Impact: By embedding radiomics and clinical data within a self-attention Transformer, we created a single, reproducible model that boosts pCR prediction accuracy without extra acquisitions or reader-intensive steps—offering clinicians a ready-to-use decision aid for tailoring neoadjuvant therapy in breast cancer.

Impact: A principled scaling study and a universal, large-capacity unrolled design offer a practical blueprint for building and deploying foundation-style MRI reconstruction models across heterogeneous clinical protocols.

Impact: This study indicates that AI denoising reconstruction can accelerate cardiac cine imaging without altering ejection fraction and strain measurements, improving confidence in these acceleration and image enhancement techniques.

Impact: This work demonstrates the first automatic IVDD detection from ultra-low-field L-spine MRI at 0.05 Tesla. The results will facilitate low-cost spinal health screening, advancing ULF MRI towards truly accessible and point-of-care clinical applications.

Impact: The self-supervised masked autoencoder model, SpectrumMAE, enabled robust super-resolution for 1H-MRSI of gliomas without the need for long scan times.

Impact: DLR can significantly improve the image quality of placental SSFSE and FIESTA imaging, shorten the diagnostic time, and increase the diagnostic accuracy of PAS disorder, which may provide potential benefits for the clinical assessment of placental abnormalities.

Impact: By modeling edge-level temporal changes, the proposed model enables enhanced and efficient prediction of longitudinal brain connectivity. This will support the identification and monitoring of neurodegenerative diseases and assist researchers in studying how brain connectivity patterns evolve over time.

Impact: The developed T2WI deep learning model accurately predicts malignant IPMN/MCN, enabling non-invasive diagnosis and guiding treatment decisions, thereby improving patient management.

Impact: This study demonstrates that a Swin-Transformer-based DCE-MRI deep learning model can accurately predict non-sentinel lymph node metastasis, enabling more precise surgical decisions, reducing unnecessary axillary dissections, and inspiring future research on AI-driven personalized breast cancer management.

Impact: The machine learning prediction model developed in this study can assist clinicians to accurately select candidates for breast-conserving surgery and improve the safety and individualization of treatment decisions.

Impact: We demonstrated the feasibility of prostate MRI at 0.55T using deep learning reconstruction. With further study to evaluate diagnostic performance in patients, the benefits of low-field MRI can be employed for prostate cancer screening in clinical practice.

Impact: This study establishes that explicit modeling of intratumoral perfusion heterogeneity significantly enhances response prediction. The proposed subregional phenotyping approach offers an interpretable biomarker for optimizing neoadjuvant therapy strategies, potentially facilitating early treatment adaptation in ER+ breast cancer.

Impact: This MRE-guided machine learning approach provides clinicians with a simple, accurate, and accessible tool for non-invasive liver fibrosis assessment, particularly excelling in early-stage fibrosis identification that is crucial for timely intervention.

Impact: AI-assisted segmentation of white matter hyperintensities provides objective, region-specific evaluation in acute ischemic stroke. This approach supports predictive modeling of white matter damage, and advances automated tools for individualized cerebrovascular evaluation and clinical decision-making.

Impact: 
DL-SR is attributed to its dual optimization mechanism: Compressed Sensing reduces raw data acquisition through sparse sampling for accelerated scanning; Dual CNN reconstruction produces high-quality images with enhanced SNR, improved clarity, larger matrix size, and reduced truncation artifacts.