1 Introduction

In medical science, proper monitoring of body function, physiological sign and diagnosis with better accuracy is a challenging task. Medical image processing plays a major role in early diagnosis of different disease. Medical imaging techniques like computed tomography (CT), magnetic resonance imaging (MRI) techniques are used by doctors for analyzing the pathological changes in the human body. The accuracy of the diagnosis and treatment depends on the quality of such images. But during acquisition and transmission, the image becomes blurred and has low contrast, [22] and complex due to presence of several overlapped objects. So, understanding of such images are difficult. This low contrast in the medical images allows the doctors to provide inaccurate judgment. But, for the reduction of deaths, the abnormality should be identified with accuracy. So, contrast improvement agents can be injected for enhancing such images, but these are not good for patients every time. So, before abnormality identification [24], it is important to make the image quality, features, and details more significant. To achieve this, medical image enhancement techniques are used by the Doctors for disease diagnosis, tumor detection and any type of abnormal detection.

Histogram Equalization (HE), [14] is a common technique for improving the contrast, by directly changing the pixel values. As the medical image has a large low gray background, So, the HE technique makes the entire image more bright [33]. As, the gray value changes to a large extent, it leads over enhancement and information loss. It causes artifacts in the enhanced images. It is also tough to improve the texture region using the HE technique alone. To modify the histogram and to find the better-enhanced image, Huang et al., [16] proposed an adaptive gamma correction technique with weighting distribution (AGCWD) function. But this technique is unable to preserve the edge details. The unclearness and low contrast is found in the medical image due to interference, noise, and low light environment. So an image enhancement method has been presented in the article, [22] by combining the both spatial domain and wavelet domain. In this method, after blurring (Sharpening), the upper-frequency part of the image is de-noised with the help of median filtering and the output image is formed by fusing all sub-images.

Another technique is proposed by Liu et al., [23] by partitioning the image into different frequency band sub-images. Then linear transformation is used in the low-frequency band and thresholding is used for de-noising the high-frequency bands. After that, all bands are reconstructed in spatial domains with the help of inverse transformation, and performance is measured in terms of entropy, and PSNR. The paper [21] presents an enhancement technique that adjusts the fractional-order as per the gradient feature of the image. It uses Otsu technique and fractional differential algorithm to enhance the image adaptively. A simple and adaptive enhancement technique is proposed by Kallel et al., [17] for dark medical images by using the concept of decomposition of image and correction factor. The author of this article claims supremacy over other techniques in terms of overall quality and visibility of local details. In modern diagnosis, medical images suffer from low-luminance and non-uniformity problems which affect the diagnosis process.

To generate a clear view of the images, a technique is presented in [36], which first enhances the visual perception with the help of luminance-level modulation and then improves the details using gradient modulation. Small and important details of the image may hide or distort because of blurring and noise in medical images. To restore the images, a new gradient-based plug-and-play algorithm has been presented in, [10]. The external denoiser is used in this paper with the help of a deep CNN, trained on gradient domain. A directional sparsely-sampled sinogram interpolation is implemented for the enhancement of the CT images in [20]. To minimize the radiation does in CT scanning, a sparsely-sampled sinogram is used in this article. For the reduction of artifact and to enhance image quality, an effective approach of sinogram interpolation method has been used which is based on directional information. The enhanced medical image suffers the problems of noise, over enhancement and information loss. So diagnosis is ineffective for diseases like a tumor. So a weighted constrained HE technique is proposed in [7] for the contrast enhancement of medical images which incorporates with Otsu’s threshold, Particle Swarm Optimization (PSO) and AGC. This algorithm focuses on preserving the information by selecting the proper threshold.

