A comparative study of feature extraction and blind source separation of independent component analysis (ICA) on childhood brain tumour 1H magnetic resonance spectra

Application of ICA to realistically simulated (1)H-MRS data

INTRODUCTION

(1)H-MRS signals from brain tissues capture information on in vivo brain metabolism and neuronal biomarkers. This study aims to advance the use of independent component analysis (ICA) for spectroscopy data by objectively comparing the performance of ICA and LCModel in analyzing realistic data that mimics many of the known properties of in vivo data.

METHODS

This work identifies key features of in vivo (1)H-MRS signals and presents methods to simulate realistic data, using a basis set of 12 metabolites typically found in the human brain. The realistic simulations provide a much needed ground truth to evaluate performances of various MRS analysis methods. ICA is applied to collectively analyze multiple realistic spectra and independent components identified with our generative model to obtain ICA estimates. These same data are also analyzed using LCModel and the comparisons between the ground-truth and the analysis estimates are presented. The study also investigates the potential impact of modeling inaccuracies by incorporating two sets of model resonances in simulations.

RESULTS

The simulated fid signals incorporating line broadening, noise, and residual water signal closely resemble the in vivo signals. Simulation analyses show that the resolution performances of both LCModel and ICA are not consistent across metabolites and that while ICA resolution can be improved for certain resonances, ICA is as effective as, or better than, LCModel in resolving most model resonances.

CONCLUSION

The results show that ICA can be an effective tool in comparing multiple spectra and complements existing approaches for providing quantified estimates.

Group independent component analysis of MR spectra

This study investigates the potential of independent component analysis (ICA) to provide a data-driven approach for group level analysis of magnetic resonance (MR) spectra. ICA collectively analyzes data to identify maximally independent components, each of which captures covarying resonances, including those from different metabolic sources. A comparative evaluation of the ICA approach with the more established LCModel method in analyzing two different noise-free, artifact-free, simulated data sets of known compositions is presented. The results from such ideal simulations demonstrate the ability of data-driven ICA to decompose data and accurately extract components resembling modeled basis spectra from both data sets, whereas the LCModel results suffer when the underlying model deviates from assumptions, thus highlighting the sensitivity of model-based approaches to modeling inaccuracies. Analyses with simulated data show that independent component weights are good estimates of concentrations, even of metabolites with low intensity singlet peaks, such as scyllo-inositol. ICA is also applied to single voxel spectra from 193 subjects, without correcting for baseline variations, line-width broadening or noise. The results provide evidence that, despite the presence of confounding artifacts, ICA can be used to analyze in vivo spectra and extract resonances of interest. ICA is a promising technique for decomposing MR spectral data into components resembling metabolite resonances, and therefore has the potential to provide a data-driven alternative to the use of metabolite concentrations derived from curve-fitting individual spectra in making group comparisons.

A hybrid method of application of independent component analysis to in vivo 1H MR spectra of childhood brain tumours

Independent component analysis (ICA) can automatically extract individual metabolite, macromolecular and lipid (MMLip) components from a series of in vivo MR spectra. The traditional feature extraction (FE)-based ICA approach is limited, in that a large sample size is required and a combination of metabolite and MMLip components can appear in the same independent component. The alternative ICA approach, based on blind source separation (BSS), is weak when dealing with overlapping peaks. Combining the advantages of both BSS and FE methods may lead to better results. Thus, we propose an ICA approach involving a hybrid of the BSS and FE techniques for the automated decomposition of a series of MR spectra. Experiments were performed on synthesised and patient in vivo childhood brain tumour MR spectra datasets. The hybrid ICA method showed an improvement in the decomposition ability compared with BSS-ICA or FE-ICA, with an increased correlation between the independent components and simulated metabolite and MMLip signals. Furthermore, we were able to automatically extract metabolites from the patient MR spectra dataset that were not in commonly used basis sets (e.g. guanidinoacetate).

Classification of single-voxel 1H spectra of brain tumours using LCModel

This study presents a novel method for the direct classification of (1)H single-voxel MR brain tumour spectra using the widespread analysis tool LCModel. LCModel is designed to estimate individual metabolite proportions by fitting a linear combination of in vitro metabolite spectra to an in vivo MR spectrum. In this study, it is used to fit representations of complete tumour spectra and to perform a classification according to the highest estimated tissue proportion. Each tumour type is represented by two spectra, a mean component and a variability term, as calculated using a principal component analysis of a training dataset. In the same manner, a mean component and a variability term for normal white matter are also added into the analysis to allow a mixed tissue approach. An unbiased evaluation of the method is carried out through the automatic selection of training and test sets using the Kennard and Stone algorithm, and a comparison of LCModel classification results with those of the INTERPRET Decision Support System (IDSS) which incorporates an advanced pattern recognition method. In a test set of 46 spectra comprising glioblastoma multiforme, low-grade gliomas and meningiomas, LCModel gives a classification accuracy of 90% compared with an accuracy of 95% by IDSS.

Non-invasive detection of glycine as a biomarker of malignancy in childhood brain tumours using in-vivo 1H MRS at 1.5 tesla confirmed by ex-vivo high-resolution magic-angle spinning NMR

Management of brain tumours in children would benefit from improved non-invasive diagnosis, characterisation and prognostic biomarkers. Metabolite profiles derived from in-vivo MRS have been shown to provide such information. Studies indicate that using optimum a priori information on metabolite contents in the construction of linear combination (LC) models of MR spectra leads to improved metabolite profile estimation. Glycine (Gly) is usually neglected in such models due to strong overlap with myo-inositol (mI) and a low concentration in normal brain. However, biological studies indicate that Gly is abundant in high-grade brain tumours. This study aimed to investigate the quantitation of Gly in paediatric brain tumours using MRS analysed by LCModel, and its potential as a non-invasive biomarker of malignancy. Single-voxel MRS was performed using PRESS (TR 1500 ms, TE 30 ms/135 ms) on a 1.5 T scanner. Forty-seven cases (18 high grade (HG), 17 low grade (LG), 12 ungraded) were retrospectively selected if both short-TE and long-TE MRS (n = 33) or short-TE MRS and high-resolution magic-angle spinning (HRMAS) of matched surgical samples (n = 15) were available. The inclusion of Gly in LCModel analyses led to significantly reduced fit residues for both short-TE and long-TE MRS (p < 0.05). The Gly concentrations estimated from short-TE MRS were significantly correlated with the long-TE values (R = 0.91, p < 0.001). The Gly concentration estimated by LCModel was significantly higher in HG versus LG tumours for both short-TE (p < 1e-6) and long-TE (p = 0.003) MRS. This was consistent with the HRMAS results, which showed a significantly higher normalised Gly concentration in HG tumours (p < 0.05) and a significant correlation with the normalised Gly concentration measured from short-TE in-vivo MRS (p < 0.05). This study suggests that glycine can be reliably detected in paediatric brain tumours using in-vivo MRS on standard clinical scanners and that it is a promising biomarker of tumour aggressiveness.