Independent Component Analysis for Optical Diagnosis of Cancer

We report here the use of the theory of independent component analysis (ICA), a recently developed technique of data representation, for development of an algorithm for diagnosis of early stage cancer of human oral cavity. The algorithm has been developed using spectral data acquired in a clinical in-vivo laser-induced fluorescence (LIF) spectroscopic study conducted on normal volunteers and patients being screened for cancer of oral cavity. The diagnostic efficacy of the algorithm has been compared with that based on classical principal component analysis (PCA), most widely employed to reduce the dimensionality of the multidimensional spectral. The ICA based algorithm provided significantly improved diagnostic performance as compared to the PCA based algorithm in discriminating the cancerous tissue sites of the oral cancer patients from the healthy squamous tissue sites of normal volunteers as well as the uninvolved tissue sites of the oral cavity of the patients with cancer. While the sensitivity and specificity values provided by the ICA based algorithm were 83% and 95% towards cancer for the training set data based on leave-one-out cross validation and 88% and 99% towards cancer for the independent validation set data, the linear PCA based algorithm provided a sensitivity and specificity of 69% and 90% respectively towards cancer for the training set data and a sensitivity and specificity of 76% and 91% respectively towards cancer for the validation set data.