The role of neural networks in improving the accuracy of MR spectroscopy for the diagnosis of head and neck squamous cell carcinoma

Citation
Rj. Gerstle et al., The role of neural networks in improving the accuracy of MR spectroscopy for the diagnosis of head and neck squamous cell carcinoma, AM J NEUROR, 21(6), 2000, pp. 1133-1138
Citations number
27
Categorie Soggetti
Radiology ,Nuclear Medicine & Imaging","Neurosciences & Behavoir
Journal title
AMERICAN JOURNAL OF NEURORADIOLOGY
ISSN journal
01956108 → ACNP
Volume
21
Issue
6
Year of publication
2000
Pages
1133 - 1138
Database
ISI
SICI code
0195-6108(200006/07)21:6<1133:TRONNI>2.0.ZU;2-Y
Abstract
BACKGROUND AND PURPOSE: MR Spectroscopy (MRS) has the unique ability to ana lyze tissue at the molecular level noninvasively. The purpose of this study was to determine if peak heights revealed by proton MRS (H-1-MRS) signals showed that neural networks (NN) provided better accuracy than linear discr iminant analysis (LDA) in differentiating head and neck squamous cell carci noma (SCCA) from muscle METHODS: In vitro 11-T H-1-MR spectra were obtained on SCCA tissue samples (n = 16) and muscle (n = 12), The peak heights at seven metabolite resonanc es were measured: olefinic acids at 5.3 ppm, inositol at 3.5 ppm, taurine a t 3.4 ppm, choline (Cho) at 3.2 ppm, creatine (Cr) at 3.0 ppm, sialic acid at 2.2 ppm, and methyl at 0.9 ppm. Using leave-one-out experimental design and receiver operating characteristic curve analysis, the ability of NN and LDA classifiers to distinguish SCCA from muscle were compared (given equal weighting of false-negative and false-positive errors). These classifiers were also compared with an existing method that forms a diagnosis by using LDA of the Cho/Cr peak area ratio. RESULTS: NN classifiers, which were identified using height data, achieved better sensitivity and specificity rates in distinguishing SCAA from muscle than did LDA using height or area data. Sensitivity/specificity for the NN analysis of the seven metabolite peak heights were 87.5% and 83.3%, respec tively, for a one-hidden-node network and 81.2% and 91.7%, respectively, fo r a two-hidden-node network, Additional nodes did not improve accuracy, The sensitivity and specificity were 81.2% and 50%, respectively, for LDA of t he seven peak heights, and 68% and 83%, respectively, for LDA of the Cho/Cr peak area ratio, CONCLUSION: NN classifiers with peak height data were superior to LDA of th e peak heights and LDA of the Cho/Cr peak area ratio for differentiating SC CA from normal muscle, These results show neural network analysis cain impr ove the diagnostic accuracy of H-1-MRS in differentiating muscle from malig nant tissue. Further studies are necessary to confirm our initial findings.