Segmentation of overlapping cells obtained from pap smear test

Abstract

The pap (papanicolaou) smear test is a manual screening test used to detect potentially pre-cancerous and cancerous changes in cervical cells. The segmentation of the microscopic cervical cells obtained from pap smear test is a challenging problem. Accurate cell segmentation and counting is important to analyze the cells and to determine whether the pap smear test results of the cells sample is satisfactory. In this paper, segmentation method of the cell clumps, cell nuclei, and cell cytoplasm is developed by using image processing algorithms on the dataset provided by ISBI (International Symposium on Biomedical Imaging) 2014 and 2015. Iterative Region based Otsu (IRO) thresholding method is used for the cell clumps segmentation. For the cell nuclei segmentation, Maximally Stable Extremal Regions (MSER) method is used. For the cell cytoplasm segmentation, Distance Regularized Level Set Evolution (DRLSE) method is used. The methodology is qualitatively evaluated by comparing the segmented images with the ground truth images pixel by pixel and quantitatively by counting the number of cells (objects) in the images. For the segmentation of cell clumps, the pixel based Dice Coefficient (DC) of 0.96 and 0.91 are obtained for synthetic and real cell images. The accuracy is 0.98 and 0.95 for synthetic and real cell images. For the segmentation of cell nuclei, the pixel based DC of 0.92 and 0.79 are obtained for synthetic and real cell images. For the counting of cell nuclei, the object based DC of 0.99 and 0.85 are obtained for synthetic and real cell images with less than 0.10 error rate. For the segmentation of cell cytoplasm, the pixel based DC of 0.84 and 0.73 are obtained for synthetic and real cell images.