A method of enhancement technique is discussed in [34] to enhance the segmented liver image with the help of HE. This technique utilizes the segmentation concept with gray stretching to achieve the best result. For brightness preservation along with improving the contrast of brain images, an adaptive HE and fuzzy logic-based technique is proposed by Subramani et al. [32]. Initially, the method of fuzzy is applied to the input image and then the output image is associated with AHE technique. A memory augmented autoencoder approach is presented in [11] for detecting anomalies in IoT data, by using reconstruction errors. For suppressing the generalization ability of this model, a memory mechanism is used in this paper. And, time series data anomaly detection has been done with the help of memory-augmented autoencoder. Kandhway et al. [18] introduced a krill herd (KH) optimization-based framework for medical images which utilizes the fitness function and plateau limits to produce the best-enhanced image.

Many other enhancement techniques [1,2,3,4,5,6, 12, 13, 15, 27, 29,30,31] have also been developed to improve the contrast, to preserve the brightness, entropy, edge and reduce the artifacts. But few techniques are able to improve the contrast of the image but they don’t produce better result in terms of information contents. Few algorithms result better entropy but cannot produce satisfactory brightness preservation, peak signal to noise ratio, and structure similarity. Few techniques are not adaptive and results artifacts. But an enhancement technique is said to be effective, if the technique is adaptive and proper balance possible among these parameters in the enhanced images, better information content, AMBE, PSNR and SSIM. So, to achieve an adaptive and effective enhancement technique for medical images, following are the main contributions and novelty, which are discussed below.

  1. 1)

    This technique presents an optimal approach to find the texture region from medical images for suppressing the effect of artifacts in the enhanced image. Its main purpose is to preserve the edge details and eliminate artifacts effectively. The threshold for finding the texture region varies from image to image. So, threshold value is optimized using Directed Searching optimization (DSO) algorithm.

  2. 2)

    The main problem associated with image enhancement techniques is over enhancement. So, it is required to control the enhancement rates, else it causes more information loss. To resolve these issues, few authors utilize the parameters like mean, median [29, 30, 35] of the histogram for clipping the histogram.

But in this paper, an adaptive clipped threshold is used to clip the histogram for controlling the enhancement rate. This technique avoids the over enhancement and preserves the details during enhancement. The appropriate optimized threshold (Vth) is selected using DSO algorithm.

  1. 3)

    In AGCWD [16] technique, the exponential parameter (ϕ) is fixed for all the images. So it may or may not result better enhancement in all types of medical images. But, to make it more adaptive and to enhance the image quality more effectively, a tunable parameter(ϕ) is used in this paper along with a weighted probability density function (PDF). The value of ϕ is not fixed in this paper and it depends on type of images. The value of ϕ is optimized in such a way that it will improve the visual quality.

  2. 4)

    The thresholds parameters like textural threshold, clipping threshold, and gamma parameters are optimized using DSO algorithm for which, the fitness function is maximized automatically without human intervention. It will make the algorithm more adaptive.

  3. 5)

    The transfer function of the enhanced image is evaluated using modified optimal weighted PDF and optimal gamma parameters.

  4. 6)

    Performance of the proposed DSOTAGC technique is measured and compared with state-of-the-art techniques in terms of visual quality, information contents (entropy), AMBE, PSNR, and SSIM.

The remaining part of the paper is arranged as follows. Section 2 describes the details of the proposed technique including edge detection, clipping technique, AGC, DSO, fitness function, threshold selection. Section 3 represents the comparative analysis of the DSOTAGC technique with respect to existing technique. The paper is concluded in Section 4.

2 Proposed DSOTAGC technique

To improve the image quality such as visibility, information contents, brightness preservation and to reduce the artifacts, an adaptive enhancement technique is proposed in this article named as DSOTAGC technique. The proposed method is based on adaptive gamma correction (AGC) technique, adaptive clipping technique and DSO algorithm. As it is required to improve such parameters in the enhanced image, so the enhancement problem is presented as an optimization (maximization) problem in this article. So, an optimal algorithm like DSO is used in this paper. As DSO utilizes the benefit of both position updating and mutation, so it can produce better optimal results by preventing premature convergence in complex medical images. For improving the contrast, AGCWD is used with an adaptive gamma parameter. The rate of enhancement is controlled with the help of an adaptive clipped threshold. The proposed framework for image enhancement is shown in Fig. 1 and discussed below.

