonstrated an AUROC of 0.8107 on the test set using the T1-subtraction volumes. With DWI volumes, a slightly decreased AuROC of 0.78 was achieved.

Rafael Navarro-González¹, Elisa Moya-Sáez¹, Rodrigo de Luis-García¹, Santiago Aja-Fernández¹, and Carlos Alberola-López^1
1University of Valladolid, Valladolid, Spain

Radiomics systems for survival prediction in glioblastoma multiforme could enhance patient management, personalizing its treatment and obtaining better outcomes. However, these systems are data-demanding multimodality images. Thus, synthetic MRI could improve radiomics systems by retrospectively completing databases or replacing artifacted images. In this work we analyze the replacement of an acquired modality by a synthesized counterversion for predicting survival with an independent radiomic system. Results prove that a model fed with the synthesized modality achieves similar performance compared to using the acquired modality, and better performance than using a corrupted modality or a model trained from scratch without this modality.