Original ArticleDetection of Specular Reflection and Segmentation of Cervix Region in Uterine Cervix Images for Cervical Cancer Screening
Graphical abstract
Introduction
Cervical cancer starts in the cervix and is mostly caused by sexually-acquired infection with Human Papilloma Virus (HPV). It covers 12% of all cancers and is the second most common cause of death among women worldwide [1]. It takes several years for the epithelial changes to progress from pre-cancerous stages to invasive cancer. Therefore, there is adequate time for screening, detection and management of the pre-cancerous stages. The screening methods such as cytology, visual examination, HPV-DNA are used for early detection of the cervical cancer. Pap smear (cytology) is considered as gold standard for cervical cancer screening. However, the requirement for well-established laboratory infrastructure has limited its suitability for resource poor regions. Visual inspection with acetic acid (VIA) is used as a low cost alternative to Pap test [2], [3], [4]. In this test, 3–5% acetic acid is applied to cervix and is left for one minute. Pre-cancerous lesions turn white on combining with acetic acid. They are called acetowhite regions and are important diagnostic features for cervical cancer detection. Cervicography [5] is analogous to VIA; in which images of cervix are acquired before and after the application of acetic acid. These images are sent to the experts who evaluate these images. The accuracy of evaluation is dependent on the expertise level of the evaluator and is highly subjective. The advances in digital imaging have enabled the acquisition of high quality cervix images at low cost. In addition, advances in image processing techniques have facilitated processing of these images using computerised tools thus can be used to compensate for the lack of expertise in low resource regions. Such automated image analysis system for cervix images can eliminate the subjectivity and improve the accuracy. Automated image analysis system for cervix images aims at identifying diagnostic features like acetowhite lesions, vascular abnormalities such as mosaics and punctations. There are various challenges for the extraction of acetowhite regions present in the cervix images. One of them is presence of Specular Reflections (SR) in the image. The specular reflections are bright white regions present on the cervix image. They are caused due to reflection of light from wet surface of the cervix. It is important to eliminate SR which otherwise would interfere in the detection of acetowhite regions. The cervix images may contain regions like vaginal walls, medical instruments and cotton swabs, which are not needed to be analysed. In addition, color and texture of these regions sometimes mimic the diagnostic features. Hence, it is desirable to identify the cervix region, which defines the Region of Interest (ROI) for cervix image analysis. An example of cervix image with various regions is shown in Fig. 1.
Specular reflection detection was dealt in many ways by the researchers [6], [7], [8], [9], [10]. Holger Lange [6] extracted a feature image, followed by adaptive thresholding to detect the SR. Zimmerman G et al. [7] utilised intensity, saturation and gradient information to detect the SR. Othmane E M et al. [8] filtered the image, converted filtered image into XYZ color plane and used luminance component and normalised luminance component for detection of SR regions. Abhishek Das [9] et al. extracted white pixels from red, green and blue channel and applied a logical AND operation on them to get the SR mask. Bianca Regeling [10] et al. used boxplot method for identification of SR on hyperspectral larynx images.
In literature ROI was detected based on color, position, curvature and shape features. Wenjing Li et al. [11] filtered the RGB image and applied Karhunen–Loeve (K–L) transform, followed by expectation maximisation algorithm to get the cervix region. They reported the extension of this work in [12]. In this work they eliminated the portion of vaginal folds present in the cervix region extracted in [11] using curve fitting [13]. A combination of color and position features were used for coarse ROI detection [14], [15], [16], [17], [18]. The cervix region detected using this method could not eliminate the vaginal walls completely. Accurate cervix boundary was detected using active contours based on curvature and color features [19], [20]. Cervix region commonly takes the shape of an ellipse or a circle. Shelly Lotenberg et al. [21] used the shape information to segment the cervix region. Shape models (ellipse and circle) were embedded in active contour framework to segment the cervix region. Costas P. et al. [22] proposed an automated method for enhancing and identifying partial curvilinear structure on geophysical images. We propose an algorithm for SR detection using a standard deviation filter and cervix region segmentation algorithm by adapting the curvilinear structure enhancement suggested in [22].
Section snippets
Data collection and ground truth
Color images of the cervix were acquired in screening programs at the health centres, conducted by Kasturba Medical College, Manipal, Karnataka, India. Married women above 25 years of age were considered for the study. Informed consent was obtained from the women participating in the study. 151 images of cervix were acquired one minute after the application of acetic acid using a Moto G Turbo cell phone with 13 MP camera. The ROI was manually annotated in all these images by a medical expert in
SR detection
Fig. 4 shows the results of SR detection methods suggested in [9], [10] and the proposed method on the cervix images 1–6 where manual marked ground truths are available.
The overlapping measures, sensitivity and Jaccard index (JC) for 6 images and methods described by [9], [10] and the proposed method are tabulated in Table 3.
Grading by the expert for 151 images is summarised in Table 4.
We have considered images with different amounts of brightness 1–5 for testing the algorithm. Fig. 5 shows the
Discussions
An algorithm for detection of specular reflection and cervix region segmentation is presented. The proposed algorithm was tested on 151 cervix images. SR detection algorithm was compared with two state of the art algorithms presented by Abhishek Das et al. [9] and Bianca R et al. [10]. Fig. 4 depicts the performance of three algorithms. Also, listed two overlapping measures sensitivity and Jaccard index for three algorithms (Table 3). It can be observed from Table 3 that the proposed algorithm
Conclusions
This paper discusses an image pre-processing approach including specularity detection and region of interest segmentation. This pre-processed image can subsequently be used for image analysis for detection of cervical cancer. Specularity detection algorithm detects most of the specularities present in the image irrespective of illumination variations. Use of novel feature image higher contrast across specular regions resulting in strong response to standard deviation filter. Robustness of SR
Conflict of interest
The authors declare that they have no conflict of interest.
Acknowledgements
This publication is made possible by a sub agreement from the Consortium for Affordable Medical Technologies (CAMTech) at Massachusetts General Hospital with funds provided by the generous support of the American people through the United States Agency for International Development (USAID Grant number 224581). The contents are the responsibility of Manipal University and do not necessarily reflect the views of Massachusetts General Hospital, USAID or the United States Government.
We would like
References (25)
- et al.
Statistical validation of image segmentation quality based on a spatial overlap index
Acad Radiol
(2004) - et al.
A multidimensional segmentation evaluation for medical image data
Comput Methods Programs Biomed
(2009) - Ferlay J, Soerjomataram I, Ervik M, Dikshit R, Eser S, Mathers C, et al. GLOBOCAN 2012 v1.0, Cancer Incidence and...
- et al.
Visual inspection of the uterine cervix after the application of acetic acid in the detection of cervical carcinoma and its precursors
Cancer
(1998) - et al.
Visual inspection with acetic acid in the early detection of cervical cancer and precursors
Int J Cancer
(1999) - et al.
Test characteristics of visual inspection with 4% acetic acid (VIA) and Lugol's iodine (VILI) in cervical cancer screening in Kerala
India Int J Cancer
(2003) - et al.
A clinical review of cervicography
Ir J Med Sci
(1990) Automatic glare removal in reflectance imagery of the uterine cervix
- et al.
Automatic detection of specular reflections in uterine cervix images
- et al.
Automatic detection and inpainting of specular reflections for colposcopic images
Cent Eur J Comp Sci
(2011)