Wavelet Based Classification for Cancer Diagnosis Bhadra Mani 1 , C. Raghavendra Rao 2 , P. Anantha Lakshmi 3* , Asima Pradhan 4 and Prasanta K. Panigrahi 5 Abstract — We make use of discrete wavelets to extract distinguishing features between normal and cancerous human breast tissue fluorescence spectra. These are then used in conjunction with discriminant analysis for the purpose of reliable tissue differentiation. The wavelet coefficients at different levels of decomposition, representing intensity variations at different scales, are selected as feature vectors wherein the multiresolution and localization properties of wavelets are optimally exploited for identifying features. This wavelet based approach, when combined with the sensitive polarized fluorescence data, yielded statistically reliable characterization of tissue types for diagnostic purpose. Analysis of a number of data sets belonging to both perpendicular and parallel polarized spectra have led to key distinctions between cancerous, benign and normal tissues. of computer aided diagnostics (CAD)[1]. Amongst various optical methods for tissue diagnostics, fluorescence spectroscopy [2] is one of the preferred choices, due to its sensitivity to molecular environment and biochemical changes that take place during the progress of disease. A number of key fluorophores have been identified, which show marked differences in behavior, in cancerous and normal tissues. Significant progress in cancer diagnosis using tissue fluorescence has been achieved, since the early work of Alfano and coworkers, on the tumor detection in rat tissues [3,4,5]. Many of these native fluorophores used for laser- induced fluorescence need to be excited through ultraviolet light; for example UV excited fluorescence probes NADH, tryptophan and tyrosine. In this work, we consider the flourescence due to fluorophores like flavin and porphyrin, which can be excited by visible light, thereby avoiding the potentially damaging effect of UV light. Key words — Cancer diagnostics, discrete wavelet transform, high-pass coefficients, linear discriminant analysis. A number of techniques, separable into two broad categories, viz., statistical analysis [6,7] and physical modeling [8], have been employed for the study of the spectral data. In this paper, we make use of wavelet transform [9] for extracting reliable features for tissue discrimination. It is worth noting that, in recent times, wavelet transform has emerged as a powerful tool for data analysis. The ability of the wavelets to provide multiresolution, in addition to their localization properties, makes them ideal tool for studying data sets having different structures. In the present context, wavelet analysis enables us to separate out the spectral fluctuations at various scales; in the process, pinpointing characteristic differences in the spectra of cancerous, benign and normal human tissues. Furthermore, it also leads to a much more transparent dimensional reduction of the data set, which is an added advantage when dealing with large data sets. Here, discrete wavelet transform (DWT) has been used and the finite extension of the wavelet basis is employed to effectively capture collective behavior and sharp changes and localize them. Furthermore, their effect at various scales can also be probed making use of the mathematical microscopic nature of the wavelets. I. INTRODUCTION In recent times, the rapid progress in the field of lasers, fiber optics, and mathematical modeling have led to the emergence 1 School of Physics, University of Hyderabad, Hyderabad – 500 046, India, e-mail: bhadramani@yahoo.com 2 Department of Mathematics and Statistics, School of Mathematics and Computer/Info. Sciences,University of Hyderabad, Hyderabad – 500 046, India, e-mail: An important aspect to be noted is that wavelets can be designed to have a number of useful properties, like being blind to regular structures; which can be particularly useful for the statistical modeling of the spectral fluctuations, since the random fluctuations from their correlated counterparts can be separated through these wavelets. As has been shown recently, wavelets are ideal for handling irregular data series [10,11,12]. Given the observed signal, y(t) = ƒ(t) + e(t), where ƒ(t) is the signal and e(t) the noise, Donoho and co-workers have shown in the above references the usefulness of the wavelets in crrsm@uohyd.ernet.in 3 School of Physics, University of Hyderabad, Hyderabad – 500 046, India e-mail: palsp@uohyd.ernet.in 4 Department of Physics, Indian Institute of Technology, Kanpur – 208 016, India, e-mail: asima@iitk.ac.in 5 Physical Research Laboratory, Ahmedabad – 380 009, India, e-mail: prasanta@prl.ernet.in *Corresponding author. Engineering Letters, 14:2, EL_14_2_4 (Advance online publication: 16 May 2007) ______________________________________________________________________________________