STATISTICAL TECHNIQUES FOR DIAGNOSING CIN USING FLUORESCENCE SPECTROSCOPY - SVD AND CART

A quantitative measure of intraepithelial neoplasia which can be made in vivo without tissue removal would be clinically significant in chem oprevention studies. Our group is working to develop such a technique based on fluorescence spectroscopy. Using empirically based algorithms , we have demonstrated that fluorescence is discriminating normal cerv ix from low- and high-grade cervical dysplasias with similar performan ce to colposcopy in expert hands. These measurements can be made in vi vo, in near real time, and results can be obtained without biopsy. Thi s paper describes a new method using automated analysis of fluorescenc e emission spectra to classify cervical tissue into multiple diagnosti c categories. First, data is reduced using the singular value decompos ition (SVD), yielding a set of orthogonal basis vectors. Each patient' s emission spectrum is then fit by linear least squares regression to the basis vectors, producing a set of coefficients for each patient. B ased on these coefficient values, the classification and regression tr ee (CART) method predicts the patient's classification. These results suggest that laser-induced fluorescence can be used to automatically r ecognize and differentially diagnose cervical intraepithelial neoplasi a (CIN) at colposcopy. This method of analysis is general in nature, a nd can analyze fluorescence spectra of suspected intraepithelial neopl asms from other organ sites. As a more complete understanding of the b iochemical and morphologic basis of tissue spectroscopy is developed, it may also be possible to use fluorescence spectroscopy of the cervix as a surrogate endpoint biomarker in Phase I and II chemoprevention t rials. (C) 1995 Wiley-Liss, Inc.