Hankel Subspace Self-Supervised Reconstruction (HSSRecon) is a scan‑specific framework for parallel MRI that partitions data by acquisition units before Hankel lifting and normalizes repeated Hankel copies. Instead of directly predicting missing k‑space data, the network learns a compact complex‑valued Hankel subspace operator, and reconstruction is performed with a conjugate‑gradient solver enforcing hard data consistency. The method is theoretically justified with proofs of positive definiteness, uniqueness, and error bounds, and demonstrates competitive performance on fastMRI brain data across multiple contrasts and sampling masks.
arXiv:2606. 00100v1 Announce Type: cross Abstract: Self-supervised deep learning-based methods have shown great promise for accelerated magnetic resonance imaging (MRI) reconstruction, achieving high image quality without requiring fully sampled data for training.
By Tongxi Song, Ziyu Li, Zihan Li, Wen Zhong, Congyu Liao, Yang Yang, Hua Guo, Wenchuan Wu, Qiyuan Tian
arXiv:2603. 04438v3 Announce Type: replace-cross Abstract: Fully unsupervised deep generative modeling (FU-DGM) offers significant potential for compressively sampled magnetic resonance imaging (CS-MRI) reconstruction.
By Qingyong Zhu, Yumin Tan, Xiang Gu, Dong Liang
The paper introduces FLEX‑MRI, a method that accelerates MRI reconstruction by partitioning the recovery process into a few finite intervals defined by progressive k‑space support changes. Instead of learning fine‑grained local dynamics, FLEX‑MRI directly learns the reconstruction change across each interval and employs stochastic exploration during inference to sample alternative transition paths. Experiments on multi‑coil MRI datasets demonstrate that FLEX‑MRI outperforms competing methods while keeping computational efficiency comparable to few‑step baselines.
By Muhammad Usama Mirza, Esin \"Ozt\"urk I\c{s}{\i}k, Tolga \c{C}ukur
The paper introduces τ+C+Mag, a physics‑driven deep learning method that incorporates auxiliary k‑space magnitude information into accelerated steady‑state dynamic MRI reconstruction. By observing strong consistency of k‑space magnitudes across time‑frames, the authors develop an ADMM‑based unrolling framework with a magnitude‑aware data‑fidelity term, using quadratically smoothed optimization and momentum updates to handle non‑differentiability and non‑convexity. Experiments on retrospectively and prospectively undersampled cine and phase‑contrast flow MRI datasets show improved artifact suppression, sharper anatomical detail, and better phase preservation compared to conventional PD‑DL approaches, as confirmed by blinded expert readers.
By Mahdi Saberi, Ya\c{s}ar Utku Al\c{c}alar, Merve G\"{u}lle, Chetan Shenoy, Mehmet Ak\c{c}akaya
The paper introduces Neuro‑JEPA, a sparse multimodal foundation model that learns unified representations of brain MRI across T1w, T2w, and FLAIR sequences using a latent predictive objective and a Mixture‑of‑Experts architecture. It was pretrained on over 1.5 million scans from 428,647 studies and systematically evaluates architectural, masking, objective, and sparsity choices for robust multimodal representation learning. Across 47 tasks from three health systems and 12 public datasets, Neuro‑JEPA consistently outperforms a simple CNN baseline, demonstrating its effectiveness for both clinical and research applications.
By Haoxu Huang, Long Chen, Jingyun Chen, Jinu Hyun, James Ryan Loftus, Kara Melmed, Daniel Orringer, Jennifer Frontera, Seena Dehkharghani, Arjun Masurkar, Narges Razavian