Alexander Giuffrida
BME PhD Defense Presentation

Date: 2025-05-23
Time: 8:30
Location / Meeting Link: Winship C4018 / https://zoom.us/j/5582282841?omn=94111537142

Committee Members:
Hyunsuk Shim, PhD (Advisor); Brent Weinberg, MD/PhD; Lee Cooper, PhD; Brian Soher, PhD; Shella Keilholz, PhD


Title: Deep Learning Methods for Magnetic Resonance Spectroscopic Neuroimaging

Abstract:
Spectroscopic MRI (sMRI) is a powerful technique for mapping brain metabolism in vivo, offering noninvasive insight into the molecular underpinnings of neurological disease. However, its clinical translation has been limited by technical challenges in spectral quantification, quality control, and interpretation. In this dissertation, I present a series of deep learning methods aimed at accelerating and improving the robustness of sMRI analysis, with a particular focus on brain tumor imaging. The core of this work is NNFit, a self-supervised neural network trained to perform spectral fitting directly on in vivo EPSI data. NNFit achieves quantification performance comparable to traditional model-based methods (e.g., FITT) while reducing inference time from 45 minutes to under 15 seconds per scan. The model is evaluated on both short and long echo time acquisitions and is integrated into a clinical workflow for glioblastoma radiation therapy planning. I also describe Onix, a custom-built visualization and evaluation platform developed to compare quantification methods, filter low-quality spectra, and compute statistical metrics across anatomical regions. To address the issue of spectral quality, I present a deep learning classifier trained to detect artifacts in sMRI data, which can be used to filter unreliable spectra prior to downstream analysis. Finally, I discuss preliminary work on synthetic spectral generation and future directions for integrating uncertainty estimation, active learning, and broader clinical deployment. Together, these contributions represent a unified framework for scalable, interpretable, and clinically relevant deep learning pipelines for spectroscopic neuroimaging.