Linear discriminant analysis of brain tumour 1H MR spectra: a comparison of classification using whole spectra versus metabolite quantification
- 26 February 2007
- journal article
- Published by Wiley in NMR in Biomedicine
- Vol. 20 (8), 763-770
- https://doi.org/10.1002/nbm.1147
Abstract
1H MRS is an attractive choice for non‐invasively diagnosing brain tumours. Many studies have been performed to create an objective decision support system, but there is not yet a consensus as to the best techniques of MRS acquisition or data processing to be used for optimum classification. In this study, we investigate whether LCModel analysis of short‐TE (30 ms), single‐voxel tumour spectra provide a better input for classification than the use of the original spectra. A total of 145 histologically diagnosed brain tumour spectra were acquired [14 astrocytoma grade II (AS2), 15 astrocytoma grade III (AS3), 42 glioblastoma (GBM), 41 metastases (MET) and 33 meningioma (MNG)], and linear discriminant analyses (LDA) were performed on the LCModel analysis of the spectra and the original spectra. The results consistently suggest improvement in classification when the LCModel concentrations are used. LDA of AS2, MNG and high‐grade tumours (HG, comprising GBM and MET) correctly classified 94% using the LCModel dataset compared with 93% using the spectral dataset. The inclusion of AS3 reduced the accuracy to 82% and 78% for LCModel analysis and the original spectra, respectively, and further separating HG into GBM and MET gave 70% compared with 60%. Generally MNG spectra have profiles that are visually distinct from those of the other tumour types, but the classification accuracy was typically about 80%, with MNG with substantial lipid/macromolecule signals being classified as HG. Omission of the lipid/macromolecule concentrations in the LCModel dataset provided an improvement in classification of MNG (91% compared with 76%). In conclusion, there appears to be an advantage to performing pattern recognition on the quantitative analysis of tumour spectra rather than using the whole spectra. However, the results suggest that a two‐step LDA process may help in classifying the five tumour groups to provide optimum classification of MNG with high lipid/macromolecule contributions which maybe misclassified as HG. Copyright © 2007 John Wiley & Sons, Ltd.Keywords
This publication has 23 references indexed in Scilit:
- Combination of feature-reduced MR spectroscopic and MR imaging data for improved brain tumor classificationNMR in Biomedicine, 2005
- Classification of brain tumours using short echo time 1H MR spectraJournal of Magnetic Resonance, 2004
- Brain tumor classification based on long echo proton MRS signalsArtificial Intelligence in Medicine, 2004
- Determination of histopathological tumor grade in neuroepithelial brain tumors by using spectral pattern analysis of in vivo spectroscopic dataJournal of Neurosurgery, 2003
- Automated classification of short echo time in in vivo 1H brain tumor spectra: A multicenter studyMagnetic Resonance in Medicine, 2002
- A new approach for analyzing proton magnetic resonance spectroscopic images of brain tumors: nosologic imagesNature Medicine, 2000
- Automated classification of human brain tumours by neural network analysis using in vivo H magnetic resonance spectroscopic metabolite phenotypesNeuroReport, 1996
- Accurate, noninvasive diagnosis of human brain tumors by using proton magnetic resonance spectroscopyNature Medicine, 1996
- In Vivo proton MR spectroscopy of human gliomas: definition of metabolic coordinates for multi‐dimensional classificationMagnetic Resonance in Medicine, 1995
- A multiparametric data analysis showing the potential of localized proton MR spectroscopy of the brain in the metabolic characterization of neurological diseasesJournal of the Neurological Sciences, 1993