Automated diagnosis of cervical pap smears

Nupur Aggarwal, Purdue University

Abstract

This thesis focuses on the automated diagnosis of cervical smears derived from the ThinPrep Pap Test. Despite the advances in cancer therapy early detection remains one of the best defenses against cancer. The Pap test provides an effective tool for early detection of both cancerous as well as precancerous lesions in the cervix. However, diagnosis by humans based on a visual analysis of the smears' micrographs involves subjective decision-making and is known to have rather low inter-observer reproducibility. An automated diagnosis system would circumvent the problem of low inter-observer reproducibility while enabling a more exhaustive quantitative analysis of the smears. To date, to the best of our knowledge, a completely automated system does not exist. Using a combination of Multiresolution Analysis for image enhancement and an optimization approach to image segmentation, we have devised a fully automated diagnosis system. In preliminary experiments on cervical smears with clinically confirmed diagnosis our method was found to have a sensitivity of 0.8 and a specificity of 0.93. The automated diagnosis system we present comprises two phases—the training phase used to devise a two-fold classifier and a testing phase in which a new smear is diagnosed based on the previously devised classifier. Each of the phases has the following four sequential steps. In the first step—the pre-processing step—ThinPrep images are enhanced to prime them for the ensuing segmentation computation. Multiresolution Analysis (MRA) is used to filter out the noise in the image, the staining artifacts and the cytoplasmic background. To the best of our knowledge our work represents the first application of the MRA for noise reduction in the micrographs of cancer. Secondly, a segmentation method is used to computationally delineate the nuclear boundaries in the pre-processed images. The novelty of our segmentation algorithm is that it is formulated as an optimization problem. Minimizing a non-trivial cost function is shown to yield the desired segmentation. In the third step, features of diagnostic interest are extracted from the segmented nuclei. In the final step of the training phase, the feature vectors derived from samples with confirmed clinical diagnosis are used to devise a classification scheme that distinguishes benign smears from those with malignancy. In the final step of the testing phase, the feature vectors derived from the nuclei of a new smear are used, together with the classification scheme devised in the training phase, to classify the lesion underlying the new smear. The performance of our approach is demonstrated on an ensemble comprising 30 smears derived from clinically confirmed benign lesions and 25 smear samples derived from clinically confirmed malignant lesions.

Degree

M.S.I.E.

Advisors

Prabhu, Purdue University.

Subject Area

Industrial engineering

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS