Keywords

1 Introduction

Cervical cancer (CC) is the fourth most recurrent women’s cancer worldwide. In line with the WHO health report, every fifth woman in the world will be impacted by it in 2050 [1]. Nearly 90% of the 270,000 deaths from CC in 2015 took place in low- and middle-income countries. Noteworthy progress in disease screening and treatment supports prevention, and prompt diagnosis may drastically reduce the CC mortality rate [1].

CC begins with abnormal modifications in the cervical tissue. The risk of having these unusual changes is concomitant with infection by the human papillomavirus (HPV). Moreover, early sexual interaction, manifold sexual partners, usage of oral contraceptives (birth control pills), unhygienic lifestyle, and misinformation are the critical factors for spreading this disease. If spotted early, CC can be cured reasonably. The most prevalent CC detection method is the Pap smear.

Nonetheless, it has inherent limitations such as sample quality, slide quality, and effectiveness of screeners. The CAD systems can help to treat this disease by analyzing an input image and, with the assistance of various image-processing algorithms, predict or detect abnormalities. The earliest and challenging step in medical image exploration is to pre-process the input image for the uncovering and removal of noise. Specular reflection (SR) is a variety of prominent noise that appears in photography and medical imaging. Once a ray of light strikes the surface, a portion of the ray is straightaway reflected from the interface amid the surface and the air, thanks to their different refractive indices. This reflected light is called SR [2]. The humidity on the cervix surface engenders the SR, which hampers early CC detection by computational systems [3]. Figure 14.1 illustrates the initial cervix image with the SR region (black box in the middle) and cropped SR region near it so that these regions undergo automatic detection, correction, and deletion according to the specialist’s needs.

Fig. 14.1
figure 1

Example of an SR affected cervix image with cropped SR region

Full size image

The rest of the manuscript goes as follows: Sect. 14.2 talks over the state-of-the-art methods for SR detection and its removal. Section 14.3 describes the components of the ASRDC scheme. The experimental outcomes appear in Sect. 14.4. Section 14.5 closes this work with remarks on the ASRDC methodology and its future.

2 State-of-the-Art Technology

Automatic recognition and removal of SR experimented a few contributions lately. The correlated literature generally embraces four categories as (i) the dichromatic reflection model (DRM) usage, (ii) kernel filtering, (iii) SR cast as classification, and (iv) thresholding procedures.

The DRM principle states that a reflection combines specularity and diffusion linearly. Yoon et al. appraised the value of the specularity-invariant pixels as well as their ratio to set apart diffuse components. Still, this maneuver suits textured imageries, and approximation in the normalization procedure bounds the accuracy of SR detection [4]. Tao et al. introduced a new metric termed line consistency for depth estimation of specular regions. They had estimated colors from multiple light sources. However, this strategy failed to distinguish saturated specularity [5]. J Suo et al. applied the DRM rationale perceiving the problem as a signal separation for SR detection and removal.

In contrast, the procedure overlooked smooth color alterations, and it failed to discern the pixels with identical hue and different saturation [6,7,8]. Das et al. advised kernel-based filters for SR detection and exclusion, e.g., filling, dilation, multiscale morphology, and IS-histogram [3]. Kudva et al. utilized morphological kernels as filters to acquire features from color images [9]. Xue et al. predefined the structuring element (SE) as a kernel. The top-hat transform treated the intensity image I (whose entries hold brightness values within some range) of input cervigrams [10]. Yet, all these schemes rest on the size and shape of the kernel applied to the database. Gao et al. proposed SR detection as a classification problem using SVM classifiers. This method’s caveat is the requirement of training every time for SR detection [11].

Zimmerman et al. suggested multiplying S and I by an arbitrary constant where S is the saturation component, which shows how much the white color taints a given color. The S component belongs to the range [0, 1] within the HSI plane. The gradient image outputs on these multiplied regions are SR pixels [12]. Akbar et al. computed the SR pixel via a chaotic clonal selection procedure [13]. The image specular degree can function as a thresholding parameter to separate diffused images and for SR detection [2, 14]. The choice of arbitrary constants throughout automatic detection of SR pixels may be contingent on the database under experimentation. However, the detection system must be entirely automated and independent of the database. Therefore, any imaging modality calls for automatic threshold selection. Table 14.1 relates the state-of-the-art SR discovery approaches and the ASRDC concept. Automatic thresholding fits in fivefold groups according to the information content they rely upon, viz.:

  1. (i)

    Histogram-based schemes analyze the primary intensity, decimation in intensity range, and nonlinear nature of the smoothed histogram.

