Optical spectroscopy combined with neural network classification improves diagnosis of cervical precancerous lesions

In this study the value of an optical reflectance spectroscopy method in diagnosis of cervical squamous intraepithelial lesions (SIL) was assessed. Single fiber reflectance (SFR) spectroscopy was used to measure reflected light from thirty two patients undergoing standard colposcopy. Seven parameters extracted from the spectra in addition to two biographic parameters were compared in biopsy-confirmed SIL versus non-SIL. The tissue classification was done using two types of neural networks including radial basis function (RBF) and feed-forward backpropagation (FFBP) networks. The classification performance was evaluated by leave - one - out (LOO) and 5-fold (5F) cross-validation methods. Also, the minimum number of neurons required for perfect discrimination with both networks was compared. The best performance was seen using FFBP network with four neurons to achieve a perfect tissue classification. However, RBF required at least 9 neurons for a similar performance although with a shorter run time. Using FFBP the best retrospective sensitivity, specificity and area under the receiver operating characteristic (ROC) curve for discrimination of diseased versus non-diseased sites were 70%, 74% and 0.71, respectively. The results showed that the SFR spectroscopy shows promise as a non-invasive, real-time method to guide the clinician in reducing the number of unnecessary biopsies. Discrimination of SIL from other abnormalities compares favorably with that obtained by fluorescence alone and by fluorescence combined with reflectance spectroscopy while the simplicity and low cost of the presented system are dominant.