Introduction
Cancer of the cervix is the second commonest life-threatening cancer among women worldwide (Arbyn et al., 2008; Behtash & Mehrdad, 2006). The squamous cell subtype is the most prevalent (Behtash & Mehrdad, 2006; Wang et al., 2004), which is often associated to Human Papilloma Virus (HPV) infection (Nogueira-Rodrigues et al., 2014; Vizcaino et al., 1998; Ferlay et al., 2014; Bosze & Syrjanen, 2010; Mateu-Aragonés, 1969). Studies also have shown that HPV infection produces inflammatory changes in the vascular pattern of the cervix (Clifford et al., 2003; Munoz et al., 2003; Munoz et al., 2004; Peirson et al., 2013; Saslow et al., 2012; Zur, 1982; Franco et al., 2001). The changes in the vascular pattern are believed to be present at early stages of cervical cancer (Guidi et al., 1995). Such changes are often diagnosed by using colposcopy.
Colposcopy developed by Hinselmann, in 1925, is commonly done by direct observation of the uterine cervix (Fusco et al., 2008-1; Zhu et al., 2008; Durdi et al., 2009; Massad et al., 2013; Bappa & Yakasai, 2013). Acetic acid topical application, or acetowhitening, on the suspicious abnormal area is used to improve the visualization of cervical tissue (Schiller, 1933). The application of the acetic acid induces changes of the colour of the area under study. The new colour of the anatomical area under study ranges from white to gray; such colour change appears when premalignant cells are present. The described chemical reaction is due to a malfunctioning of epithelial cells of the cervix (Reid et al., 1984-1; Mateu-Aragonés, 1964; Coppleson et al., 1993), which is induced by a number of agents, including HPV. This infection disorientates keratin filaments in the superficial layers of the squamous epithelium (Zijlstra et al., 1991) hardly observed by colposcopy done by direct observation; further, the abnormal lesions are not easily detected by the Schiller test (Fusco et al., 2008-1; Reid et al., 1984-1), a test that uses iodine (Schiller, 1933), often added to acetowhitening.
Hinselmann created an optical stereoscopic microscope to visualize the cervix using the aforementioned technique. Such microscope used incandescent light or white light-emitting diode (LED) type. It allowed visualizing the morphology of the blood vessels of the cervix (Fusco et al., 2008-2). Kraatz (1939) used a green filter on the light beam to highlight the vascular pattern, which this author contrasted with the tissue background. More recently, Mateu-Aragonés (1964; 1979) proposed the vascular pattern of the suspicious tissue to classify lesion severity (Coppleson et al., 1993; Zijlstra et al., 1991; Lickrish, 2000; Reid et al., 1984-2); however, this classification was mainly based on clinical observations. Limitations of current optic methods or disturbed visual acuity preclude the detection of critical fine vascular changes that are key elements of lesioned tissue to make an accurate and timely diagnosis (Vlachokosta et al., 2013). Thus, state-of-the-art imaging techniques such as the ones we demonstrate in this research may be of value to improve early diagnosis of cervical cancer (Durdi et al., 2009; Xue et al., 2007-1; Xue et al., 2007-2).
As a fact, digital techniques using appropriate filters make feasible to observe key elements of the cervical tissue (Lee & Park, 1990; Shamir et al., 2008), modulating technical details such as brightness and contrast. Germane to this, segmentation techniques have allowed differentiating vascular anomalies of the cervix including vascular loops and webs (Srinivasan et al., 2009; Xue et al., 2010; Mehlhorn et al., 2012; Lee & Park, 1990; Shamir et al., 2008; Srinivasan et al., 2007; Dvir et al., 2006; Lehmann & Palm, 2006; Zimmerman-Moreno & Greenspan, 2006). However, at present, none of the above mentioned biotechnological improvements have properly made possible to highlight the vascular pattern of the uterine cervix. This research aims to fill this gap. Thus, we developed a novel method using digital image processing. This method looks for enhancing the vascular patterns of uterine cervical tissue in images obtained by digital colposcopy; the results of this investigation will help to detect lesions of human cervix at earlier times than done at present.
Methods
Selection of the picture (step Io, figure 4)
For this research, which was approved by the Institutional Review Board of Mediciencias Research Group, we selected 64 colposcopic images from our databank. The images were zoomed at 15x (Leisegang Optik TM model 2 colposcopy, Berlin, Germany). The digital camera that took the photographs had a resolution of four mega-pixels. The images were stored in Joint Photographic Experts Group (JPEG) format. Then, the pictures were pseudorandomly selected according to defined criteria. The inclusion criteria of the samples we selected took into account any injury observable during the colposcopic procedure, including inflammation or regions with suspected noninvasive cancer. An exclusion criterion was that the images had less than 40 kB; we also excluded images found more than once in the list searched at the data bank (Figure 1). The final sample was composed of four images.