  2. (ii)

    Clustering-related strategies split the gray levels from the input image into the background and foreground pixels.

  3. (iii)

    Entropy-based methods employ local entropy, cross-entropy of the foreground and background regions, original and binary images, etc.

  4. (iv)

    Object attribute-based methods focus on the similarity between the gray level and its black and white versions.

  5. (v)

    Statistical relation-founded schemes rely on higher-order moments and/or the correlation among pixels for threshold selection [15].

Table 14.1 State-of-the-art SR detection categories
Full size table

These threshold-picking strategies can be either bi-modal or multi-modal. Nonetheless, the application demands to get SR pixels, which are always bright. Hence, bi-modal distribution is the best choice, along with a histogram-based approach. Table 14.2 abridges a review of automatic threshold determination practices built on histograms.

Table 14.2 State-of-the-art SR detection categories
Full size table

In 2004, Sezgin and Sankur reviewed the performance of thresholding techniques using five quality measures, viz., misclassification error (ME), edge mismatch (EMM), relative foreground area error (RAE), modified Hausdorff distance (MHD), and region nonuniformity (NU). They calculated the average score of each scheme, ranked individual quality measures, and, finally, concluded that Kittler and Kapura were the superlative adaptive thresholding procedures [15]. Donald Bailey also investigated adaptive thresholding techniques for performance analysis and found that Kittler’s minimum error is the best [20]. The ASRDC methodology overcomes the limitations above due to:

  1. (a)

    Its complete independence on size and shape of the kernel

  2. (b)

    No requirement for the training process

  3. (c)

    Fully automatic threshold calculations

The catchline features of the ASRDC methodology are:

  1. 1.

    Use of the lightness as a no-reference quality measure for selection of appropriate algorithm

  2. 2.

    Automatic selection of threshold by a modified Kittler’s method

  3. 3.

    Automatic enhancement of low-quality images before the SR treatment

3 The ASRDC Methodology

The ASRDC block diagram appears in Fig. 14.2. The section further describes the threefold contribution of the chapter.

Fig. 14.2
figure 2

Block diagram of the ASRDC system

Full size image

3.1 Selection of Optimum Threshold Detection Technique

The authors picked the Kittler minimum error thresholding scheme for automatic SR detection [19] from the approaches talked over in Sect. 14.1.

3.2 Automatic SR Detection

A histogram exemplifies the distribution of the pixel intensities, where an SR is a bright spot on an image, which agrees with the maximum part of the intensity range (close to white). Nevertheless, few non-SR pixels may also possess high brightness. SR pixels occur at the dark side of the S in addition to the bright side of the I images [14]. As a consequence, the SR occurrence in the intensity saturation (IS) histogram is a foremost ASRDC concern. The automatic SR recognition is carried out by a simple variant of the Kittler method to attain the optimum threshold on the S and I images. The modified Kittler method (MKM) understands the threshold (T) differently from the original tactic. This work considers the span going from minimum to maximum intensity (i.e., over complete dynamic range) as opposed to starting with a random T. An optimum threshold matches the minimum Jaccard Index (JI) value (aka criteria function) of the MKM.

The Jaccard Index J is a statistic that explains the similarities between finite sample sets. J is defined properly as

$$ J=\frac{\left|A\cap B\right|}{\left|A\cup B\right|}=\frac{\left|A\cap B\right|}{\left|A|+|B|-|A\cap B\right|}, $$
(14.1)

i.e., the size of the intersection |A∩B| divided by the size of the union |A∩B| of the sample sets A and B.

The ASRDC methodology employs half of the intensities for two different inputs. The thresholds, starting from 1 to 128, assist in calculating the minimum error with the MKM criterion function and are known as the left-side threshold (TL). Similarly, the right-side threshold (TR) results from all combinations of pixels from 128 to 256. The modified algorithm is given below.

The MKM Algorithm

  1. 1.

