BaMRSDomainAdaptation

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

THIS WEBPAGE IS DEPRECATED - please visit our new website

Bachelor thesis: Improving Generalizability through Generative Adversarial Domain Adaptation for MR Spectroscopy

Advisor: Bjoern Menze
Supervision by: John LaMaster

Abstract

Spectroscopy is a technique that, when used in medicine, uses magnetic properties to evaluate the chemical composition of tissue of interest in a noninvasive manner. Analyzing quantified metabolite ratios can help physicians to differentiate between physiological and pathological tissues. This is an advanced topic that can be simplified to signal processing. In collaboration with a MRS research group at University of California, San Francisco, we are looking to develop a tool to aid in the training of DL models for spectra analysis. High quality, processed, and annotated medical data is expensive in general. This is even more problematic with MRS data. Current physics models can generate synthetic spectra, but due to the nonlinearities inherent in spectroscopy data and its acquisition protocols, models trained on these spectra do not perform well when tested on real spectra.

Project Description

PDF: Improving Generalizability through Generative Adversarial Domain Adaptation for MR Spectroscopy

Tasks

Develop and optimize a GAN-based framework to convert clinically recorded MRSI data to a synthetic feature space in order to improve the performance of machine learning quantification models trained on synthetic spectra.

Requirements

  • Machine Learning background, especially with GANs
  • Proficient with PyTorch?
  • Signal processing is beneficial
  • Ambitious with a strong motivation
  • Ability to work independently

Magnetic Resonance Spectropy

MRS is a quantitative MR technique used to characterize the metabolism or metabolic composition of tissue in vivo. This is done by analyzing the specific frequencies present in the FID, instead of just considering the magnitude which is used for qualitative imaging. These frequencies are metabolite dependent, making them easy to identify. While all protons resonate at the same Larmor frequency, the precessions that occur when excited by the RF pulse cause small changes in the magnetic field of the nucleus. MR-visible molecules are visible because there are many of these small perturbations that cumulatively change the local effective magnetic field of the molecule. This change in the magnetic field changes the molecule's resonant frequency. In the frequency domain, this change in resonant frequency shifts the frequency peak to the left or the right. This chemical shift is unique to each molecule and by comparing peaks of different metabolites, maps of relative metabolite concentrations can be generated. This type of information is useful for characterizing tissue, including potential tumors.

Contact

John LaMaster
Bjoern Menze

Students.ProjectForm
Title: Improving Generalizability through Generative Adversarial Domain Adaptation for MR Spectroscopy
Abstract: Spectroscopy is a technique that, when used in medicine, uses magnetic properties to evaluate the chemical composition of tissue of interest in a noninvasive manner. Analyzing quantified metabolite ratios can help physicians to differentiate between physiological and pathological tissues. This is an advanced topic that can be simplified to signal processing. In collaboration with a MRS research group at University of California, San Francisco, we are looking to develop a tool to aid in the training of DL models for spectra analysis. High quality, processed, and annotated medical data is expensive in general. This is even more problematic with MRS data. Current physics models can generate synthetic spectra, but due to the nonlinearities inherent in spectroscopy data and its acquisition protocols, models trained on these spectra do not perform well when tested on real spectra.
Student: Linus Kreitner
Director: Bjoern Menze
Supervisor: John LaMaster
Type: DA/MA/BA
Area: Machine Learning, Medical Imaging, Computer Vision
Status: finished
Start: Dec 2020
Finish: Mar 2021
Thesis (optional):  
Picture:  


Edit | Attach | Refresh | Diffs | More | Revision r1.2 - 08 Jul 2021 - 17:57 - JohnLaMaster