Automated nuclei and cytoplasm segmentation of multiple overlapping cervical cells

Abstract

Pap smear test is a screening technique to detect pre-cancerous and cancerous changes in a sample of cells collected from the cervix and deposited onto a glass slide for microscopic examination. Automated screening technique attempts to improve the performance of sensitivity and specificity by detecting, segmenting and classifying the cervical cells on a slide, compared to the tedious, time-consuming and error-prone manual screening procedure. Accurate segmentation of overlapping cells is the critical step in the automated screening system that produces a larger sample of cells per specimen (compared to the currently prevailing segmentation of isolated cells) to allow a more robust subsequent analysis process. However, current methods cannot undertake such a complete segmentation, which is still an open issue for medical image analysis community. This application is notoriously difficult because of the large number of cells in a single image, the overlap among these cells, the poor contrast of the cell cytoplasm, and the presence of mucus, blood and inflammatory cells. In this thesis, we present two new algorithms for segmenting multiple overlapping visual objects using a methodology based on a joint optimisation of several level set functions. These methods are applied to the segmentation of individual cytoplasm and nucleus from clumps of overlapping cervical cells. We start from discussing the delineation of each individual nuclei and cytoplasm from a clump of overlapping cervical cells using a joint level set framework with unary and pairwise constraints. Our approach initially performs a scene segmentation to highlight both free-lying cells, cell clumps and their nucleus. Then cell segmentation is performed using a joint optimisation of several level set functions on all of the detected nuclei and cytoplasm pairs, where each of these functions represents one of the cervical cells present in the image. This optimisation is constrained by the length and area of each cell, a prior on cell shape, the amount of cell overlap and the expected gray values within the overlapping regions. Experiments present promising quantitative segmentation results on a database comprising 18 synthetic overlapping cell images constructed from real freelying cervical cell images. We also perform a qualitative assessment based on a small number of complete fields of view containing multiple cells and clumps of cells. Finally, using this database, we also show competitive nuclei detection results compared to the state of the art. The second contribution attempts to decrease the false negative rate (FNR) and increase the true positive rate (TPR) in individual cytoplasm segmentation. Our solution initially generates an initial estimation for each cell that roughly delineates the cytoplasm boundary, according to its geometric feature. A shape prior that represents the chance of each point to be the cell boundary pixel is formed according to its distance from the initial estimation centroid. Complete cell segmentation is performed through a joint level set optimisation. This optimisation is constrained by a unary term of a cell boundary length and the binary term computed from the area of overlapping region. We evaluate the methodology on a larger database of images generated by synthetically overlapping real cervical cells, where the number of cells within a clump is varied from 2 to 10 and the overlap coefficient between pairs of cells from 0 to 0.5. We also show the representative visual results of our approach on a database of 16 EDF cytology images and synthetic images. In brief, our theoretic and experimental findings show these two joint level set optimisation based methods can be exploited in multiple overlapping cervical cell segmentation. As the early attempt, the first method provides an acceptable performance in the segmentation for overlapping cytoplasms. The second work makes this application more close to the practical automated screening with a more robust segmentation capability. We have confidence in the two algorithms that can be adapted to improving the performance of an automated screening system for cervical cancer in the future.

Date of Award	14 Feb 2014
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Qing LI (Supervisor) & Wenyin LIU (Supervisor)