Mixture Model Based Efficient Method for Magnetic Resonance Spectra Quantification

Abstract

Magnetic resonance spectroscopy (MRS) is a popular technique used in oncology to identify type of tumor and its progress with respect to its specific metabolism different than in normal tissue. In this study a complete pre-processing pipeline resulting in identification and quantification of chemical compounds has been proposed comprising novel method based on Gaussian mixture model (GMM). Model parameters were estimated with use of modified EM algorithm initialized by a new idea of spectrum segmentation. In order to make such solution ready to use in clinical applications an implementation based on GPU calculations is introduced. On simulated dataset we analyzed proposed methods by computational speed and data transfer time. On phantom data we compared our method to the two popular solutions: LC Model and Tarquin. It was observed that proposed algorithm outperforms both methods in sense of accuracy and precision of estimated concentration. The most efficient implementation was based on GPU with single precision calculations giving huge speed-up and satisfactory model accuracy comparing to CPU-based algorithm.