Selection of the wavelength of interest (WOI) (step I1, figure 4)
The electromagnetic band between 500 nanometers (nm) and 600 nm (Figure 2) has two hemoglobin related peaks above the baseline (Figure 3). The first peak appears at 541 nm, and the second peak at 577 nm (Zijlstra et al., 1991) (Figure 2). Accordingly, we selected WOIs of the image localized at the green band of the electromagnetic spectrum, which has between 500 nanometers (nm) and 600 nm (Figure 2) to properly detect oxyhemoglobin from the colposcopic images, in a red-green-blue (RGB) model.
Unsharped mask (step I2, figure 4)
To better analyze the blood vessels of the cervix, unsharped mask was added. To accomplish this, a Laplacian filter was used. This filter enhances image regions having that display high frequency levels; in making so, it is possible to achieve a much better contrast of the image. The Laplacian filter was constructed as follows: First, a negative Laplacian matrix He was defined [Equation 1]. Second, a vector named to as alpha (α), which is a constant number that ranges between zero and one, was used to control the shape of the Laplacian filter (Zimmerman-Moreno & Greenspan, 2006). The Unsharped image ( I2 ) obtained was the result of the convolution of He and the original image ( I1 ) [Equation 2].
Treatment of specular highlights (step I3,figure 4)
A specular highlight is the spot of light that appears on objects when they are lighted up. In our case, the fluids of the cervical mucosa produced such specular highlights (Dvir et al., 2006). In image processing techniques, methods that improve the quality of the image such as lighting, saturation and OTSU threshold method (Srinivasan et al., 2007; Lehmann & Palm, 2001) are commonly used. However, such methods lack specificity to correct non uniform illumination of the image under study. Therefore, better methods have to be considered. In this line, top-hat methodology allows extracting the information needed from pictures in a better way than the mentioned above. Thus, we used it. The equation 3 shows the difference between the images obtained in the step I1 minus the opening operation using a circular structure element EETH ; it was done assuming that the brightness approximated to spotted shapes.
The opening operation produced mathematical erosion; for this reason, the image was expanded. Of note, this operation allowed obtaining a better definition of the object’s edges, with a simultaneous increase in the pixel size.
Specular highlights are those ITH that have values higher than the threshold named to as Uo. This threshold was determined by the OTSU threshold method, which consists in the separation of pixels in two forms named to as foreground and background pixels (Lehmann & Palm, 2001). Thus, the threshold value was calculated taking the minimal intra-class variance (Otsu, 1979). The obtained data were then masked by M [Equation 4].
To improve the quality of the image, a Laplace interpolation method was also used (Zimmerman-Moreno & Greenspan, 2006). It was done by dilating the M mask with a structuring element named to as EED [Equation 5 ]. This manipulation ensured that the M1 mask covered the highlights, before to use the Laplace interpolation.
The Laplace interpolation method (∇2) allows assign an intensity function named to as y(x), using known points y1 of I1 stage, as shown in equation 6. After this, we solved the equation (7) for each unknown points. The Laplacian operator ∇2 was interpolated to improve the quality of the image.
Low-pass filtering (step I4, figure 4)
Following the technique described by Tomazevic et al (2002), we created an image called IPB, which was obtained after convoluting it. Such convolution was the result of multiplying I1i times H as shown in equation 8. Hm in equation 9, corresponded to the average low-pass filter, and Hg in equation 10, corresponded to the Gaussian filter. M, N and σ values were not changed at the frequency values of interest, as shown in equation 11. This conserved both, the smooth changes coming from the illumination procedure, and the changes produced by variations of the intensity in the tissue.
Adjustment of intensity levels of pixels (step I5, figure 4)
In this step, we used a linear transformation function T (l) [equation 12]. The equation 12 represents the light intensity of each pixel. ai and bi are two values obtained at the initial moment of the light intensity. They corresponded to the minimum and maximum of the intensities range, where the normalized histogram exceeded 5% of the total of the pixels under study. These data were converted in two new values of intensity named to as ao and bo (they corresponded to the values obtained among those that desired redistribute the new histogram). Such conversion was done for redistributing the intensity of the pixels within the range of available intensities.
Li showed a new intensity distribution, which was obtained from Lo levels as shown in (Equation 13).
