Segmentation And Classification Of Optical Microscopic Cervical Cell Images

Optical microscopic cell image is one kind of the widely used medical images in clinical diagnostic applications. This dissertation focuses on optical microscopic cervical cell image segmentation and classification, which are key techniques in the application of computer-assisted cervical cytology screening. The main efforts are dedicated to solve several key problems in the existing methods in the area of cervical cell image segmentation and classification. The main contributions are summarized as follows:In the area of cell image segmentation, to solve the problem that the existing edge-based single-cell image segmentation methods have broken edges and can not guarantee closed contours, Chapter 2 first develops a variational and Partial Differential Equation(PDE)-based method to cope with the edge linking problem, and then proposes a novel closed contour extraction method based on an ending point restrained Gradient Vector Flow field. This method is further extended to extract the cell contours in single-cell images. Experiments on real cervical cell images with obscure edges validate the proposed single-cell image segmentation method.To solve the problem that classical cell image segmentation methods can not effectively utilize the color information in cell images, Chapter 3 presents a color difference vector field to model the colors of cell images. Based on analyzing the characteristics of the cell edges in the color difference vector field, a sequential match method is developed to make segmentation of color single-cell images. The experiments have been performed on real color cell images, which were acquired by an automatic microscopic image acquisition system. The segmentation results show the effectiveness of the proposed method.To solve the problem that the existing cell image segmentation methods are not able to effectively address the overlapping cytoplasm segmentation, Chapter 4 proposes a novel overlapping cells segmentation method, which is based on dynamic sparse contour searching and Gradient Vector Flow(GVF) Snake model. This method first approximately represents the cell contour as a set of sparse contour points, which can be further partitioned into two parts: the strong contour points and the weak contour points. Then, it considers the cell contour extraction as a contour point locating problem and develops a dynamic sparse contour searching(DSCS) algorithm to locate the cell contour points. The contour points located by the DSCS algorithm are very close to the true cell boundary. Finally, using the located contour points, the GVF Snake model is employed to extract the accurate cell contour. Experiments have been performed on two cervical smear image datasets(including the publicly available Herlev dataset). The high accuracy of the cell contour extraction result validates the effectiveness of the proposed method.To solve the problem that classical segmentation methods are prone to lose the cell nuclei with low intensity contrast, Chapter 5 proposes a course-to-fine nuclei segmentation method. In this method, the cell nuclei are first enhanced by a nuclei enhancement method combing the intensity and the color information of the image. Then, the morphological reconstruction is employed to extract the rough segmented regions of the nuclei. Finally, in each rough extracted region, the GVF Snake model is performed to make fine segmentation of the nuclei. In addition, in order to address the overlapping cytoplasm segmentation in multiple cells image, which is a challenging topic in medical image processing, Chapter 5 proposes a cell contour supporting region to extend the overlapping cells segmentation method proposed in Chapter 4 to cope with multiple cells image. Various multiple cells images have been tested to show the validity and effectiveness of the proposed method.In the area of cell image feature extraction and classification, to solve the drawbacks of classical cell features, Chapter 6 first proposes 7 cell features that do not depend on accurate segmentation of the cell cytoplasm. Then, based on the aforementioned color difference vector filed, another 6 cell features are introduced. Using these cell features and 9 classical cell features, a new cervical cell image feature set is constructed. Finally, in order to classify the cervical cells based on this new feature set, Chapter 6 proposes a two-stage classification method, which is inspired by the idea of the decision tree classification method. On contrast to the direct classification method, the proposed two-stage classification method is able to achieve higher correct classification rate. Classification experiments on two cervical smear image datasets(including the publicly available Herlev dataset) validate the effectiveness of the proposed method.