In Vivo Cervical Precancer Classification Through Multifractal Analysis of Spectral Fluctuations in Intrinsic Fluorescence Spectra.

Abstract

Spectral fluctuations in fluorescence spectroscopy, often ignored as noise, contain significant information about the fluorophore microenvironments. We present a discrete wavelet transform (DWT)-based technique to extract spectral fluctuations from the intrinsic fluorescence signals and utilize them to classify normal and precancerous patients. The fluctuations are extracted by applying the inverse DWT after zeroing the approximation and noisy detail coefficients. Multifractal detrended fluctuation analysis revealed stronger multifractality for precancer signals manifested in the singularity spectrum. The Hurst exponent ( $H$ ) and the Hausdorff dimension $\left(\Delta \alpha \right)$ clearly distinguish two groups. Random Forest classification of generalized Hurst and Holder exponents achieves 96% sensitivity, specificity, and accuracy with an AUC of 0.98. This indicates that the spectral fluctuations derived from the intrinsic fluorescence data capture the subtle, distinctive features, resulting in better classification between the two grades. Further, a comparison among various mother wavelet functions reveals the best performance for the "bior2.4" wavelet.