    Go through every possible threshold (T), i.e., grey level from 1 to 128 or from 128 to 256.

  2. 2.

    Consider the groups (i) 1 or T, and (ii) 2 or (T+1), i.e., highest intensity.

  3. 3.

    Compute the histograms of these groups and mark their sums as P1 and P2.

  4. 4.

    Determine the mean and standard deviation for the histograms.

  5. 5.

    Compute the Jaccard Index (J) criterion function for all possible T.

  6. 6.

    The finishing threshold is the position with minimum J.

Thresholds generated from step 6 (TL and TR) work on saturation (S) and intensity (I) images, respectively. The SR pixels as given below.

$$ SR=\left(S\left(i,j\right)< TL\right)\&\&\left(I\left(i,j\right)> TR\right) $$
(14.2)

where (i, j) symbolizes the pixel location (for row i and column j) in an picture and “&&” is the logical AND.

3.3 Image Quality Assessment (QA)

The overall input image condition dramatically impacts any algorithm performance. Therefore, the input image quality assessment (QA) is vital to the deployment of an adaptive system. The objective QA of the examined picture consists of the computation of the threefold quality metric categories, viz., full reference (FR), reduced reference (RR), and no reference (NR). As ground reality images are not available when it comes to cervigrams, the NR metric is preferred in input quality testing. Several investigators commended many color quality parameters or attributes for NR-based QA like brightness, colorfulness, sharpness, contrast, and entropy. The colorfulness illustrates the color information perceived by the human eye. The sharpness gives the amount of preservation of edges. The contrast addresses the emphasis on the foreground and background association. The average image information corresponds to the brightness measures, whereas the lightness promotes the distortion in intensities of pixels [22].

It is essential to use quality measures related to distortion for SR detection. Hence, the ASRDC scheme takes account of the lightness parameter together with colorfulness (C1), contrast (C2), and sharpness (C3). Their grouping forms a quality measure if and only if they are correlated, which, consequently, leads to the calculation of the correlation between the lightness and the three attributes. The validation of the null hypothesis can confirm the possibility of this combination, i.e., the lightness is uncorrelated with all three attributes. The Pearson correlation coefficient (p-value) gives the acceptance probability of the null hypothesis. In general, the significance level of the p-value is 0.05, namely, if the p-value is less than 0.05, attributes are correlated with the refutation of the null hypothesis [23]. The tryouts from Sect. 14.4 (B) substantiate the dominance of lightness features among the color attributes to deal with an eventual image enhancement. The experimental investigation of the ASRDC method indicates that if the lightness is greater than 1, then image quality amelioration is required.

The low-quality images must be enhanced before applying the ASRDC methodology. This scheme has a histogram-based automatic threshold selection (Sect. 14.3 (B)) that also enhances pictures by altering the histogram shape.

Most prevalent histogram-centered measures to enhance pictures are histogram equalization (HE) [17], bi-HE (BHE) [24], adaptive HE (AHE) [32], contrast-limited AHE (CLAHE) [33], and brightness preserving BHE (BBHE) [24]. Among these approaches, the BBHE scheme preserves the image characteristics after pre-processing. Hence, this work treated low-quality images with BBHE before employing the ASRDC method to manage SR in an input image adaptively as:

  1. (a)

    Initial quality check of the input picture utilizing the lightness parameter

  2. (b)

    Low-quality image enhancement through BBHE

  3. (c)

    Application of the ASRDC technique to the input image

3.4 SR Inpainting

Further SR inpainting is carried out to generate the SR free image. An iterative nonzero averaging filter creates an SR free image [10]. The complete flow of the ASRDC method arises in Fig. 14.3 with its algorithm.

Fig. 14.3
figure 3

ASRDC workflow

Full size image

ASRDC Algorithm

  1. 1.

    Take the input cervix image and calculate the lightness.

  2. 2.

    If lightness < 1,

    1. (i)

      RGB to HSI conversion to separate saturation (S) and intensity (I) images

    2. (ii)

      Compute left (TL) and right threshold (TR) using the ASRDC method

    3. (iii)

      Collect pixels lying between intensities of S less than TL and intensities of I greater than TR, as SR pixels

    4. (iv)

      Perform mask enlargement on the output of step 2 (iii)

  3. 3.