Fig. 1
figure 1

Proposed framework of DSOTAGC technique

Full size image
  1. a)

    As medical images are textural images, so, Sobel operator with an optimal textural threshold is used to find out the texture region.

  2. b)

    Compute the textured histogram.

  3. c)

    Clip the histogram using the adaptive clipping threshold.

  4. d)

    Find the transfer function of the enhanced image with the help of adaptive gamma parameter and modified PDF.

  5. e)

    Define the fitness function.

  6. f)

    Optimize the texture threshold, clipping threshold, gamma parameter using DSO algorithm to maximize the fitness function automatically.

2.1 Input the image and detect the texture region

The MRI and CT scan images are taken from Medpix [9, 25] and MRTIP [26] data base to improve their quality for early diagnosis. In general, medical images contain both non-texture and texture region. But the pixels of the non-texture region biased the enhanced images and produce artifacts in the enhanced image. So, the effect of the non-texture region should be minimized. For which, the texture region of each image is 1st identified in this article. For identifying the texture region, edge pixels of such images should be detected. This has been done with the help of Sobel operator and a textural threshold (Tsp) in this paper. It evaluates the gradient of each pixel. Here, the input image is represented by I. Intensity value of pixel at coordinate (x, y) is depicted by I(x, y). Two 3 × 3 masks have been used for measuring gradient at each pixels. The gradient ( Gx) in the x direction as,

$${G}_x=\left[\begin{array}{ccc}+1& 0& -1\\ {}+2& 0& -2\\ {}+1& 0& -1\end{array}\right]\times {I}_{x,\kern0.5em y}^{3\times 3}$$
(1)

Gradient (Gy) in y-direction as,

$$G_y=\begin{bmatrix}-\text{1}&-\text{2}&-\text{1}\\\text{0}&\text{0}&\text{0}\\+\text{1}&+\text{2}&+\text{1}\end{bmatrix}\times I_{x,\kern0.5emy}^{\text{3}\times\text{3}}$$
(2)

The gradient (Gx, y) at each pixel as,

$${\textrm{G}}_{\textrm{x},\textrm{y}}={\left({G}_x^2+{G}_y^2\right)}^{0.5}$$
(3)

Then, an image edge map Is(x, y) is constructed using the measured gradient at each pixel and a particular threshold Tsp.

$${I}_s\left(x,y\right)=\left\{\begin{array}{c}1,\kern0.5em {\textrm{G}}_{\textrm{x},\textrm{y}}\ge {T}_{sp}\kern0.5em \\ {}0,\kern0.5em {\textrm{G}}_{\textrm{x},\textrm{y}}<{T}_{sp}\end{array}\right.$$
(4)

It is considered as the reference image for finding the texture region. If the gradients value of that image is greater than that of the threshold, then these values are considered for texture histogram computation. The threshold for finding the texture region varies from image to image. So, after performing a lot of experiments, the range of threshold is considered in between 0 to 9, and the threshold is optimized with the help of DSO algorithm.

2.2 Histogram computing and clipping

After extracting the textural region, the histogram of the textural image is computed. Then an adaptive clipped threshold is evaluated for clipping the histogram. The main purpose of clipping threshold is to control the enhancement rate. This technique avoids over enhancement and preserves the details during enhancement. The appropriate optimized threshold (Vth) is selected using DSO algorithm. The value of this parameter is taken in between 0 to 1. The clipped normalized histogram is formed using Eq. (5).

$$Pd(k)=\left\{\begin{array}{c}{V}_{th},\kern1.75em if\ X(k)\ge {V}_{th}\kern2em \\ {}X(k),\kern0.5em if\ X(k)<\kern0.5em {V}_{th}\kern2.75em \end{array}\right.$$
(5)

In Eq. (5), the variables Pd and X represent the PDF of clipped histogram and PDF of the textured image respectively, whereas k is the index, varies up to maximum gray level of the image.

