Classification of normal and precancerous cervical tissues using nonlinear maximum representation and discrimination features (NMRDF) on polarized reflectance data

Seema Devi; Asha Agarwal; Kiran Pandey; Asima Pradhan

doi:10.1117/12.2080682

Translator Disclaimer

5 March 2015 Classification of normal and precancerous cervical tissues using nonlinear maximum representation and discrimination features (NMRDF) on polarized reflectance data

Seema Devi, Asha Agarwal, Kiran Pandey, Asima Pradhan

Author Affiliations +

Proceedings Volume 9318, Optical Biopsy XIII: Toward Real-Time Spectroscopic Imaging and Diagnosis; 93180R (2015) https://doi.org/10.1117/12.2080682
Event: SPIE BiOS, 2015, San Francisco, California, United States

Abstract

Reflectance spectroscopy contains information of scatterers and absorbers present inside biological tissues and has been successfully used to diagnose disease. Success of any diagnostic tool depends upon the potential of statistical algorithm to extract appropriate diagnostic features from the measured optical data. In our recent study, we have used the potential of the classification algorithm, Nonlinear Maximum Representation and Discrimination Features (NMRDF) to extract important diagnostic features from reflectance spectra of normal and dysplastic human cervical tissue. This NMRDF algorithm uses the higher order correlation information in the input data, which helps to represent the asymmetrically distributed data and provides the closed form solution of the nonlinear transform for maximum discrimination. We have recorded unpolarized, co and cross-polarized reflectance spectra from 350nm to 650nm, illuminating the human cervical tissue epithelium with white light source. A total of 139 samples were divided into training and validation data sets. The input parameters were optimized using training data sets to extract the appropriate nonlinear features from the input reflectance spectra. These extracted nonlinear features are used as input for nearest mean classifier to calculate the sensitivity and specificity for both training as well as validation data sets. We have observed that co-polarized components provide maximum sensitivity and specificity compared to cross-polarized components and unpolarized data. This is expected since co-polarized light provides subsurface information while cross-polarized and unpolarized data mask the vital epithelial information through high diffuse scattering.

Citation Download Citation

Seema Devi, Asha Agarwal, Kiran Pandey, and Asima Pradhan "Classification of normal and precancerous cervical tissues using nonlinear maximum representation and discrimination features (NMRDF) on polarized reflectance data", Proc. SPIE 9318, Optical Biopsy XIII: Toward Real-Time Spectroscopic Imaging and Diagnosis, 93180R (5 March 2015); https://doi.org/10.1117/12.2080682

ACCESS THE FULL ARTICLE