Segmentation of cervical cancer CT-scan images using K-nearest neighbors method

Abstract

World Cancer Research Fund stated that there were over 500.000 cervical cancer cases in 2018. In Indonesia, CT-Scan is a common method in screening cervical cancer. However, CT-Scan images tend to have a low contrast thus making it difficult to differentiate normal organs and the cancer, which may lead to misinterpretation. This research focuses on developing a CAD scheme for the segmentation of cervical cancer CT-Scan images to assist doctors and radiologists in cervical cancer screening. The algorithm developed consisted of feature extraction and pixel classification in the CT-Scan image using K-Nearest Neighbors classifier. Experiments were done by using two different feature extractions (pixel minimum, maximum, mean HU values and direct pixel HU values) with three different K values (K=3, K=5 and K=9). Results showed that the first and second experiment had balanced accuracy of 59.484% and 58.552% respectively. Moreover, the increased K values showed to decrease the balanced accuracy by 0.287-2.227%. This CAD system needs to be further developed in order to reach a higher accuracy. However, the CAD system itself is not expected to make a solid 100% accurate diagnosis, but to assist radiologists and doctors in screening cervical cancer CT-Scan images.