2.3 Evaluation of adaptive gamma parameter and transfer function of enhanced image

A weighting distribution function is used in this article to modify the statistical histogram and to form a weighted histogram. To enhance the image quality, a tunable parameter (ϕ) is used along with a weighted PDF function. The value of ϕ is not fixed in this paper and it depends on type of the images. The value of ϕ is selected within the range [0, 1] for maximizing the objective function. The weighted PDF (pdw) is computed as,

$${pd}_w(i)=\max (Pd)\times {\left(\frac{Pd(i)-\min (Pd)}{\max (Pd)-\min (Pd)}\right)}^{\phi }$$
(6)

In Eq. (6), Pd represents the PDF which is computed using the clipped histogram of textured image. Then, the weighted cumulative density function (CDF) is calculated using in Eq. (7). The adaptive gamma (𝛶) is evaluated using weighted CDF and presented in Eq. (8).

$${cd}_w(i)=\sum_{i=0}^{i_{max}}\frac{pd_w(i)}{\sum {pd}_w}$$
(7)
$$\textrm{Y}=1-{cd}_w(i)$$
(8)

Then the enhanced image is constructed using the adaptive gamma parameter (Υ), intensity value of the image (i) and maximum gray level (imax) of the image. The transfer function (T(i)) of the enhanced image is represented as,

$$T(i)={i}_{max}{\left(i/{i}_{max}\right)}^Y$$
(9)

2.4 Directed searching optimization (DSO) algorithm

Zou et al., [37] introduced DSO algorithm for solving constrained optimization problems. It is performed in two main steps, such as position updating and genetic mutation. Position updating is required for the convergence of the DSO. Mutation enhances the diversity of the individual and is a benefit for preventing premature convergence. A penalty function is adapted for balancing the constraint and objective and violations. It includes the parameters like, population size (PS), forward probability (Pα), iterations (iter), forward co-efficient (α), mutation probability (Pm), and backward co-efficient (β). Initialize the population matrix (PM), where the variable (DM) represents the dimension (DM) of the problem space.

$$PM=\left[\begin{array}{ccc}{P}_1^1& {P}_2^1& \cdots \kern0.5em {P}_{DM}^1\\ {}{P}_1^2& {P}_2^2& \begin{array}{cc}\cdots & {P}_{DM}^2\end{array}\\ {}\begin{array}{c}\vdots \\ {}{P}_1^{PS}\end{array}& \begin{array}{c}\vdots \\ {}{P}_2^{PS}\end{array}& \begin{array}{c}\begin{array}{cc}\ddots & \vdots \end{array}\\ {}\begin{array}{cc}\cdots & {P}_{DM}^{PS}\end{array}\end{array}\end{array}\right]$$
(10)

The position updation using DSO algorithm is shown in Fig. 2. In this figure, the ith component of jth solution vector of kth iteration is depicted by \({P}_i\kern-.5pt^j(k).\)The variable \({\textrm{P}}_i\kern-.3pt^{{\textrm{j}}_{\textrm{g}}}\left(\textrm{k}\right)\) represents ith component of jgth global solution vector of kth iteration and jg is the index of global solution vector. As per this strategy, \({P}_i\kern-.3pt^j\) is inclined to reach \({P}_i\kern-.5pt^{j_g}\), so the main searching region is considered as the region nearer to \({P}_i\kern-.3pt^{j_g}\). So the forward region is represented by PF. The backward region is shown by PB. Here, PiL and PiU are the lower and upper bound of ith component.

