Alfredo De Goyeneche¹, Shuyu Tang¹, Nii Okai Addy¹, Bob Hu¹, William Overall¹, and Juan Santos^1
1HeartVista, Inc., Los Altos, CA, United States

We developed a deep-learning-based approach for motion correction in quantitative cardiac MRI, including perfusion, T1 mapping, and T2 mapping. The proposed approach consists of a segmentation network and a registration network. The segmentation network was trained using 2D short-axis images for each of the three sequences, while the same registration network was shared between all three sequences. The proposed approach was faster and more accurate than a popular traditional registration method. Our work is beneficial for building a faster and more robust automated processing pipeline to obtain CMR parametric maps.

Junshen Xu¹ and Elfar Adalsteinsson^1
1Massachusetts Institute of Technology, Cambridge, MA, United States

Image denoising is of great importance for medical imaging system, since it can improve image quality for disease diagnosis and further image processing. In cardiac MRI, images are acquired at different time frames to capture the cardiac dynamic. The correlation among different time frames makes it possible to improve denoising results with information from other time frames. In this work, we propose a self-supervised deep learning framework for cardiac MRI denoising. Evaluation on in vivo data with different noise statistics shows that our method has comparable or even better performance than other state-of-the-art unsupervised or self-supervised denoising methods.

Jiarui Xing¹, Sona Ghadimi², Mohammad Abdishektaei², Kenneth C Bilchick³, Frederick H Epstein², and Miaomiao Zhang^1
1Electrical and Computer Engineering, University of Virginia, Charlottesville, VA, United States, ²Department of Biomedical Engineering, University of VIrginia, Charlottesville, VA, United States, ³School of Medicine, University of Virginia, Charlottesville, VA, United States

Implantation of the left ventricular pacing lead at the area with delayed activation is critical to Cardiac Resynchronization Therapy (CRT) response. Current approaches of detecting late-activated regions of left ventricles (LV) are slow with unsatisfied accuracy, particularly in cases where scar tissues exist in the patient’s heart. This work presents a multi-task deep learning algorithm to automatically identify late-activated regions of LV, as well as estimating the Time to the Onset of circumferential Shortening (TOS) using spatio-temporal cardiac DENSE MR images. Experimental results show that our algorithm provides ultra-fast identification of late-activated regions and estimated TOS with increased accuracy.  

Ziyu Li¹, Qiyuan Tian², Chanon Ngamsombat^2,3, Samuel Cartmell⁴, John Conklin^2,4, Augusto Lio M. Gonçalves Filho^2,4, Wei-Ching Lo⁵, Guangzhi Wang¹, Kui Ying⁶, Kawin Setsompop^7,8, Qiuyun Fan², Berkin Bilgic², Stephen Cauley², and Susie Y Huang^2,4
1Department of Biomedical Engineering, Tsinghua University, Beijing, China, ²Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, United States, ³Department of Radiology, Siriraj Hospital, Mahidol University, Bangkok, Thailand, ⁴Department of Radiology, Massachusetts General Hospital, Boston, MA, United States, ⁵Siemens Medical Solutions, Boston, MA, United States, ⁶Department of Engineering Physics, Tsinghua University, Beijing, China, ⁷Department of Radiology, Stanford University, Stanford, CA, United States, ⁸Department of Electrical Engineering, Stanford University, Stanford, CA, United States

Highly accelerated high-resolution volumetric brain MRI is intrinsically noisy. A hybrid generative adversarial network (GAN) for denoising (entitled HDnGAN) consisting of a 3D generator and a 2D discriminator was proposed to improve the SNR of highly accelerated images while preserving realistic textures. The novel architecture benefits from improved image synthesis performance and increased training samples for training the discriminator. HDnGAN's efficacy is demonstrated on 3D standard and Wave-CAIPI T₂-weighted FLAIR data acquired in 33 multiple sclerosis patients. Generated images are similar to standard FLAIR images and superior to Wave-CAIPI and BM4D-denoised images in quantitative evaluation and assessment by neuroradiologists.

Suphachart Leewiwatwong¹, Junlan Lu², David Mummy³, Isabelle Dummer^3,4, Kevin Yarnall⁵, Ziyi Wang¹, and Bastiaan Driehuys^1,2,3
1Biomedical Engineering, Duke University, Durham, NC, United States, ²Medical Physics, Duke University, Durham, NC, United States, ³Radiology, Duke University, Durham, NC, United States, ⁴Bioengineering, McGill University, Montréal, QC, Canada, ⁵Mechanical Engineering and Materials Science, Duke University, Durham, NC, United States

Quantifying hyperpolarized ¹²⁹Xe MRI of pulmonary ventilation and gas exchange requires accurate segmentation of the thoracic cavity. This is typically done either manually or semi-automatically using an additional proton scan volume-matched to the gas image. These methods are prone to operator subjectivity, image artifacts, alignment/registration issues, and SNR. Here we demonstrate using a 3D convolutional neural network (CNN) to automatically and directly delineate the thoracic cavity from ¹²⁹Xe MRI alone. This 3D-CNN uses a combination of Dice-Focal, perceptual loss, and training with template-based data augmentation to demonstrate thoracic cavity segmentation with a Dice score of 0.955 vs. expert readers.

