Tumor margin identification and prediction of the primary tumor from brain metastases using FTIR imaging and support vector machines

Abstract

Infrared spectroscopy enables the identification of tissue types based on their inherent vibrational fingerprint without staining in a nondestructive way. Here, Fourier transform infrared microscopic images were collected from 22 brain metastasis tissue sections of bladder carcinoma, lung carcinoma, mamma carcinoma, colon carcinoma, prostate carcinoma and renal cell carcinoma. The scope of this study was to distinguish the infrared spectra of carcinoma from normal tissue and necrosis and to use the infrared spectra of carcinoma to determine the primary tumor of brain metastasis. Data processing follows procedures that have previously been developed for the analysis of Raman images of these samples and includes the unmixing algorithm N-FINDR, segmentation by k-means clustering, and classification by support vector machines (SVMs). Upon comparison with the subsequent hematoxylin and eosin stained tissue sections of training specimens, correct classification rates of the first level SVM were 98.8% for brain tissue, 98.4% for necrosis and 94.4% for carcinoma. The primary tumors were correctly predicted with an overall rate of 98.7% for FTIR images of the training dataset by a second level SVM. Finally, the two level discrimination models were applied to four independent specimens for validation. Although the classification rates are slightly reduced compared to the training specimens, the majority of the infrared spectra of the independent specimens were assigned to the correct primary tumor. The results demonstrate the capability of FTIR imaging to complement histopathological tools for brain tissue diagnosis.