Fig. 2
figure 2

Position updating strategy of DSO

Full size image

To update the position of solution vector, the following conditions must be verified. The variable j lies in the range [1, PS] and j ≠ jg and i must vary in between [1 DM]. Then, the condition (rand () < Pα) must be checked. If this is satisfied then the position updating is carried out as per Eq. (11).

$${P}_V={\textrm{P}}_{\textrm{i}}\kern-.3pt^{\textrm{j}}\left(\textrm{k}\right)+\left(1+\upalpha \right)\left({\textrm{P}}_i\kern-.3pt^{{\textrm{j}}_{\textrm{g}}}\left(\textrm{k}\right)-{\textrm{P}}_{\textrm{i}}\kern-.3pt^{\textrm{j}}\left(\textrm{k}\right)\right)$$
(11)

In Eq. (11), the adaptive step size is represented by \({P}_i\kern-.3pt^{j_g}(k)-{P}_i\kern-.3pt^j(k)\). This step size becomes small when the solution vectors are nearer to each other and it is benefit for local search. But at first, the solution vectors are far away from each other in solution space, so the step size is more and is required for global search. Then the following conditions is required to check.

$$\textrm{If}\kern0.5em {P}_V>{P}_{iU}=>{P}_V={P}_{iU}$$
(12)
$$\textrm{If}\kern0.5em {P}_V<{P}_{iL}=>{P}_V={P}_{iL}$$
(13)

Then the new updated position is,

$${P}_i\kern-.3pt^j\left(k+1\right)={P}_i\kern-.3pt^j(k)+r\times \left({P}_V-{P}_i\kern-.3pt^j(k)\right)$$
(14)

If, rand () < Pα is not satisfied, then position is updated in backward direction.

$${P}_s={P}_j\kern-.3pt^i(k)-\left(\beta \right)\times \left({P}_i\kern-.3pt^{j_g}(k)-{P}_i\kern-.3pt^j(k)\right)$$
(15)

Again the following conditions are checked in backward direction

$$\textrm{If}\kern0.5em {P}_S>{P}_{iU}=>{P}_S={P}_{iU}$$
(16)
$$\textrm{If}\ {P}_S<{P}_{iL}=>{P}_S={P}_{iL}$$
(17)

Then the position in backward direction is updated using the equation,

$${P}_j\kern-.3pt^i\left(k+1\right)={P}_i\kern-.3pt^j(k)+r\times \left({P}_s-{P}_i\kern-.3pt^j(k)\right)$$
(18)

For improving the performance of the DSO algorithm, mutation operator is used after position updating. Then the mutation probability condition is checked. If rand () < Pm then the new updated position is,

$${P}_i\kern-.3pt^j\left(k+1\right)={P}_{iL}+r\times \left({P}_{iU}-{P}_{iL}\right)$$
(19)

If the population size is too small, DSO has few possibilities to find the optimal solution, as only a small part of the search space is explored. If the population size is more, then DSO slows down the convergence rate. So, it cannot be faster than a moderate sized population and results in wastage of resources. So, moderate size of iterations and population are taken and each of size 50 to reduce the complexity, to reduce the wastage of resources, and to find the best convergence performance.

Mutation enhances the diversity of individuals and exploits the unseen location of the searching space. But, a higher value of mutation probability may lead the solution to more diverge. So the probability of mutation is taken as 0.01. As the forward region is nearer the global position, so the forward coefficient and forward probability are considered as 1 and 0.8 respectively. The variable r is a random number that lies in between 0 to 1. The backward region is used for slowing down the convergence rate and backward co-efficient (β) is taken as 10. The upper and lower bound of ith component PiU and PiL depends on the threshold values taken in this paper.

