ISMRM25 - AI-Based Image Reconstruction & Enhancement

ISMRM & ISMRT Annual Meeting & Exhibition • 10-15 May 2025 • Honolulu, Hawai'i

AI-Based Image Reconstruction & Enhancement

Power Pitch

AI & Machine Learning

Thursday, 15 May 2025

Power Pitch Theatre 1

08:15 - 10:15

Moderators: Cemre Ariyurek & Yusuke Jo

Session Number: PP-03

No CME/CE Credit

08:15		Screen Number: 1 1184. A quest for deep learning MR image reconstruction loss functions in k-space View Presentation Video S. Xu, W. Huang, K. Hammernik, M. Terpstra, D. Rueckert, S. Gatidis, T. Kuestner Medical Image and Data Analysis (MIDAS.lab), Department of Diagnostic and Interventional Radiology, University of Tuebingen, Tübingen, Germany Impact: We propose a loss function that is more suitable for k-space, taking into account its characteristics and enhancing the accuracy of k-space regression. This loss function can be applied to any tasks that require calibration of k-space similarity.
08:17		Screen Number: 2 1185. Resolving ambiguous space: Leveraging side information with deep learning to extend the limits of MR image reconstruction View Presentation Video A. Atalik, S. Chopra, D. Sodickson New York University, New York, United States Impact: By leveraging readily available sources of information which may not generally be used for image reconstruction, our approach reduces ambiguities, enabling more accurate solutions even with highly-sparse measurements.
08:19		Screen Number: 3 1186. The role of nonlinear activations in Fourier-domain neural networks: Noise resilience, regularization, blurring and autocorrelation artifact View Presentation Video P. Dawood, F. Breuer, I. Homolya, M. Gram, P. M. Jakob, M. Zaiss, M. Blaimer University Hospital Erlangen, Erlangen, Germany Impact: This work illuminates the role of nonlinearity in robust artificial neural networks for Fourier-domain interpolation via activated convolution layers at limited training data and shows possible implications of noise resilience, image blurring and autocorrelation artifacts in the image center.
08:21		Screen Number: 4 1187. Efficient deep-learning-based reconstruction of Ferumoxytol-enhanced whole-heart 5D cardiac MRI View Presentation Video K. Borsos, A. Ogier, C. Roy, L. Romanin, M. Stuber, M. Prsa, T. Küstner, J. Yerly Lausanne University Hospital, Lausanne, Switzerland Impact: FreeNet demonstrates potential for rapid inline reconstruction of motion-resolved free-running CMR images for the first time. Our work is a preliminary step towards addressing current roadblocks, bringing “single-click” free-breathing CMR to wider patient populations.
08:23		Screen Number: 5 1188. SELFIE: SElf-supervised Learning for Fast dynamIc golden-anglE radial MRI reconstruction with auto-extracted representations View Presentation Video M. Schellenberg, A. Mekhanik, V. Murray, R. Otazo Memorial Sloan Kettering Cancer Center, New York, United States Impact: SELFIE can achieve comparable performance to supervised deep learning without the limitation of using a compressed sensing reference, which is promising for challenging clinical applications where acquiring a reference is impractical.
08:25		Screen Number: 6 1189. 4D Cardiac Shape Reconstruction Using Image-Aided Neural Signed Distance Correction Fields View Presentation Video Z. Zhang, Z. Liu, Z. Zhang, Z. Cui ShanghaiTech University, Shanghai, China Impact: The 4D cardiac shape sequence of a new patient can be obtained using our method with less time required and the best shape precision among the deep learning-based methods so far.
08:27		Screen Number: 7 1190. FlowMRI-Net: A generalizable self-supervised physics-driven 4D Flow MRI reconstruction network for aortic and cerebrovascular applications View Presentation Video L. Jacobs, M. Piccirelli, V. Vishnevskiy, S. Kozerke ETH and University Zurich, Zurich, Switzerland Impact: FlowMRI-Net facilitates higher undersampling factors than the current state-of-the-art for aortic and cerebrovascular 4D flow MRI within clinically feasible reconstruction times, improving clinical adaptation particularly for cerebrovascular applications which are otherwise too time-consuming.
08:29		Screen Number: 8 1191. Two-minute whole-heart MRI at 0.9 mm3 using a combined super-resolution and motion-corrected undersampled deep-learning reconstruction View Presentation Video A. Phair, S. Littlewood, A. Fotaki, T. Fletcher, L. Felsner, C. Prieto, R. Botnar King's College London, London, United Kingdom Impact: Extending Super-MoCo-MoDL for combined super-resolution and undersampled reconstruction allows sharp whole-heart 3D 0.9-mm³ isotropic-resolution images to be obtained from low-resolution 2-minute scans with 18-fold overall acceleration. This represents a promising approach towards achieving fast high-resolution 3D clinical CMR.
08:31		Screen Number: 9 1192. Spatiotemporal Deformable Unrolling Networks for Cardiac Cine MR Reconstruction View Presentation Video Y. Zhu, J. Cheng, Z-X Cui, D. Gan, Y. Wang, J. Zeng, C. Wang, D. Liang University of Nottingham Ningbo China, Ningbo, China Impact: The proposed method SDUN provides an effective scheme for accurately capturing and learning complex cardiac motions and deformations in cardiac cine MR reconstruction through employing a deformable and adaptive CNN in proximal network.
08:33		Screen Number: 10 1193. A Multi-task Implicit Neural Representation Model for Reconstruction, Motion Field Estimation, and Segmentation of Cardiac Cine MRI View Presentation Video P. D. Negho, N. Vogt, P-A Vuissoz, J. Oster Université de Lorraine, Nancy, France Impact: The proposed model has the potential to speed up cardiac function assessment, by diffusing a single manual segmentation across successive temporal frames and slices, while also providing motion information and offering the possibility to reconstruct a super-resolved volume.
08:35		Screen Number: 11 1194. High spatial resolution cardiac T1 mapping on 1.5T and 3T utilizing deep learning-based image reconstruction View Presentation Video D. Amsel, J. Wetzl, D. Giese, R. Gebker, C. Tillmanns, A. Lingg, P. Krumm, K. Chow, T. Küstner University of Tuebingen, Tuebingen, Germany Impact: The acquisition of higher spatial resolution T1 maps is achieved for both 1.5T and 3T systems. The proposed method may improve the detection of small focal lesions without increasing the required scan time or breath hold duration.
08:37		Screen Number: 12 1195. Machine Learning-Optimized Sampling Enables High-Resolution 3D MR Fingerprinting for Dynamic Quantification of Solute Transport in CSF View Presentation Video Y. Zhu, G. Wang, J. Zhu, R. Boyacioglu, C. Flask, X. Yu Case Western Reserve University, Cleveland, United States Impact: We present a novel method to leverage machine learning to improve sampling for MR fingerprinting. This method achieves simultaneous mapping of T₁ and T₂ in the whole mouse brain at 200-µm isotropic resolution in 4.3 min.
08:39		Screen Number: 13 1196. Enhance-then-Synthesize: A Deep Learning Acceleration Pipeline for Spinal MRI, Reducing Scan Time by 60% View Presentation Video Z. Zhang, J. Li, S. Jiang, Y. Xia, B. Jiang, L. Wang, L. Xiang, L. Fan, S. Liu Subtle Medical Inc., Shanghai, China Impact: The DL-based enhance-then-synthesize pipeline reduced the scanning time for spine MRI protocols up to 60%, while delivering image quality comparable to or higher than that of the standard scanning sequence. The generated sequences proved to be viable alternatives for diagnosis.
08:41		Screen Number: 14 1197. Synthetic FDG PET from functional MRI for epilepsy: development and external validation View Presentation Video C. Yao, J. Lu Xuanwu Hospital Capital Medical University, Beijing, China Impact: Deep learning can generate high-fidelity synthetic PET based on functional MRI Radiologists confirm high imaging fidelity between synthetic and actual PET The imaging features of synthetic PET are similar to actual PET Synthetic PET improves epilepsy diagnosis and prognosis
08:43		Screen Number: 15 1198. Training Diffusion Probabilistic Models with Limited Data for Accelerated MRI Reconstruction with Application to Stroke MRI View Presentation Video S. Kumar, Y. Arefeen, H. Saber, J. Tamir The University of Texas at Austin, Austin, United States Impact: Training diffusion probabilistic models with limited data across various MRI contrasts holds substantial potential to accelerate diverse MRI protocols, addressing a critical unmet need in time-sensitive care scenarios, such as stroke diagnosis.
08:45		Screen Number: 16 1199. Universal MR Image Restoration with Diffusion Models as Plug-and-Play Priors View Presentation Video M. Mostapha, R. Miron, N. Janardhanan, M. Nadar, O. Darwish, T. Huelnhagen, T. Würfl, H. Chandarana, D. Grodzki, R. Schneider Siemens Healthineers, Princeton, United States Impact: The Proposed Diffusion PnP method enables fast and accurate MRI reconstructions using a pre-trained diffusion prior, without the need for fine-tuning or retraining. This approach demonstrates strong potential for diverse clinical applications in MRI.
08:47		Screen Number: 17 1200. Harmonizer: A Deep Learning Framework for Transforming Low-resolution Clinical dMRI to Research-grade Quality View Presentation Video S. Cetin-Karayumak, R. Zurrin, W. Consagra, L. O'Donnell, Y. Rathi Harvard Medical School and Brigham and Women's Hospital, Somerville, United States Impact: his work enhances clinical dMRI data for multi-site neuroimaging studies, enabling high-quality, research-compatible analyses. The Harmonizer algorithm opens opportunities to study various disorders and pathologies using clinical data, supporting detailed white matter analyses and cross-site comparisons.
