Automatic Detection of Cervical Region from VIA and VILI Images using Machine Learning

Abstract

Cervix is lower part of uterus in the female reproductive system. Images of the cervix are captured at the vaginal cavity using a speculum. Such images may be used for detecting cervical abnormalities manually or using auto detection software. However, these images usually contain extra data outside of the cervix for example, part of speculum or other non-cervix tissues. Cervix segmentation is a process of extracting cervix from image, thereby localizing the cervical region by eliminating the non-cervix regions in the image. In this work we have used object detection algorithms-SSD and Faster-RCNN for extracting cervix from the images obtained by performing VIA (Visual Inspection with acetic acid) and VILI (Visual Inspection with Lugol's Iodine) tests. We have also used classical image processing approach of clustering and thresholding for cervix segmentation in VIA images. Finally, a comprehensive analysis and a comparative study of machine learning and traditional image processing approaches on factors such as better fit, time for execution and accuracy of abnormality prediction for VIA cervix segmentation has been presented.