2.5 Fitness function

In this paper, a fitness function is defined to select the optimum value of the tunable parameters using a heuristic optimization approach. It creates fitness criterions [7, 8, 18, 19] for the enhanced image. The diagnosis process becomes more complicated if the image is having less information contents and signal strength. So it is required to improve the entropy and PSNR. So, the fitness function taken in this paper, is a multi-objective function and represented as the summation of both entropy and PSNR. It depends on the constraint parameters used in this paper such as textural threshold, clipping threshold and optimal value of ϕ which are required for constructing the enhanced image. The defined fitness function is.

$$f={\beta}_1\kern.5pt{f}_1+{\beta}_2\kern.5pt{f}_2$$
(20)

Whereas, β1and β2 are the constants and their values represents the significance of individual objective function. Both objective functions are considered with equal importance, so the value of β1, β2 are taken as 0.5. As in the enhanced medical image, both entropy and PSNR are required to be improved, so the fitness function should be maximized and shown in Eq. (21).

$$J=\arg \left(\max (f)\right)$$
(21)

In Eq. (20), f1 represents the entropy or the average information contents and measured using Eq. (22). f2 is the PSNR of the enhanced image and evaluated using Eq. (23–24).

$${f}_1=H\left({I}_e\right)=-\sum\limits_{l=0}^{L-1}P(l)\kern2pt{\mathit{\log}}_2P(l)$$
(22)
$${f}_2= PSNR=10{\mathit{\log}}_{10}\left(\frac{{\left(L-1\right)}^2}{MSE}\right)$$
(23)

In Eq. (22), P(l) represents the PDF at intensity level l and the maximum gray level is presented by variable L. Whereas, the mean square error (MSE) is represented as,

$$MSE=\frac{1}{r\ast c}\sum_{i=1}^r\sum_{j=1}^c{\left({I}_o\left(i,j\right)-{I}_e\left(i,j\right)\right)}^2$$
(24)

Whereas, Io(i, j), and Ie(i, j) represents the original and enhanced image respectively.

3 Results and discussions

To verify the feasibility and robustness, the proposed method is applied to medical images (200 MRI and CT images) collected from various databases [9, 25, 26]. Then the performance of this method is investigated in both qualitative and quantitative approaches by comparing with existing methods including HE [14], Dualistic sub-image histogram equalization (DSIHE) [33], median mean based sub-image clipped histogram equalization (MMSICHE) [30], Exposure based sub-image histogram equalization (ESIHE) [29], AGCWD [16], dominant-orientation based texture histogram equalization (DOTHE) [31], triple clipped dynamic histogram equalization based on standard deviation (TCDHE-SD) [35] and Particle swarm optimization based enhancement technique [19]. MATLAB R2018a is used to perform the simulation work.

3.1 Performance Parameter

To evaluate the performance of the proposed DSOTAGC technique, parameters like entropy, PSNR, absolute mean brightness error (AMBE), and structure similarity index measure (SSIM) are taken into consideration.

Entropy is used to evaluate the information contents and the richness of the details within the image [8]. More value of the entropy [8] indicates better information contents within the images. It is measured using Eq. (22). As medical images are associated with noise, so it is required to suppress the noise. The performance index PSNR [1] is measured in this article to show the noise suppression quality. It is defined by the ratio of maximum signal power to distorted noise power, It is inversely proportional to MSE and measured using Eq. (23). Higher PSNR presents the better noise suppression capability.

The parameter SSIM [28] provides consistent image visual quality in terms of structural information. The SSIM value ranges from 0 to 1. Better is the similarity in enhanced image with respect to input image, when SSIM is nearer to 1. It considers three terms, related to luminance, contrast, and structure. It is measured using Eq. (25).

$$SSIM\left(x,y\right)=\frac{\left(2{\mu}_x{\mu}_y+{C}_1\right)\left(2{\sigma}_{xy}+{C}_2\right)}{\left({\mu_x}^2+{\mu_y}^2+{C}_1\right)\left({\sigma_x}^2+{\sigma_y}^2+{C}_2\right)}$$
(25)

