Development of a multivariate statistical algorithm to analyze human cervical tissue fluorescence spectra acquired in vivo.

Abstract

BACKGROUND AND OBJECTIVE: A general multivariate statistical algorithm has been developed to analyze the diagnostic content of cervical tissue fluorescence spectra acquired in vivo. MATERIALS AND METHODS: The primary steps of the algorithm are to: (1) preprocess the data to reduce inter-patient and intra-patient variation of tissue spectra within a diagnostic category, without a priori information, (2) dimensionally reduce the preprocessed fluorescence emission spectrum with minimal information loss and use it to select the minimum number of the original emission variables of the fluorescence spectrum required to achieve classification with negligible decrease in predictive ability, and (3) assign a posterior probability to the diagnosis of each sample, so that samples with relative uncertain diagnosis can be reevaluated by a clinician. The algorithm was tested retrospectively and prospectively on cervical tissue spectra acquired from 476 sites from 92 patients at 337 nm excitation. RESULTS: The algorithm based on the entire fluorescence spectrum differentiates squamous intraepithelial lesions (SILs) from normal squamous epithelia and inflammation with an average sensitivity and specificity of 88% +/- 1.4 and 70% +/- 1, respectively. The average sensitivity and specificity of the identical algorithm based on intensity selected at only two emission wavelengths is 88% +/- 1.4 and 71% +/- 1.4, respectively. CONCLUSION: The multivariate statistical algorithm based on both types of spectral inputs at 337 nm excitation has a similar sensitivity and significantly improved specificity relative to colposcopy in expert hands.