    If lightness > 1, Apply BBHE on the input image and extract red component

    1. (i)

      Calculate TR from BBHE image

    2. (ii)

      Collect pixels of the input image which are greater than TR, as SR pixels

  4. 4.

    Inpaint detected SR pixels by mean color replacement.

4 Results and Discussions

4.1 The Dataset

The present exploration involves digitized uterine cervix pictures collected by the National Cancer Institute (NCI) from four epidemiological studies made in the USA, e.g., “Costa Rican Natural History Study of HPV and Cervical Neoplasia (NHS),” “ASCUS LSIL Triage Study (ALTS),” “Biopsy Study,” and “Costa Rica Vaccine Trial (CVT)” on HPV and CC screening [25]. The trials comprise a total of 612 images from all 4 datasets distributed as NHS (200), ALTS (200), Biopsy (50), and CVT (162). This research work selects images randomly from the available databases with resolutions for pictures in the ALTS and NHS equal to 2891 × 1973 and for the Biopsy and CVT, 4256 × 2832.

The acceptance of the image enhancement output entails the evaluation of the quality of the input image. As talked over in Sect. 14.3.3, the p-value is calculated to obtain the correlation between image attributes. Table 14.3 displays the p-values for various arrangements of images from the datasets.

Table 14.3 p-values between C4 and C1, C2, and C3
Full size table

C1, C2, C3, and C4 represent the colorfulness, sharpness, contrast, and lightness attributes, respectively. The p-values of C3 are very low (i.e., <0.05 for all combinations of the dataset). However, C1 and C2 show significant p-values concerning the significance level. This implies a correlation between lightness and contrast. Thus, the lightness can be combined with contrast to test the image quality. The present analysis selected 80 images (20 from each dataset) to determine the dominant feature between lightness and contrast. This experiment aims to confirm the necessity of image enhancement through no-reference image attribute. The contrast of all 80 images ranges between 0.45 and 0.5, which did not give a noteworthy threshold as a decision parameter. However, a significant change in the value of lightness is observed for all 80 images, as below and above value 1. Thus, lightness is chosen as a dominant feature of the cervix color image to decide the input quality. Experimentation concluded the adaptability condition as if the lightness is less than 1, input image quality is satisfactory, and the ASRDC algorithm should get applied without image enhancement. For an image with lightness greater than 1, it should be enhanced before applying the ASRDC algorithm.

4.2 Experiments

The S and I images are normalized to the original grayscale range of 0 to 255. TL is calculated for the S image, and TR is calculated for I the image, as explained in Sect. 14.3.2. The original Kittler method and the MKM are applied to normalized S and I images. The final SR pixels are detected using Eq. 14.2. Table 14.4 compares the average threshold for S and I images using both methods. The thresholds given by the original Kittler method show nonuniformity over different sets of cervix images and detect a very less number of SR pixels. This affects the accuracy of SR pixel detection. However, the approximate range of difference between TL and TR by the suggested modification (TR-TL = dynamic range of intensities of non-SR pixels) is constant for all four datasets under experimentation. This, in turn, increases the accuracy of SR as well as non-SR pixel detection.

Table 14.4 Average left-side and right-side threshold
Full size table

The SR is detected using the recommended modification explained in Sect. 14.3.2 (Fig. 14.4) that contemplates the fact that the SR pixels are heterogeneous regarding other image pixels and that they can be easily observed by the human eye. The performance of enhancement relying on SR detection for low-quality images corroborates the prerequisite of adaptability from Sect. 14.3.4. Figure 14.5b shows the detected SR pixels from the original low-quality image, whereas accurate SR detection from the enhanced high-quality image after the ASRDC application appears in Fig. 14.5c showing that the SR detection is effective to the adaptive enhancement of the low-quality image.

Fig. 14.4
figure 4

SR pixel detection

Full size image
Fig. 14.5
figure 5

Performance of enhancement-based SR detection for low-quality image

Full size image

4.3 SR Inpainting

The iterative mean color replacement from Sect. 14.3.2 takes away the detected SR pixels (refer to Fig. 14.6). The ASRDC method adaptively selects the SR detection tactic to be applied with or without input image enhancement based on the lightness measure of an image. The enhancement initiates automatically for low-quality images before applying the ASRDC technique.

