Magnetic Resonance Spectroscopy (MRS) for GABA Quantification: Understanding Spectrogram Representation to Improve the Training of a Deep-Learning-Based Reconstruction Model

• Grant number: 24/01294-2
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): June 01, 2024
• Effective date (End): December 31, 2024
• Field of knowledge: Engineering - Electrical Engineering
• Principal Investigator: Leticia Rittner
• Grantee: Letícia Levin Diniz
• Host Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

The quantification of gamma-aminobutyric acid (GABA) has been a topic of interest due to the importance of this neurotransmitter and its link to different psychiatric diseases. Magnetic resonance spectroscopy (MRS) is the main method for in-vivo measurement of this metabolite. However, the application of MRS for the quantification of GABA requires the acquisition of multiple resonance signals (transients), which are processed to obtain a GABA-edited spectrum. Typically, 320 transients are acquired to obtain the spectrum, which implies long acquisitions that are uncomfortable for the patient, costly and difficult to plan. Given the need to obtain a good quality GABA-edited spectrum using fewer transients, thus considerably reducing data acquisition time, the Edited-MRS Reconstruction Challenge proposed the use of learning models to reconstruct the spectrum using only 80 transients. The winning group presented the SpectroViT model, which uses as input data the spectrogram derived from the combination of GABA-edited transients. Despite SpectroViT's good performance, the impact of the spectrogram's characteristics on its behavior has not yet been explored. In order to better understand the input data and how it relates to the model, this work aims to understand how the acquisition characteristics present in the GABA-edited spectrum are represented in the spectrogram domain, and how this knowledge can be applied to improve SpectroViT's performance through strategies such as data augmentation. This in-depth characterization of the spectrogram of GABA-edited signals is crucial for a better understanding of the proposed model and the factors that influence its effectiveness, a fundamental aspect for evaluating its possible clinical applications.