THE FLUORESCENCE SPECTROSCOPY OF MEDICAL SAMPLES AND THE ANALYSIS OF SPECTRA USING MULTI-CURVE RESOLUTION METHODS

Jurgita Stankevičiūtė1, Vilmantas Gėgžna2, 3, Aurelija Vaitkuvienė3

1 Faculty of Physics, Vilnius University, Lithuania

2 Faculty of Natural Sciences, Vilnius University, Lithuania

3 Institute of Applied Research, Vilnius University, Lithuania

In the medical practice the evaluation of indistinct premalignant cervical lesions is based on examination of cytological smear samples. Although the examination is performed in laboratory the accuracy of it is influenced by the human factor. The sensitivity of such tests ranges from 50% to 80%, while specificity – from 70% to 90% [1]. Methods of fluorescence spectroscopy have huge potential as an alternative to non-invasive clinical detection of changes in cells. Until now only preliminary researches of cervical smear auto-fluorescence has been made comparing low-grade squamous intraepithelial lesions (LSIL) group to one without pathological changes (Normal) and the number of samples investigated was very low. Thus the goal established for this work is to investigate a larger quantity of Normal and LSIL group samples for purposes of medical diagnostics using autofluorescence based measurements and spectra decomposition methods.

The tasks of this research were to:

1) examine investigate a larger quantity of fluorescence spectra; 2) improve data analysis system involving spectra decomposition algorithms called principal component analysis (PCA), mutual information least-dependent component analysis (MILCA), multi curve resolution – alternating least squares(MCR-ALS); 3) calculate the accuracy of cervical Normal and LSIL group autofluorescence-based identification in comparison to cytology-based evaluation.

Methods. 37 samples of Normal, 33 samples of LSIL group were investigated registering 5 spectra per undried specimen. Spectra were preprocessed, divided to 2 subsets (for classification model creation and for model validation). The components were extracted combining PCA, MILCA, MCR-ALS algorithms and statistical Mann-Whitney and receiver operating characteristic (ROC) methodologies.

The best results of Normal and LSIL group classification were received using PCA method: the Mann-Whitney test showed statistically reliable difference (p<0.05), the accuracy measure called AUC (area under the ROC curve) = $0.64(0.56\div0.72)$, sensitivity = 58.8% $(51.2\div66.0)$, specificity = 62.1% $(54.6\div69.2)$, positive predictive value =58.1%$(50.6\div65.4)$, negative predictive value = 62.8% $(55.2\div69.8)$, overall accuracy=60.6% $(53.0\div67.7)$. All measures indicated with their 95% confidence intervals in brackets.

Resume. Mann-Whitney test indicates statistically reliable differences between spectra of cytological Normal and LSIL groups. Unfortunately these differences are not sufficient to fulfill requirements for medical diagnostics, i.e. AUC > 0.75 [2].It should be mentioned that the "golden standard" in medicine is histological evaluation. Due to medical rules, the histological evaluation is not available for both cytological Norma and LSIL cases. Thus the next step of analysis should bethe development of auto-fluorescence based classification algorithm to distinguish between cytological Normal/LSIL group and histologically evaluatedgroup of cytological HSIL (high-grade squamous intraepithelial lesions) cases.


[1] A. Vaitkuvienė, V. Gėgžna, R. Kurtinaitienė, and J. V. Vaitkus, Cervical smear photodiagnosis by fluorescence, Photomed and Laser Surg, 30, 268-274, 2012.

[2] J. Fan, S. Upadhye, and A. Worster, Understanding receiver operating characteristic (ROC) curves. CJEM8(1), 19-20, 2006.