To detect the small lesion and identify the margin of observable tumors, in vivo, the potential of light-induced fluorescence (LIF) spectroscopic imaging was evaluated to improve the accuracy of conventional white light endoscopy. In this study, further investigations were carried out using a general multivariate spectral classification algorithm....[ Read more ]

To detect the small lesion and identify the margin of observable tumors, in vivo, the potential of light-induced fluorescence (LIF) spectroscopic imaging was evaluated to improve the accuracy of conventional white light endoscopy. In this study, further investigations were carried out using a general multivariate spectral classification algorithm.

A conventional endoscopic system with a multiple channel spectrometer was used to measure the autofluorescence of nasopharyngeal tissue in vivo. Classification was based on the spectral difference between the carcinoma and normal tissue. A sophisticated algorithm based on Principal Component Analysis (PCA) was developed to differentiate between the nasopharyngeal carcinoma from the normal tissue. Firstly, preprocessing was done to reduce noise and to calibrate the different measurement distances and geometry about which there was no prior information. Secondly, processing using Principal Component Analysis was done to effectively reduce the variable dimensions while maintaining useful information for analysis. Thirdly, various post-processing techniques were investigated and the classification performance was compared. Algorithms using the ratio of auto-fluorescence spectra intensity at two wavelength and three wavelength bands were applied for diagnostic performance comparisons[1]. In addition, the robustness of the PCA based algorithm in noisy environment was investigated. Finally, the possibility of applying of the optical processing techniques based on the PCA-based algorithm using optical filter in real time diagnosis were analyzed.

Without prior knowledge of tissue optics and blood absorption characteristics, the application of the PCA-based method gives significantly better diagnostic performance than the previous two-wavelength and three-wavelength algorithms. Based on the entire spectra, the two-wavelength ratio algorithm gives a sensitivity of 88% and a specificity of 92% in respect to the detection of nasopharyngeal carcinomas. A sensitivity of 92% and specificity of 96% are achieved by the PCA-based algorithm. For the three-wavelength algorithm, a sensitivity of 88% and specificity of 95% are achieved. Also, the PCA-based algorithm shows less than 1% degradation in specificity with 5% additional random noise on the collected spectra. In addition, the investigation into the application of the algorithm using optical filters for real-time diagnosis showed that the diagnostic performance degradation is insignificant compared to the theoretical results.

In conclusion, the PCA based multivariate statistical algorithm is an efficient way to improve the spectral classification performance of nasospharyngeal carcinoma. The results of which can be read by non-experts. The algorithm is also robust in respect to spectra collected in a noisy environment and it is feasible to implement the algorithm using optical processing techniques for real time diagnosis.