Supervised Machine Learning Algorithms for Early Detection of Oral Epithelial Cancer Using Fluorescence Lifetime Imaging Microscopy
MetadataShow full item record
In this study, the clinical potential of the endogenous multispectral Fluorescence lifetime imaging microscopy (FLIM) was investigated to objectively detect oral cancer. To this end, in vivo FLIM imaging was performed on a hamster cheek pouch model with an oral epithelial cancer. The autofluorescence emissions of the hamster tissue were recorded in three different spectral bands which were determined based on the peak emission wavelength of three major fluorophores of hamster mucosal tissue: collagen (390±20 nm), NADH (452±22.5 nm), and FAD (>500 nm). Then, a total of 7 features pertaining to FLIM were extracted from each channel, providing 21 features overall. To design a classifier in a supervised approach, a training set is required, in which each pixel is labeled with one of the four groups. In this study, we utilized a total of 65 regions of interest (ROI) from the imaged cheek pouch of seven hamsters, for which the histopathological diagnosis could be correlated. The resulting database was used to train a K-Nearest-Neighborhood (KNN) algorithm aimed to detect benign from pre-malignant/malignant lesions. In addition, a Sequential Floating Forward Selection (SFFS) was applied to optimize the KNN algorithm and identify a subset of features that would maximize the classification performance. The best performance corresponded to the 3-NN algorithm with the (1/e) lifetime in the NADH channel and the normalized intensity in FAD channel as features. The overall accuracy, sensitivity and specificity for detecting pre-malignant and malignant lesions were 92.2%, 87.3%, and 94%, respectively, assessed using a cross-validation method. It has to be noted that the feature selection algorithm suggested both lifetime parameter and intensity parameter for an optimal feature set, which validates the need to utilize endogenous FLIM for the objective detection of oral cancer. At last, all data from the 65 ROIs were used to train the 3NN classifier to classify the full tissue areas. The results suggest that multispectral endogenous FLIM has a potential to screen malignant oral epithelial tissue. This technology, however, still needs to be evaluated in human patients.
Lee, Joohyung (2014). Supervised Machine Learning Algorithms for Early Detection of Oral Epithelial Cancer Using Fluorescence Lifetime Imaging Microscopy. Master's thesis, Texas A & M University. Available electronically from