Multispectral Digital Colposcope for the Detection of Cervical Neoplasia
Abstract
As the fourth most common cancer among women worldwide, cervical cancer has a high mortality rate. The implementation of screening methods has decreased mortality by 50% and thus has been key to decrease the burden of the disease. The current screening system aims to detect precancerous lesions that are easily treatable. In developed areas, these screening tests include a Papanicolau or human papilomavirus (HPV) test followed by an examination for positive cases where the cervix is viewed using a colposcope to identify any regions of concern for biopsy. However, these are currently too resource-intensive for many developing countries and improvement is needed in the accuracy to detect the result of a biopsy (considered the gold standard) and reproducibility. A newly developed device, a multispectral digital colposcope (MDC), may aid in the screening of cervical cancer by taking pictures of the cervix similar to what a colposcopist views for their examination. The images are acquired using varying levels of processing and using both fluorescence and reflectance filters and thus the high-dimensionality poses a challenge for traditional modeling strategies. This study is a comparative study of different machine learning methods using features derived from these images to predict whether the patient has a high-grade biopsy. Using data on 470 patients recruited at two sites (El Paso, Texas and Vancouver, BC), we trained and tested support vector machines, k-Nearest-neighbors, random forests, lasso, and logistic regression. Accuracy was estimated using the area under the ROC curve (AUC) in independent data. Several algorithms had very good accuracy in the training set, including perfect separation using random forest, but no algorithm had an AUC greater than 0.7 in the test set. Random forests and logistic regression, with stepwise variable selection, generally had the highest AUCs in the test set (0.65). This study suggests that the technology needs further development and research before its implementation is feasible.^
Subject Area
Biostatistics
Recommended Citation
Dong, Zhe, "Multispectral Digital Colposcope for the Detection of Cervical Neoplasia" (2017). Texas Medical Center Dissertations (via ProQuest). AAI10279825.
https://digitalcommons.library.tmc.edu/dissertations/AAI10279825