Impact: A specific type of artifacts is observed for scan-specific real-valued neural networks. The source of these artifacts is discussed.

Impact: This work enables accelerated ultra-low-field MRI by integrating physics-based consistency with a time-conditioned self-supervised unrolled network. The method reconstructs 3D images from undersampled data without paired training datasets, paving the way for affordable, portable MRI in resource-limited settings.

Impact: With dual-domain self-supervised learning applied to the 3D non-Cartesian mcUTE sequence, whole-brain myelin quantification and high-resolution anatomical images could be attained within 4 minutes at 3T. In vivo and phantom results indicate that proposed methods outperform conventional reconstruction techniques.

Impact: This work introduces the first INR model to predict high-resolution voxel-level SUVR from regional data (R=0.959). This offers a scalable and data-efficient solution, providing clinicians with the detailed maps they demand for early detection and individualized assessment in Alzheimer's Disease.

Impact: Coil-STRAINER establishes a calibrationless, self-supervised, subject-specific reconstruction paradigm that efficiently models both inter-coil dependencies and coil-specific features, overcoming the long training burden of conventional INRs and advancing practical, patient-specific parallel MRI without large datasets or fully-sampled ACS lines.

Impact: This work enables high-quality, low-cost pseudo-high-field MRI from low-field systems, improving diagnostic accuracy and accessibility. It offers scalable solutions for resource-limited settings and opens avenues for future self-supervised learning in medical imaging.

Impact: This work proposes the first joint multi-contrast reconstruction method capable of handling arbitrary heterogeneous imaging protocols. As it does not use population-derived priors and functions independent of image resolution, it is a significant step towards generalizability of advanced reconstruction methods.

Impact: A single rapid MPnRAGE scan with multiple inversion contrasts enables quantitative T1 mapping, generation of tissue-suppressed images, and synthesis of a standard MPRAGE, all from the same dataset, reducing overall scan time and enhancing diagnostic versatility.

Impact: The proposed method achieves high-quality MRI reconstruction at high acceleration factors using only undersampled measurement, demonstrating broad potential for clinical application.

Impact: By jointly estimating the image and coil sensitivities, our training-data-free approach overcomes a key limitation in accelerated MRI. This enables higher fidelity reconstructions without external data, offers a robust pathway to improved diagnostic quality, and advances deep learning reconstruction research.

Impact: We present a zero-shot framework for learning annihilation relations. Our temporal-annihilation-filter model improves zero-shot myocardial perfusion reconstruction, potentially enabling deep-learning gains, such as greater spatial coverage, spatiotemporal resolution, and SNR, without external training data, mitigating generalization concerns for patient-specific imaging.

Impact: This study pioneers the first self-supervised four-dimensional magnetic resonance fingerprinting (SS-4DMRF) framework, enabling precise, motion-resolved tissue quantification. It empowers clinically accessible four-dimensional quantitative imaging for radiotherapy planning and unlocks new possibilities for data-driven biomarker discovery in oncology.

Impact: Our proposed Lorentz Subspace Encoding (LSE) can accurately reconstruct sparsely-sampled CEST Z-spectra, enabling reliable quantification of parameters, such as APT, NOE, MT.

Impact: ImMAP provides an interpretable and robust denoising-diffusion based reconstruction baseline that is practical for real clinical data. It enables improved image quality at high accelerations while avoiding the instability and parameter sensitivity common in current diffusion MRI methods.