Identification of primary tumors of brain metastases by Raman imaging and support vector machines

Abstract

Vibrational spectroscopic imaging techniques are new tools for visualizing chemical components in tissue without staining. The spectroscopic signature can be used as a molecular fingerprint of pathological tissues. Fourier transform infrared imaging which is more common than Raman imaging so far has already been applied to identify the primary tumor of brain metastases. The current study introduces a two level discrimination model for Raman microspectroscopic images to distinguish normal brain, necrosis and tumor tissue, and subsequently to determine the primary tumor. 22 Specimens of normal brain tissue and brain metastasis of bladder carcinoma, lung carcinoma, mamma carcinoma, colon carcinoma, prostate carcinoma and renal cell carcinoma were snap frozen, and thin tissue sections were prepared. Raman microscopic images were collected with 785nm laser excitation at 10μm step size. Cluster analysis, vertex component analysis and principal component analysis were applied for data preprocessing. Then, data of 17 specimens were used to train the discrimination model based on support vector machines with radial basis functions kernel. The training data were discriminated with accuracy better than 99%. Finally, the discrimination model correctly predicted independent specimens. The results were superior to discrimination by partial least squares discriminant analysis and support vector machines with linear basis function kernel that were applied for comparison.