The Role of Neural Networks in Improving the Accuracy of MR Spectroscopy for the Diagnosis of Head and Neck Squamous Cell Carcinoma
Ronald J. Gerstlea,
Stephen R. Aylwarda,
Sharon Kromhout-Schiroa and
Suresh K. Mukherji
,a
a From the Departments of Radiology (R.J.G., S.R.A., S.K-S., S.K.M.) and Surgery (S.K.M.), University of North Carolina School of Medicine, Chapel Hill, NC.
View larger version (7K):
[in a new window]
|
FIG 1. Examples 1H-MR spectra of normal muscle (A) and SCCA (B). Peaks measured in this study were olefinic acids (5.3 ppm), inositol (3.5 ppm), taurine (3.4ppm), Cho (3.2 ppm), Cr (3.0 ppm), sialic acid (2.2 ppm), and methyl (0.9 ppm). The height of the methylene (1.3 ppm) peak was used as an internal standard. (curved arrow indicates Cho resonance) nodes. Through training, a collection of nodes can adjust their interconnections so as to make complex decisions in consideration of a large number of inputs. This capability makes this ideally suited for the analysis of spectral resonanaces.
| |
View larger version (7K):
[in a new window]
|
FIG 2. A 2D example of LDA. A, Two features (F1 and F2 axes) describe two classes (black and white dots). B, Using the assumption of a normal distribution for each of the classes. LDA computes a linear boundary (dotted line) in the 2D space to separate the two classes completely
| |
View larger version (7K):
[in a new window]
|
FIG 3. Schematic of a node, the fundamental unit of an NN. A node weights its input to make a simple yes/no decision. An NN is trained by adjusting the weights used by its nodes. Through training, a collection of nodes can adjust their interconnections so as to make complex decisions in consideration of a large number of inputs. This capability makes this ideally suited for the analysis of spectral resonanaces
| |
View larger version (14K):
[in a new window]
|
FIG 4. A 2D example shows the effect of the configuration of the NN on the boundary. Two features (F1 and F2 axes) describe two classes (black and white dots) (A). In this example, the boundary is not linear. Boundaries (dotted lines) and schematics of NNs that would create them are shown for networks with one (B), two (C), and three (D) hidden nodes.
| |
View larger version (14K):
[in a new window]
|
FIG 5. A schematic of a two-hidden-node NN used for the analysis of multiple peak height data. Seven peak heights for a sample serve as input to the nodes of the input layer. The computations propagate through the network, and the classification as tumor or nontumor is read from the nodes of the output layer.FIG 6. ROC curves, computed using the binormal method, for each classifier tested (NN-2=two-node NN, NN-1=one-node NN, LDA-Cho/Cr=LDA for Cho/Cr area ratio, LDA-Heights=LDA for Cho/Cr peak height ratio). TPR=true-positive rate (sensitivity); FPR=false-positive rate (1-specificity). ROC analysis demonstrates that NN analysis outperfoms linear discriminant analysis for measuring the diagnostic accuracy of 1H-MR in attempting to differentiate SCCA from muscle
| |
