[Defense] Investigating Deep Learning for Enhancing and Predicting Low-Quality MRI: From Loss Functions to Temporal Modeling
Wednesday, April 16, 2025
1:00 pm - 2:30 pm
In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
Mohammad Javadi
will defend his proposal
Investigating Deep Learning for Enhancing and Predicting Low-Quality MRI: From Loss Functions to Temporal Modeling
Abstract
Magnetic Resonance Imaging (MRI) is a powerful tool in diagnostic imaging, yet its accessibility is often limited by scan time, hardware cost, and infrastructural demands. These constraints give rise to a wide range of low-quality (LQ) images, commonly caused by aggressive k-space undersampling or inherently low signal strength in low-field (LF) MRI systems (e.g., the FDA-approved 64mT Hyperfine Swoop). This dissertation investigates deep learning (DL) methods for enhancing and predicting high-quality MRI from such degraded inputs, with a focus on the technical assessment of generative models and their training strategies. The first direction explores the role of adversarial loss in medical image enhancement. While adversarial components are frequently included in fusion loss functions alongside pixel and perceptual terms, their clinical impact remains unclear. We perform a systematic evaluation of adversarial loss weightings and introduce lesion-focused metrics to capture clinically relevant improvements in sharpness and structure. The second direction investigates diffusion models, specifically SR3, as an alternative to GAN-based approaches. These models offer stable training and superior reconstruction of high-frequency features, making them well-suited for LF-to-HF image translation, yet their effectiveness in pathologically diverse clinical settings is largely untested. The final part of the dissertation outlines two ongoing research threads. The first is a longitudinal prediction framework that leverages dual-encoder networks and tumor evolution modeling to forecast future HF-quality scans. The second explores meta-learning techniques aimed at building robust, domain-adaptive models for diverse LQ imaging conditions
Wednesday, April 16, 2025
1:00 PM - 2:30 PM
PGH 550 and MS Teams
Dr. Nikolaos Tsekos, proposal advisor
Faculty, students, and the general public are invited.