Identification and segmentation of pre-cancerous cervix lesions for automated diagnosis of Cervical Neoplasia
MetadataAfficher la notice complète
The design and implementation of an automated diagnostic tool for Cervical Intraepithelial Neoplasia (CIN), a detectable and treatable precursor pathology of uterine cervical cancer, is being considered. The pathological configurations and implications of CIN are researched, and the shortcomings of current screening and diagnostic procedures for CIN are analyzed. An image processing framework is then developed to automate the detection and characterization of abnormalities due to CIN in cervicographic and colposcopic images of the cervix. The problem of translating a digital image of the cervix into clinical decisions is broken down into six modules – 1) cervix Region of Interest (ROI) segmentation, 2) Specular Reflection (SR) removal, 3) AcetoWhite (AW), Columnar Epithelium (CE) and Squamous Epithelium (SE) segmentation, 4) classification of AW sections, 5) segmentation of mosaicism and punctations, and 6) assessment of disease severity. Algorithms based on mathematical morphology, and clustering based on Gaussian mixture modeling in a joint color and geometric feature space, are used to segment macro regions, like the cervix ROI, SR, AW, CE and SE regions. The performance of the cervix ROI segmentation and AW segmentation algorithms are evaluated by comparison with ground truth segmentations of expert graders. The classification of vascular abnormalities, such as mosaicism and punctations, is modeled as a texture classification problem, and a solution is attempted by characterizing the neighborhood gray-tone dependences and co-occurrence statistics of the textures. Texture content-based retrieval of similar regions, using a seed region, is also presented as an alternative to unsupervised classification of regions. The precise segmentation of mosaicism and punctations from textured regions using Gaussian modulated rotating structuring elements and Gaussian matched filtering are also demonstrated. A thorough analysis of the inter-grader variability between experts in resolving the AW region and grading images containing vascular abnormalities is also presented to compare the performance of the automated segmentation methods presented in this research. The most significant contribution of this work is the development of a unified framework for a fully automated diagnostic system for CIN, and it is the first such attempt addressing the feasibility of such a system. In the past, research has been devoted to the detection of one or more specific abnormalities (AW, mosaicism or punctations) but not of all of them. The model presented in this research provides a sequential framework for translating digital images of the cervix into a complete diagnostic tool, with minimal human intervention. In its current form the research presented in this work may be used to aid physicians to locate abnormalities due to CIN and decide the best areas for a biopsy. The methodologies and framework presented in this work promise deployment of a fully automated diagnostic system for CIN, especially suitable for resource-poor regions.