Cape Town - 2026 ISMRM-ISMRT Annual Meeting and Exhibition - Deep Learning Meets k-Space: New Frontiers in Reconstruction

Deep Learning Meets k-Space: New Frontiers in Reconstruction

Digital Poster

Acquisition & Reconstruction

Wednesday, 13 May 2026

Digital Posters Row B

14:35 - 15:30

Session Number: 561-04

No CME/CE Credit

Similar Sessions

This session highlights AI-powered image reconstruction methods.

		561-04-001. Accelerating MRI reconstruction with cross feature fusion variational network Xihui Ju, Jian He, zhongjie jia, zhenzhen wang, wang du, jingchao zhang, shuangfeng dai, Hongbin Wang Beijing Wandong Medical Technology Co., Ltd, Beijing, China Impact: This study aimed to develop a novel deep learning method named cross feature fusion variational network (CFFVN) for accelarated MR reconstruction to benefit radiologists by enabling fast and high-quality MR imaging, reducing patient scan time and enhancing diagnostic accuracy.
		561-04-002. Hybrid 3D CNN-Transformer for Accelerated and Robust Cerebral Blood Volume (CBV) Reconstruction from PWI-MRI Akito Yamauchi, Zongyu Li, Zhonghui Qie, Jia Guo Columbia University, New York, United States of America Impact: Our hybrid 3D CNN-Transformer (3D ConvFormer) reconstructs motion-robust CBV maps from PWI using >2.5-fold fewer time-points (15 vs. 40). It provides superior artifact correction and quantitative accuracy compared to conventional methods and a baseline spatiotemporal CNN (ST-CNN).
		561-04-003. MR Deep Learning Reconstruction Method with Controllable Frequency Weight After Learning Satoshi ITO, Kotaro ADACHI, Fumiya CHUBACHI Utsunomiya University, Utsunomiya, Japan Impact: A novel method is proposed that enables diverse representations of detailed structures by introducing a new technique for controlling the frequency components of images. It is expected to provide a wealth of information in image diagnosis and enhance diagnostic accuracy.
		561-04-004. Mitigating Divergence in PINN(Physics-Informed Neural Network)-MREPT using Stepwise Training and Collocation Enhancement Ruian Qin, Junqi YANG, shaoying huang, Wenwei Yu Chiba University, Chiba, Japan Impact: Enhances MREPT reconstruction stability and accuracy through physics-guided, stepwise learning process without requiring ground truth data.
		561-04-005. Deep Learning Reconstruction for Rapid Ankle MRI: A Clinical Feasibility Study youbin Ding, Mo Xianfu, Shaoyong Hu, Yiming Wang, Zhongping Zhang, Peng Wu, Xiaodong Zhang The Third Affiliated Hospital Southern Medical University, Guangzhou, China Impact: DL protocols can significantly accelerate clinical scans while maintaining image quality, making them clinically valuable. However, further research is needed to determine whether DL reconstruction can surpass conventional imaging in detecting subtle features for earlier disease diagnosis.
		561-04-006. From Thick to Thin: a High-fidelity and Robust Reconstruction Framework for Brain Tumor MRI Liqin Yang, Caohui Duan, Xin Lou, Dinggang Shen, Kaicong Sun ShanghaiTech University, Shanghai, China Impact: We propose a clinical solution to accelerating brain tumor MRI. Our reconstruction framework supports arbitrary view of 2D acquisitions and multiple interpolation rates. It delivers both thick-slice reconstruction for tumor diagnosis and thin-slice reconstruction for surgical resection without 3D acquisition.
		561-04-007. On the Clinical Value of Deep Learning Image Reconstruction to Accelerate Submillimeter Resolution Imaging at 7T Jocelyn Philippe, Kevin Battistini, Caterina Bernetti, Arsany Hakim, Marwan El-Koussy, Natalia Pato Montemayor, Marcel Dominik Nickel, Patrick Liebig, Robin Heidemann, Felix Kurz, Jean-Philippe Thiran, Tom Hilbert, Tobias Kober, Gabriele Bonanno, Gian Franco Piredda, Piotr Radojewski, Thomas Yu Siemens Healthineers International AG, Lausanne, Switzerland Impact: This study demonstrates that deep learning reconstruction can accelerate high-resolution 7T brain imaging while preserving diagnostic quality, enabling shorter scan times. These findings may influence clinical workflows and future research on reliable AI-based imaging in patients with brain pathologies.