08:49		Screen Number: 18 1201. 7T MRI-Synthesized Iron and Myelin Histology by Deep Learning View Presentation Video S. Pittayapong, S. Hametner, B. Bachrata, W. Bogner, R. Höftberger, G. Grabner Carinthia University of Applied Sciences, Klagenfurt, Austria Impact: Our research shows that deep learning can synthesize myelin and iron stainings from multi-contrast MRI. This technique enhances understanding of brain development, function, and diseases, promising advances in medical imaging and histology.
08:51		Screen Number: 19 1202. Super-Resolution Image Enhancement for 3D Morphological Lung MRI with bSTAR at 0.55T View Presentation Video P. Panos, M. Pradella, K. Hostettler, O. Bieri, G. Bauman University of Basel, Allschwil, Switzerland Impact: Due to physical and physiological challenges lung MRI is currently rarely used in clinical routine¹. Our study demonstrates that a deep learning-based super-resolution image enhancement results in improved visualization of MRI lung morphology without introducing new anatomical structures.
08:53		Screen Number: 20 1203. Volumetric Measurement Comparisons for Deep Learning Improved Ultra-Low Field MRI View Presentation Video K. T. Islam, S. Zhong, P. Zakavi, H. Kavnoudias, S. Farquharson, G. Durbridge, M. Barth, K. McMahon, P. Parizel, A. Dwyer, G. Egan, M. Law, Z. Chen Monash University, Clayton, Australia Impact: SynthSR and LoHiResGAN were evaluated for their unique approaches to enhancing ultra-low-field MRI images. Their ability to improve volumetric accuracy highlights their potential to support and expand access to high-quality neuroimaging in settings with limited access to high-field MRI.
08:55		Screen Number: 21 1204. Deep learning reconstruction based rapid T2-weighted imaging improved the image quality for primary rectal cancer staging View Presentation Video L. Zhu, B. Shi, K. Wang, W. Feng, J. Dai, Y. Xia, H. Zhang Ruijin Hospital, Shanghai Jiao Tong University of Medicine, Shanghai, China Impact: The application of DLR would be beneficial for rectum T2WI in terms of shorten scan time, improved tumor boundary delineation, and better node characteristic presentation which are important for primary rectal cancer TN staging.
08:57		Screen Number: 22 1205. An End-to-End Deep Learning Method for Reconstructing SMS-PI Accelerated Musculoskeletal MRI View Presentation Video M. Mostapha, G. Koerzdoerfer, E. Raithel, N. Janardhanan, M. Nadar, J. Fritz Siemens Healthineers, Princeton , United States Impact: Proposed methods facilitate the reconstruction of 8-fold accelerated 2D-TSE-MR images across various planes, contrasts, and MSK regions. Preliminary results indicate the feasibility of DL reconstruction at 12-fold acceleration, potentially allowing for significant speed-ups compared to the slower standard of care.
08:59		Screen Number: 23 1206. 2-Minute 3D FSE Knee MRI with 10-fold accelerated Sonic DL – Rapid morphometric and qualitative assessment of Cartilage and Meniscus View Presentation Video L. Carretero, B. Nunes, X. Zhu, E. Sanchez-Lacalle, D. Sundaran, J. Dholakia, M. Fung, M. Padron GE HealthCare, Madrid, Spain Impact: Sonic DL technique enables 10-fold acceleration of 3D FSE, delivering accurate cartilage and meniscus morphometry. Its superior sharpness and lesion conspicuity offer potential for routine clinical use, improving both diagnostic confidence in 3D knee imaging and patient throughput.
09:01		Screen Number: 24 1207. Improved Image Quality, Diagnostic Performance, and Reading Efficiency Based on Deep Learning-Reconstructed Accelerated Rectal MRI W. Peng, S. wang, H. zhang National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Beijing, China Impact: This study offered a comprehensive and viable perspective on the application of DLR in rectal MRI, which facilitated improved image quality and reading efficiency, while reducing acquisition time. Moreover, it enhanced the accuracy of T-staging for junior radiologists.
09:03		Screen Number: 25 1208. Micro Transformer Cascades for Accelerated Multi-Coil MRI Reconstruction with Overlapped SwinV2 Attention View Presentation Video T. Rahman, A. Bilgin, S. Cabrera The University of Texas at El Paso, El Paso, United States Impact: A smaller than conventional SwinV2-Micro Transformer architecture is proposed to facilitate the incorporation of powerful overlapped attention in Transformer cascades for MRI reconstruction. This allows the formation of longer cascades enabling higher reconstruction quality at different acceleration factors.

Similar Session(s)