MechE PhD Prospectus Defense: Mengyu Li

TITLE: DATA EFFICENT AND TRANSFERABLE ARTIFICIAL INTELLIGENCE FOR MEDICAL MRI: FROM IMAGE RESTORATION AND REPRESENTATION LEARNING TO CLINICAL DIAGNOSISABSTRACT: Artificial intelligence (AI) has demonstrated strong potential for magnetic resonance imaging (MRI), yet many existing methods remain task-specific, data-intensive, and sensitive to changes in acquisition conditions, downstream tasks, datasets, and clinical populations. This prospectus develops a unified framework for data-efficient and transferable AI in medical MRI, progressing from image restoration and general representation learning to clinically guided characterization and longitudinal prediction. The first part investigates generalizable and transferable MRI representations. The Magnetic Resonance Image Processing Transformer (MR-IPT) was developed for accelerated MRI restoration across heterogeneous undersampling patterns and acceleration factors, demonstrating that shared pretrained representations can generalize to previously unseen acquisition settings. Building on this concept, a masked autoencoder was pretrained on approximately 31 million heterogeneous brain MRI slices and adapted under few-shot conditions to classification and segmentation tasks. These studies demonstrate that large-scale self-supervised MRI pretraining can support accurate downstream analysis while substantially reducing dependence on task-specific annotation. The second part extends transferable representation learning toward clinically guided MRI interpretation. NeuroBridge integrates large-scale self-supervised pretraining with complementary structural objectives, including hippocampal segmentation, hippocampal atrophy classification, and reconstruction, followed by adaptive representation fusion for neurodegenerative disease characterization. The framework demonstrated improved current-state classification and cross-cohort transfer and supported an exploratory retrospective simulation of confidence-aware opportunistic screening. Continuing work expands this clinically guided representation with white matter hyperintensity information from FLAIR MRI. The proposed longitudinal work will determine whether structural information available in an index brain MRI can predict subsequent cognitive decline. Index MRI examinations will be linked to later Clinical Dementia Rating outcomes within baseline cognitive strata, while accounting for follow-up interval and cohort heterogeneity. Complementary paired-MRI analyses will characterize progression-associated representation changes. Collectively, this research seeks to establish transferable MRI representations that remain useful across heterogeneous acquisition conditions, analytical tasks, clinical populations, and temporal prediction endpoints, providing a foundation for more adaptable and clinically meaningful AI in medical MRI. COMMITTEE: ADVISOR/CHAIR Professor Xin Zhang, ME/ECE/BME/MSE, Professor Chad Farris, Radiology, Professor Sean Andersson, ME/SE, Professor Lei Tian, ECE