Development and Evaluation of a Multi-Modal Optical Imaging System for Early Identification of Oral Neoplasia
Doctor of Philosophy
Over the last decade, the five-year survival rate for oral cancer has remained at only 64%. Despite easy access to the oral cavity, most patients with oral cancer are diagnosed at an advanced stage when treatment is more invasive and likely to be less successful. Imaging tools that can rapidly and accurately identify oral neoplasia could improve early detection of malignant oral lesions. This dissertation describes research to develop and evaluate a multi-modal optical imaging system with automated image processing to improve early detection of oral neoplasia. The multi-modal optical imaging system is comprised of two modalities, a high-resolution microendoscope (HRME) and a wide-field autofluorescence imager (AFI) to identify suspicious areas and to confirm whether suspicious areas contain neoplasia. A tablet-interfaced HRME with automated image analysis was developed and characterized to improve early detection of esophageal squamous cell carcinoma which has similar histologic patterns to oral neoplasia; results showed the tablet HRME can acquire comparable images to the first generation HRME design at a fraction of the cost and size. Training and validation was performed using a previously published dataset from a study of 177 patients referred for screening or surveillance endoscopy in China. Results showed that the automated image processing could differentiate between neoplastic and non-neoplastic images with a sensitivity of 95% and 91% in an independent validation set compared with 84% and 95% achieved in the original study. Additionally, automated image processing tools were developed to analyze wide-field autofluorescence images. The diagnostic performance of this approach was compared to previous results from a pilot study of 30 patients scheduled for surgical resection of a clinically suspicious oral lesion. The automated analysis method achieved a comparable area under the receiver operating characteristic curve (AUC) to the previous results based on manual analysis (0.862 automated vs. 0.877 manual) while minimizing dependence on user input. The automated analysis algorithms for AFI and HRME were then evaluated together to analyze images acquired from a population of 100 patients scheduled for surgical resection of a clinically suspicious oral lesion. A classification algorithm based on image metrics derived from AFI and HRME was able to correctly classify 100% of sites taken from biopsies pathologically diagnosed as normal and 85% of sites taken from biopsies diagnosed as moderate/severe dysplasia or cancer. These results provide evidence that multi-modal optical imaging with automated image analysis could be a valuable diagnostic adjunct for early detection of oral neoplasia.
optical imaging; oral cancer; fluorescence