A study on diagnostic image analysis for the detection of precancerous lesions using multi-spectral digital images

This dissertation explores a diagnostic image analysis framework using multispectral digital colposcopy for real-time in vivo detection of cervical cancer. In the first part of the dissertation, the clinical feasibility of a previously developed multispectral digital colposcope (MDC) is demonstrated using a hamster cheek pouch model of carcinogenesis. Various studies on MDC applications to cervical cancer detection in human subjects are then presented. First, an automated diagnostic image analysis algorithm for cervical cancer using white light reflectance images is presented. The algorithm can identify pre-neoplastic tissue areas from an entire cervix based on intensity changes feature in the reflectance images induced by acetic acid treatment. Then, the information about tissue type is incorporated into the diagnostic image analysis framework. For this purpose, a Markov Random Field (MRF) model is adopted and the results are discussed. One of the practical difficulties of utilizing a MRF model in unpolarized white light reflectance imaging is the specular reflection problem since the effect of specular reflection extends into surrounding tissue areas. Through the use of cross polarized imaging, the effects of specular reflection reduced and the ability to segment images based on tissue types is enhanced, leading to better diagnostic performance. The diagnostic performance of polarized imaging is compared to that of unpolarized imaging. In order to assess the performance of the proposed approach, a gold standard for the entire cervical image is constructed using histopathology results from a whole cervix specimen. The results presented in this dissertation indicate that an automated diagnostic image analysis framework for early detection of cervical cancer has the potential to be clinically applied as a low cost alternative screening technique in developing countries. Advances in imaging technology as well as in image analysis algorithms will continue to reduce the cost of diagnostic imaging systems and improve the imaging and diagnostic capability, leading to an inexpensive, real-time, minimally-invasive alternative to conventional screening techniques for early detection of cervical cancer in developing countries.