Fig. 14.6
figure 6

SR Inpainting using mean color replacement

Full size image

4.4 Quantitative Evaluation of Proposed Method

Most of the reported literature spoke about visual comparisons of various SR detection and removal methods [3,4,5, 13, 14, 17]. Due to the unavailability of ground reality images captured with proper illumination, the quantitative evaluation is complicated.

However, the ASRDC system compares results utilizing NR image quality attributes of the original and inpainted image on the dataset under experimentation. Table 14.5 provides the average calculations of mean and standard deviation, respectively. It proves that an inpainted image has a low mean as compared to the original image that is to say SR (bright intensity pixels) are removed. The SR free image is homogeneous due to uniformity in image intensities and has less deviation from the mean, i.e., a decrease in the standard deviation. Table 14.6 illustrates the attributes of color image, viz., colorfulness, sharpness, and standard deviation. The SR removal decreases the colorfulness due to mean color replacement, makes the input image sharper, and decreases the proportion of distortion present in the input image, i.e., decrease in lightness. These observations depict a close agreement with the theoretical concepts of NR color image quality metrics.

Table 14.5 Comparison of statistical characteristics of original and inpainted image
Full size table
Table 14.6 Comparison of color image attributes of original and inpainted image
Full size table

4.5 Qualitative Analysis for State-of-the-Art Methods

Figure 14.7 illustrates the comparative visual difference between the ASRDC method and other state-of-the-art implementations aiming at SR detection and mitigation suggested in [9, 12, 14]. It follows that the novel ASRDC technique outsmarts the reported practices in terms of SR detection for images collected from different databases. Kudva et al. suggested the use of the Jaccard Index (JI) to measure the quantitative performance of SR detection techniques with manually marked SR pixels for images having practically visible SR pixels [9]. The JI value must be higher for the selected image to have accurate detection.

Fig. 14.7
figure 7

Comparative analysis of state-of-the-art methods for SR detection. (a) Original image; (b) SR detection with [12]; (c) SR detection using [14]; (d) SR detection via [9]; (e) SR detection using the ASRDC method

Full size image

Table 14.7 parallels outcomes for the maximum JC for the ASRDC scheme and other state-of-the-art techniques to validate the new approach. The present analysis considered only four images for the JC evaluation. However, the method can be extended for the entire database assessment. Recently, efforts relying on artificial intelligence (AI), data mining, and fuzzy-based methodologies addressed the SR detection issue [26, 27]. Health 4.0 protocols have also provided new insights regarding the usage of medical resources to handle various medical emergencies [28,29,30,31]. These lines of attack may lead to a revolution in CC detection and treatment.

Table 14.7 Qualitative analysis with state-of-the-art methods of SR detection
Full size table

5 Conclusions

This research puts forward the ASRDC as an adaptive method for detection and removal of SR from input cervix image, i.e., cervigrams. Experimentation was carried out on 612 images collected from NCI. The ASRDC method overcomes significant limitations of current SR detection techniques, i.e., dependency on shape and size of the kernel, selection of arbitrary constant, and every time training of the system. The ASRDC methodology uses the lightness as an NR quality measure to check the necessity of image enhancement, automatic enhancement of low-quality images before SR detection technique, and automatic selection of threshold by the MKM. Subjective and objective quality evaluation over different datasets highlights the ASRDC significance. Noise and resolution of the biomedical image largely depend on the quality of the equipment used for the capture and the skills of the expert (human intervention). In general, noise content and resolution of any biomedical images can be improved by using equipment that is more sophisticated. In addition to this, the ASRDC methodology will be an additional tool to enhance the grade of the biomedical images under study.

The inpainted images generated by the ASRDC adaptive system can be passed to further stages of early CC detection for additional feature extraction, segmentation, and classification techniques.

The authors are aware that when it comes to analyses of 3D images and 2D or 3D video, other shortcomings may affect SR detection as well as correction. For the cases when the dimensionality is high and several imaging modalities become necessary, soft computing strategies may lessen the processing time, help with more challenging settings, and work with other objective metrics [32,33,34,35,36,37,38,39,40,41,42]. It should be pointed out that SR detection and removal can benefit from the knowledge obtained in other similar image-processing tasks that share some characteristics and caveats with this problem.