		561-04-008. Multi-Domain Adaptive Fusion Cascade Network (AFCN) with Metabolite-Aware Loss for Accelerated MRSI Reconstruction Nate Tran, Sana Vaziri, Abdullah Bas, Jenny Lee, Jacob Ellison, Irvane Kamga, Angela Jakary, Yan Li, Esin Ozturk Isik, Janine Lupo University of California San Francisco, San Francisco, United States of America Impact: Our metabolite-aware, multi-domain adaptive fusion cascade network provides faster and more reliable reconstruction of 4-fold undersampled MRSI compared to other current approaches. This enables clinically practical acquisition of high-resolution, long-echo 3D-MRSI by significantly shortening scan times for routine use.
		561-04-009. Breast DCE-MRI Lesion Delineation: Interactive Self-Correcting AI Vs. Dual-Expert Manual Contouring in a Multi-Reader Study Zhitian Guo, Moyun Zhang, Shuo Wang, Xinyue Yin, Haonan Guan, Lina Zhang The First Affiliated Hospital of Dalian Medical University, Dalian, China Impact: Interactive AI reduces inter-reader dispersion and improves geometric consistency while preserving expert-level volumetry, supporting standardized, auditable breast DCE-MRI contouring in clinical practice and trials.
		561-04-010. Overcoming the Spatial-Temporal Trade-off in DCE MRI Using Recurrent Inference Machines Dilara Tank, Desirée van den Berg, Nienke Wassenaar, Eric Schrauben, Matthan Caan, Oliver Gurney-Champion Amsterdam University Medical Center, University of Amsterdam, Amsterdam, Netherlands Impact: Quantitative DCE biomarkers have the potential to drive personalized cancer treatments. However, DCE imaging faces a trade-off between accurate parameter mapping and high resolution. Here, we use AI reconstructions to enable both as a first step towards personalized cancer care.
		561-04-011. The impact of deep learning reconstruction on T1-weighted structural image quality (MRIQC) and brain morphometry (FreeSurfer) Synne Johnsen, Robin Antony Birkeland Bugge, Dag Alnæs, Stener Nerland, Ellen Olsrud, Henning Rise, Lars Westlye, Wibeke Nordhøy Norwegian University of Science and Technology, Trondheim, Norway Impact: Using deep learning reconstruction improves the quality and efficiency of structural T1-weighted MRI, as confirmed by quantitative measures using MRIQC. Enhanced image quality can impact morphometric measurements with FreeSurfer, potentially affecting the consistency and comparability of longitudinal data collection.
		561-04-012. Accelerating Ultra-High Field T2 SPACE Acquisitions by Combining Coherent and Incoherent Undersampling Thomas Yu, Marcel Dominik Nickel, Constantin von Deuster, Jürgen Herrler, Robin Heidemann, Patrick Liebig, Tom Hilbert, Gian Franco Piredda Siemens Healthineers International AG, Lausanne, Switzerland Impact:* This study demonstrates the effectiveness of combining incoherent and coherent undersampling with deep learning reconstruction for further accelerating T2 SPACE acquisitions and reconstructions with submillimeter resolution at 7T, allowing up to 40% decrease in acquisition time without compromising image quality.
		561-04-013. Cross-Cascade Feature Aggregation for Improved Spatio-Temporal Reconstruction in Cardiac Cine MRI Donghang Lyu, Marius Staring, Yiming Dong, Jochen Keupp, Hildo Lamb, Mariya Doneva Leiden University Medical Center, Leiden, Netherlands Impact: This study explores how integrating features from multiple preceding cascade blocks in unrolled networks can improve cine MRI reconstruction, potentially enabling more accurate and temporally consistent cardiac imaging.
		561-04-014. Impact of highly accelerated deep learning reconstruction on common brain segmentation software pipelines Kain Kyle, Taylor Emsden, Daniel Cornfeld, Paul Condron, Sergio Dempsey GE HealthCare (Sydney, AUS), Sydney, Australia Impact: For research workflows using structural T1w images for atlas-based segmentation, based on an n of 1, DLSpeed acceleration factors of 18 may retain excellent agreement in larger structures, while smaller structures and clinical use may require a factor of 10.
		561-04-015. Deep Learning Based Reconstruction Method for T2 Mapping via Multiple Overlapping-Echo Detachment Acquisition Che Wang, Linyu Fan, Yingying Wang, Chenyang Dai, Qizhi Yang, Jianfeng Bao, Liangjie Lin, Zhong Chen, Congbo Cai, Shuhui Cai Fujian Provincial Key Laboratory of Plasma and Magnetic Resonance, Xiamen University, Xiamen, China Impact: We propose a parallel imaging reconstruction method for high-quality METMOLED images in single-shot METMOLED T2 mapping, which may also be applicable to other magnetic resonance quantitative reconstruction.