AbstractsMathematics

Methods for cytological image analysis

by Μαρίνα Πλησίτη




Institution: University of Ioannina; Πανεπιστήμιο Ιωαννίνων
Department:
Year: 2012
Keywords: Κυτταρολογικές εικόνες τεστ Παπ; Κατηγοριοποίηση κυττάρων; Ανίχνευση πυρήνων κυττάρων; Κατάτμηση πυρήνων κυττάρων; Κατάτμηση επικαλυπτόμενων πυρήνων κυττάρων; Μαθηματική μορφολογία – Αλγόριθμος υδροκριτών; Παραμορφώσιμα μοντέλα; Επιλογή Χαρακτηριστικών; Pap smear images; Cell classification; Cell nuclei detection; Cell nuclei segmentation; Overlapping cell nuclei segmentation; Mathematical morphology -Watersheds; Deformable models; Feature Selection
Record ID: 1153063
Full text PDF: http://hdl.handle.net/10442/hedi/27134


Abstract

This thesis is focused on the development of image segmentation methods in combination with classification techniques for efficiently addressing the specific problems presented in Pap smear images. The several steps that must be followed for the effective analysis of such images in an automated manner are described. The goal is to achieve accurate identification of the regions of interest, and as a result to obtain reliable conclusions about the contents of the Pap smear. The first issue that we have successfully addressed is the correct detection of the locations of the nuclei in images containing both isolated cells and cell clusters.. In this scope, techniques based on mathematical morphology are developed. The elimination of the undesirable findings is achieved in two steps: the application of a distance dependent rule on the resulted centroids and the application of classification algorithms, employing features of the neighborhood of the candidate nuclei. Furthermore, we have developed an automated method for the boundary determination of cells nuclei based on the watershed transform. For the elimination of false positive findings, features characterizing the shape, the texture and the image intensity are extracted from the candidate nuclei regions, which are used as input in a classification step, performed to determine the true nuclei. These features are also tested for their discriminative ability. Concerning the separation of partially overlapped nuclei, we have developed an automated method which is based on training a physically based deformable model. A deformable model whose behavior is driven by physical principles is trained on images containing single nuclei, and attributes of the shapes of the nuclei are expressed in terms of modal analysis. Based on the estimated modal distribution and driven by the image characteristics, we developed a framework, to detect and describe the unknown nuclei boundaries in images containing two overlapping nuclei. The problem of the estimation of an accurate nucleus boundary in the overlapping areas is successfully addressed with the use of appropriate weight parameters that control the contribution of the image force in the total energy of the deformable model. Moreover, we have investigated the case of the successful classification of cells in normal and abnormal categories based on features extracted exclusively from the nucleus area and ignoring the contingent cytoplasm features. We examined the ability of non-linear dimensionality reduction schemes to produce accurate representation of the features manifold, along with the definition of an efficient feature subset, and their influence on the classification performance. Η παρούσα διατριβή εστιάζεται στην ανάπτυξη αυτόματων μεθόδων κατάτμησης κυτταρολογικών εικόνων, οι οποίες σε συνδυασμό με τεχνικές κατηγοριοποίησης, αντιμετωπίζουν διεξοδικά τα προβλήματα που υπάρχουν στις μικροσκοπικές εικόνες από τεστ Παπ. Σε αυτές τις εικόνες, τα αντικείμενα ενδιαφέροντος είναι οι πυρήνες των κυττάρων, οι οποίοι…