Classification of Cervical Cell Images based on Adaptive Thresholding Segmentation

This paper presents a new method of automatically screening cervical cancerous cell images. The proposed method first enhanced the cervical cell images by a morphological filtering and adaptive histogram equalization method. Then, an Experiential-Factor-Weighted Otsu Thresholding algorithm, which solves the biases of traditional Otsu thresholding method due to the overlapping of cells in images, is presented for segmentation of the cell nuclei. To extract the largest cell nuclei, the algorithm uses four features, which are area, perimeter, ratio of area and convex area, ratio of length and width of the segmented cell nuclei. Finally, to classify the cell images into normal and abnormal ones, the K-means clustering algorithm is employed on the basis of two cell nuclei features: area and mean gray level, which are extracted from the largest cell nuclei. Experiments were done on 233 cervical cell images including 49 cancerous cell images and 184 normal cell images. The experiment results validated the proposed method.