Nucleus Region Segmentation Towards Cervical Cancer Screening Using AGMC-TU Pap-Smear Dataset
Mrinal Kanti Bhowmik, Sourav Dey Roy, Niharika Nath, Abhijit Datta · 2018
Considering global health impact, cervical cancer stands in the second position after breast cancer. Detection of any abnormality at an early stage can lead to complete cure of the disease. Higher mortality rate in rural areas symbolize lack of awareness, poor medical facilities, lack of resources and lack of efficient screening programmes. Hence to tackle these issues, automatic computer-aided diagnosis of cervical cancer is needed which should be observer-independent and less time consuming. Abnormal Pap-Smear cells are mainly distinguishable from the normal ones based on their nucleus shapes. However due to the variation in size, shape and intensity overlapping of the nucleus area with the surrounding normal cells, accurate extraction of nucleus portion is a very challenging task. In this paper, effectiveness of some state-of-the art segmentation methods for nucleus region extraction and classification of normal and abnormal cells is highlighted depending on accuracy and quantification of segmentation output. Experiment has been conducted in our own created AGMC-TU Pap-Smear dataset. In these experiments, classification is carried out based on 12 shape features extracted from the segmented nucleus region. The classification accuracy obtained with SVM-Linear (SVM-L) classifier is 92.83% based on combination of all the extracted feature set whereas classification based on discriminative feature set is 97.65% and increases the accuracy rate by almost 5% using most effective segmentation method (i.e. FCM). The Area Under the Curve (AUC) using the outer performed classifier i.e. (SVM-L) based on combined feature set and discriminative feature set are 0.93 and 0.96 respectively for FCM segmentation method. Both accuracy and AUC reveals that accurate segmentation of nucleus region from the whole Pap-Smear cell increases the accuracy rate of classification and hence indicating its effectiveness for predicting the abnormal Pap-Smear cells.