Whereas μx, μy are the mean intensity, σx, σy are the standard deviations of gray-scale images x and y respectively. The sample correlation coefficient between x and y is represented by σxy . And c1 and c2 indicate the constant and are used here to avoid instability. Again to measure the brightness preservation, AMBE is measured using Eq. (26). It represents mean intensity difference between input and enhanced images. Minimum value of AMBE indicates better enhancement and brightness preservation [30]. Whereas, in Eq. (26),  mean(Io) and mean((Ie) represent the mean intensity value of the input and enhanced image.

$$AMBE=\left| mean\left({I}_o\right)- mean\Big(\left({I}_e\right)\ \right|$$
(26)

3.2 Performance analysis

Because of the complexity problem in medical imaging, few existing techniques cannot produce satisfactory results. The rate of change of information is very rapid from region to region because of the overlapping of tissues in medical imaging. So the proper processing with satisfactory results is a challenging task in medical imaging. So, to achieve better enhancement effect, DSOTAGC technique is presented in this paper, whose simulated results are shown in Figs. 3, 4, 5, 6, 7, 8, 9 and 10 Then, qualitative and quantitative approaches are used to evaluate its performance. The qualitative approach comprises visual quality and histogram analysis, whereas quantitative approach involves the performance parameters like entropy, PSNR, SSIM, and AMBE.

Fig. 3
figure 3

Image-1 and enhanced by enhancement techniques

Full size image
Fig. 4
figure 4

Image-2 and enhanced by enhancement techniques

Full size image
Fig. 5
figure 5

Image-3 and enhanced by enhancement techniques

Full size image
Fig. 6
figure 6

Image-4 and enhanced by enhancement techniques

Full size image
Fig. 7
figure 7

Image-5 and enhanced by enhancement techniques

Full size image
Fig. 8
figure 8

Image-6 and enhanced by enhancement techniques

Full size image
Fig. 9
figure 9

Histogram of input and enhanced image-1

Full size image
Fig. 10
figure 10

Average plot (a) Entropy, (b) PSNR (c) AMBE, (d) SSIM

Full size image

The original medical images are shown in Figs. 3, 4, 5, 6, 7 and 8, in which organs are not clearly visible. HE and DSIHE technique improves the contrast, but some details are over-enhanced and look worse than the original image. Figure 9 represents the histogram of image-1 and its enhanced image. It indicates, how the shape of enhanced image is similar with input image. From these figures, it is cleared that, by HE, DSIHE, ESIHE, techniques, the gray distribution of the images is fully changed, which affects the visual quality of the resultant image. MMSICHE technique results a better enhanced image in terms of AMBE which indicates better brightness preservation. But it cannot produce satisfactory information contents within the enhanced image. And the resultant image is not much clear.

DOTHE technique enhances the edges but produces noise in the enhanced image. In DOTHE algorithm, some objects can be easily observed, but quality of the image is not good. AGCWD technique enhances image but the edge contents is not preserved. As it is not adaptive and results information loss, So it is not suitable for all types of medical images. TCDHE-SD technique also produces better quality image but not perform better in terms of similarity with respect to original image. TCDHE-SD technique is only based on histogram division and doesn’t use any parameter to improve the edge contents in the enhancement image. PSO-based technique improves the edge details and entropy better as compared to other techniques as it utilizes the fitness function as the summation of entropy and edge contents.

But the proposed DSOTAGC method preserves the image edges and produces better enhancement effect than other existing methods. It also produce optimal results by making a balance among information contents and PSNR. It is because of the equal weightage of each objective function defined in the fitness function (Eq. (20)). The enhancement of the contrast, preservation of brightness, and entropy is also better in the proposed technique over other techniques. The better contrast in the enhancement image is because of the adaptive gamma parameter. As initially, Sobel operator is used to determine the edges, then the non-texture effect is suppressed, so it results better preservation of edges in the enhanced images.

Existing methods like DOTHE, AGCWD show the effect of over-enhancement and are evident from Figs. 3, 4, 5, 6, 7, 8 and 9. But, an adaptive clipped threshold is proposed in this paper for controlling the enhancement rate. So, proposed technique results better in terms of over enhancement and controlling the enhancement rate. Proposed method is adaptive because the textural threshold, clipped threshold, and adaptive gamma value have been optimized with the help of DSO algorithm.

