Acetowhite region segmentation in uterine cervix images using a registered ratio image

• Published in: Computers in biology and medicine Vol. 93; pp. 47 - 55
• Main Authors: Liu, Jun, Li, Ling, Wang, Lei
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.02.2018; Elsevier Limited
• Subjects: Acetic acid; Acids; Adult; Algorithms; Automation; AW segmentation; Cancer; Cancer screening; Cervical cancer; Cervical cancer screening; Cervix; Cervix image; Cervix Uteri - diagnostic imaging; Clustering; Colposcopy - methods; Developing countries; Experiments; Female; Human papillomavirus; Humans; Image processing; Image Processing, Computer-Assisted - methods; Image segmentation; Internal Medicine; LDCs; Level set algorithm; Medical screening; Methods; Other; Patients; Uterine Cervical Neoplasms - diagnostic imaging; Uterus; Cervical cancer screening; Cervix image; Level set algorithm; AW segmentation
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2017.12.009

Visual inspection with acetic acid (VIA) remains a main cervical cancer screening tool in developing countries. However, it depends on the operator's experience, and its utility is often limited by the lack of trained doctors. Smart colposcope devices to automatically detect the cervical intraepithelial neoplasia (CIN, the early stage of cervical cancer) may provide a promising alternative. As the acetowhite (AW) region is the most important feature of CIN during VIA, its segmentation is considered an important procedure in the automatic detection of CIN. In this study, an automatic AW region segmentation algorithm based on the pre-acetic-acid and post-acetic-acid test images was developed. The cervix region was extracted according to a clustering algorithm from the pre-acetic acid test image. A ratio image was then obtained after registering the pre- and post-acetic-acid test images to facilitate the segmentation of the AW region using a modified level set algorithm. The results showed that although the developed algorithm yielded a mean sensitivity of 71.86%, which was lower than that of the fuzzy C-means (FCM) algorithm by 12.08% and the classical CV model-based level set algorithm (CV-LSA) by 4.04%, a high mean specificity (92.76%) was achieved that was greater than those of FCM and CV-LSA by 46.61% and 31.34%, respectively. Additionally, a high Jaccard index (JI) mean accuracy of 61.51% was achieved, which was greater than those of FCM and CV-LSA by 18.74% and 17.14%, respectively. This new algorithm, with an improved segmentation performance over traditional algorithms, may serve as a promising tool to advance the clinical prognosis of cervical cancer. •A method by registering pre- and post-acetic test images to automatically segment the AW region in cervigrams was developed.•A modified CV level set algorithm was developed that showed improved segmentation accuracy over three traditional algorithms.•The evaluation of segmentation performance was based on a heterogenous patient cohort with varying CIN degrees.