The histogram of the resulting image from the illumination correction (I3, figure 4 and figure 5) was concentrated in a particular area; it did not occupy the entire range of intensity levels. Using the transformation function T (L), an image with higher contrast was obtained (I4, figure 4). As a result, the levels of the I3 image are distributed in a wider range of intensities currents as shown in Figure 5. Lastly, the values obtained by top-hat method were statistically compared with the values obtained by direct threshold method. Chi-square was used to compare proportions. P value was setup at 0.05.
Results
A full description of the results is presented in figure 6. In (a) a green layer of the original image is shown; (b) shows the Interpolation of specular highlights; (c) presents the filtering process; (d) allows knowing Correction of illumination, and (e) reports adjustment of intensity levels.
In figure 7, two original colposcopic images were included. The first image (1a) showed highlights whereas the second image (2a) did not. In figure 7b, it was observed the data obtained after treating specular highlights using the direct method. In figure 6c, the image was analyzed using the top-hat method. It was found that with the direct method, highlights were not detected in the darkest region (figure 7, 1b). In absence of highlights, artifacts were detected (figure 7, 2b). Using the top-hat method the opposite occurred.
In figure 8, sections of vascular images obtained from colposcopy showed hairpined (a), clothed (b) and dotted patterns (c, d). After performing the enhancement process in the three layers of the RGB model, a color image was obtained. In (c) the vascularization pattern was highlighted in a dotty-shape like form that was not evident prior to processing. In (d), a lesion on the squamous epithelium was also observed. Of note, this lesion was unseen in the original image obtained, due to the lack of observation of internal vascularization. It must be emphasized that once the enhancement was made, the vascular structures on the affected region were better observed; therefore, the contour of the injury delineates better. Additionally, when comparing average filter and Gaussian filter, there were not differences in the drawing of the vascular network; hence, any of these filters would offer similar results.
When Tomazevic et al (2002) compared different methods to correct non-uniform illumination, aimed for the method of entropy minimization as the best in performance. This method enables the identification of artifactual highlights, without modifying the actual intensities of tissue itself. In our study, when the direct threshold method was compared with top-hat method, the highlights were not detected at the darkest region; while using top-hat method, those highlights can be detected independent of illumination. Additionally, when evaluating highlights of the whole image the true positive findings obtained by using the top-hat method allowed to identify 93,1 % of them, whereas the direct threshold method identified 85,1 % of such highlights (p < 0.05). For this reason, this work used the top-hat method for detection of specular highlights in the images.
For the images under study, we found that the average filter without approach has less noise, but in regions where there is variation in intensity, shadows or artifacts that affect the reading of the test is presented, as in the contour of the lesion in Figure 7 (b). Gaussian filter with greater emphasis and contrast of the vasculature was achieved by reducing the presence of artifacts; however, a slight increase in noise was observed.
Discussion
A novel way of enhancing vascularization in colposcopic images has been uncovered in this research by using the top-hat method. It is accepted that the top-hat methodology may similarly use average filter or Gaussian filter. Of note, the average filter has less noise than the Gaussian filter. However, some caveats have to be kept in mind when using the average filter. This filter picks up more artifacts in regions of interest where uneven intensity occurs, and also, it retains the shape and size of blood vessels simultaneously accentuating the noise present in the image.
To accurately use the techniques reported here, it has be kept in mind that there a number of intensities that are visible when picturing colposcopic images. One of the intensities refers to highlights (Lehmann & Palm, 2001); other intensities refer to actual lesions in uterine cervical tissues. Such unwanted and mixed visual information can give inaccurate results. Because of these shortcomings, computational analysis of the colposcopic images has been worked out during recent years (Zimmerman-Moreno & Greenspan, 2006) leaving still some open questions due to the differences in the imaging patterns displayed by the aforementioned intensities. In this regards, top-hat methodology has filled out the gaps, as demonstrated here.
It should be noticed that the top-hat method used by us in this research, allowed improving in a very significant way the detection of the abnormal vascular pattern as well as the actual morphology and size of the blood vessels of cervical tissue with a low cost-technology. These results might help to increase the likelihood of making the diagnosis of uterine cervical cancer in humans at earlier stages of the disease than done until present. These findings will also help to do a better characterization of cervical lesions than previously done (Xue et al., 2007-2) in both develop and developing countries; moreover, this research could be the basis of improving the visualizing of vascular patterns elsewhere in the body.
Conclusions
The digital analysis of the images expressing oxyhemoglobin, uncovered with this research, accurately defines the size and shape of the blood vessels of the uterine cervix. This novel methodology improves the visualization of the vascular pattern in colposcopic images, allowing making a more accurate staging of uterine cervical cancer.