Wavelet and physical parametric analysis of the AcetoWhitening optical effect: Comparative evaluation of performances in non-invasive diagnosis of cervical Neoplasia

• Published in: Proceedings of the 10th IEEE International Conference on Information Technology and Applications in Biomedicine pp. 1 - 4
• Main Authors: Margariti, V, Zervakis, M, Balas, C
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2010
• Subjects: Artificial neural networks; Discrete wavelet transforms; Optical imaging; Optical sensors
• ISBN: 1424465591; 9781424465590
• ISSN: 2168-2194; 2168-2208
• DOI: 10.1109/ITAB.2010.5687609

The quantitative assessment of the dynamic scattering characteristics associated with the Aceto-Whitening (AW) effect provides a powerful tool for the in vivo detection, grading and mapping of cervical neoplasia. In this study we perform wavelet analysis of the in vivo measured Intensity of the Backscattered Light (IBSL) vs. time curve with the purpose of improving the diagnostic performance of the method. The performance of the Wavelet analysis is comparatively evaluated with the classification performed with the aid of the area under the IBSL vs. time curve parameter (BS-AUC). A total of 371 Intensity of the Back Scattered Light (IBSL) vs. time curves corresponding to 73 tissue points obtained from 64 patients, that have been subsequently biopsied were used for developing and for evaluating the performance of the method. The data set included cases with No Evidence for Disease (NED), Inflammation, HPV, Cervical Intraepithelial Neoplasia (CIN) grade 1, 2, 3 and Cancer cases. Wavelet transformation combined with a 1-Nearest Neighbor (NN) classifier was used for the first time for classifying the IBSL vs. time curves. The clinical data set was selected from a larger sample with the selected subset to include only cases classified in full agreement between all standard methods (pap-test, colposcopy, histology). This was done for the purpose of establishing a trustful training set and for facilitating comparison between different methods. It was found that both methods display similar performance in identifying (histology confirmed) high grade curves, with sensitivities and specificities exciding 90%. The sensitivity of the wavelet method varies between 63%-85% and the specificity between 94%-97% referring to distinguishing each class from all the other classes. Another important finding of this study is that each method misclassifies different cases, which indicate that their combination, in a data fusion scheme, would further improve the diagnostic performance of the dynamic optical imaging method.