Malte Hoffmann^1,2, Benjamin Billot³, Juan Eugenio Iglesias^1,2,3,4, Bruce Fischl^1,2,4, and Adrian V Dalca^1,2,4
1Department of Radiology, Harvard Medical School, Boston, MA, United States, ²Department of Radiology, Massachusetts General Hospital, Boston, MA, United States, ³Centre for Medical Image Computing, University College London, London, United Kingdom, ⁴Computer Science and Artificial Intelligence Laboratory, MIT, Cambridge, MA, United States

We introduce a novel strategy for learning deformable registration without acquired imaging data, producing networks robust to MRI contrast. While classical methods repeat an optimization for every new image pair, learning-based methods require retraining for accurate registration of unseen image types. To address these inefficiencies, we leverage a generative strategy for diverse synthetic label maps and images that enable training powerful networks that generalize to a broad spectrum of MRI contrasts. We demonstrate robust and accurate registration of arbitrary unseen MRI contrasts with a single network, thereby eliminating the need for retraining models.

Aniket Pramanik¹ and Mathews Jacob^1
1Electrical and Computer Engineering, The University of Iowa, Iowa City, IA, United States

We present a novel framework for the joint reconstruction and segmentation of parallel MRI (PMRI) brain data. We introduce an image domain deep network for calibrationless recovery of undersampled PMRI data. It is deep-learning based generalization of local low-rank approaches for uncalibrated PMRI recovery including CLEAR. The image domain approach exploits additional annihilation relations compared to k-space based approaches and hence offers improved performance. To minimize segmentation errors resulting from undersampling artifacts, we combined the proposed scheme with a segmentation network and trained it end-to-end. It offers improved reconstruction with reduced blurring and sharper edges than independently trained reconstruction network. 

Annika Liebgott^1,2, Louisa Fay¹, Viet Chau Vu², Bin Yang¹, and Sergios Gatidis^2
1Institute of Signal Processing and System Theory, University of Stuttgart, Stuttgart, Germany, ²Department of Radiology, University Hospital of Tuebingen, Tuebingen, Germany

The treatment of malignant melanoma with immunotherapy is a promising approach to treat advanced stages of the disease. However, the treatment can cause serious side effects and not every patient responds to it, which means crucial time may be wasted by an ineffective treatment. Assessment of the possible therapy response is hence an important research issue. The research presented in this study focuses on the investigation of the potential of medical imaging and machine learning to solve this task. To this end, we trained and compared different deep learning models on multi-modal PET/MR images to differentiate non-responsive from responsive patients.

Jonas Denck^1,2,3, Jens Guehring³, Andreas Maier¹, and Eva Rothgang^2
1Pattern Recognition Lab, Department of Computer Science, Friedrich-Alexander University Erlangen-Nürnberg, Erlangen, Germany, ²Department of Industrial Engineering and Health, Technical University of Applied Sciences Amberg-Weiden, Weiden, Germany, ³Siemens Healthcare, Erlangen, Germany

We trained an image-to-image GAN that incorporates the sequence parameterizations in terms of the acquisition parameters repetition time and echo time into the image synthesis. We trained our model on the generation of synthetic fat-saturated MR knee images from non-fat-saturated MR knee images conditioned on the acquisition parameters, enabling us to synthesize MR images with varying image contrast. Our approach yields an NMSE of 0.11 and PSNR of 23.64, and surpasses the performance of a pix2pix [1] benchmark method. It can potentially be used to synthesize missing/additional MR contrasts and for customized data generation to support AI training.

Zongpai Zhang¹, Huiyuan Yang¹, Yanchen Guo¹, Lijun Yin¹, David C. Alsop², and Weiying Dai^1
1State University of New York at Binghamton, Binghamton, NY, United States, ²Beth Israel Deaconess Medical Center & Harvard Medical School, Boston, MA, United States

Convolutional neural network (CNN) has demonstrated its accuracy and speed in image registration of structural MRI. We designed an affine registration network (ARN), based on CNN, to explore its feasibility on image registration of perfusion fMRI. The six affine parameters were learned from the ARN using both simulated and real perfusion fMRI data and the transformed images were generated by applying the transformation matrix derived from the affine parameters. The results demonstrated that our ARN markedly outperforms the iteration-based SPM algorithm both in simulated and real data. The current ARN is being extended for deformable fMRI image registration.