Machine learning methods for fluorescence lifetime imaging (FLIM) based automated detection of early stage oral cancer and dysplasia (Conference Presentation)

Abstract

Despite of the ease accessibility of the oral cavity, only ~30% of oral cancer patients are diagnosed at early stages. Some of the factors that contribute to this low rate of early detection are: asymptomatic oral cancer lesions, similarity to benign lesions, and sampling error during biopsy procedures. Progression of oral cancer is accompanied by alterations in the intrinsic fluorescence properties of the oral tissue, making fluorescence lifetime imaging (FLIM) suitable for the diagnosis of oral cancer. In this study, in vivo human oral lesions from 70 patients were imaged using a multispectral FLIM endoscopy system. The collected database consisted of 50 benign lesions, and 20 dysplastic and early stage cancerous lesions, as determined by histopathological diagnosis. For each pixel, three fluorescence decays were collected corresponding to three emission bands (390 nm, 450 nm, 500 nm), and analyzed using a biexponential decay model. Selected parameters of this fitting algorithm along with the normalized intensities at each emission band were used as features for a quadratic discriminant analysis (QDA) classifier. The classification performance was estimated using a 10 fold cross-validation approach, resulting on levels of sensitivity and specificity >85%, and an ROC AUC of 0.9 for detecting dysplastic and cancerous oral lesions from benign lesions. These results demonstrate the potential of endogenous FLIM endoscopy for automated early detection of oral cancer.