This proposed algorithm results better adaptive-ness because of both position updating strategy and mutation of DSO algorithm. So the pre-matured convergence has been prevented and produces the best optimal result. From Figs. 3, 4, 5, 6, 7 and 8, it is also observed that, the artifact content is more in HE, DSIHE, ESIHE, AGCWD, DOTHE, TCDHE-SD, and PSO based technique. It is because of non-texture pixels to the enhanced image and over enhancement. As over enhancement is controlled and non-texture effect is minimized, so proposed technique results artifact free images.

The measured entropy and PSNR values are presented in Table 1 and 2 respectively. It is observed that the proposed DSOTAGC method results better information content and PSNR over other methods. Low entropy is observed in HE technique, as it loses much information because of over enhancement. The proposed DSOTAGC technique yields entropy, nearly equal to that of input images shown in Table 1. It indicates, information content is preserved by the proposed method.

Table 1 Measured Entropy
Full size table
Table 2 Measured PSNR, AMBE, SSIM
Full size table

Higher PSNR obtained by DSOTAGC method results reduction in noise components and produce an artifacts free images. So, the resultant image is of pleasant quality with improved contrast. The measured AMBE value is shown in Table 2 and it is noticed that the brightness error is less in the proposed method over other discussed techniques. It represents, the brightness is better preserved by our proposed technique, and over enhancement is also controlled. The enhancement technique is more suitable, if SSIM is very close to one. From Table 2, it is noticed that proposed DSOTAGC technique results better SSIM as compared to existing techniques. So shape of the histogram of the enhanced image obtained by proposed method is very similar to the histogram of the original image, shown in Fig. 9. It is because of better contrast, texture as well as edge enhancement.

To verify the performance of the enhancement techniques, the experiment has been performed using MRI and CT images, but only few images are shown in this paper. The average values of entropy, PSNR, AMBE, and SSIM have been measured using all two hundred medical images and shown in Fig. 10a-d. After performing this experiment, it has been observed that, the proposed DSOTAGC technique results better average entropy, PSNR, AMBE, and SSIM, over other existing techniques. So, from the qualitative analysis (Figs. 3, 4, 5, 6, 7, 8 and 9) and quantitative analysis (Table 1 and 2), average plot (Fig. 10), it is observed that, proposed DSOTAGC based technique outperforms in terms of the visual quality, brightness, information, edge preservation, noise suppression, and structure similarity.

Because of its adaptive-ness and better performance, the proposed DSOTAGC technique can be used as a preprocessing technique for medical imaging applications such as coronavirus detection, brain tumor detection, brain tissue segmentation and monitoring the disease in the human body.

4 Conclusions

In this paper, a novel medical image enhancement technique is proposed with the objectives of improving the contrast, entropy, PSNR and making the image artifacts free. The most challenging task in the image enhancement technique is removing the artifacts along with preserving image texture. So, initially, texture region from the image has been identified using Sobel operator and suppressed the non-texture effect. The problem of the over enhancement in medical images is eliminated using adaptive clipping technique. To make the algorithm adaptive, the texture threshold, clipping threshold, and exponent parameters have been optimized using DSO algorithm. As, the fitness function is a combination of entropy and PSNR and each having equal importance, so DSOTAGC technique produces a balanced adaptive optimal result with better entropy and PSNR value. Contrast is improved with the help of adaptive gamma parameters. Further, the proposed approach is not only good for artifact suppression but also better for structure similarity. After performing more number of experiments, it is found that the DSOTAGC technique achieves the best enhancement effect by enhancing the details of the non-contrast medical images, improving the contrast and avoiding artifacts. Overall quality of the image and visibility of the local details by the proposed DSOTAGC technique is better than the other discussed techniques.