Denoising and contrast-enhancement approach of magnetic resonance imaging glioblastoma brain tumors

Abstract.

We investigate a new preprocessing approach for MRI glioblastoma brain tumors. Based on combined denoising technique (bilateral filter) and contrast-enhancement technique (automatic contrast stretching based on image statistical information), the proposed approach offers competitive results while preserving the tumor region’s edges and original image’s brightness. In order to evaluate the proposed approach’s performance, quantitative evaluation has been realized through the Multimodal Brain Tumor Segmentation (BraTS 2015) dataset. A comparative study between the proposed method and four state-of-the art preprocessing algorithm attests that the proposed approach could yield a competitive performance for magnetic resonance brain glioblastomas tumor preprocessing. In fact, the result of this step of image preprocessing is very crucial for the efficiency of the remaining brain image processing steps: i.e., segmentation, classification, and reconstruction.

1. Introduction

Magnetic resonance imaging (MRI) modality could be considered as the main modality dedicated to providing brain tissue images. Such modality uses the interactions between electromagnetic field and nuclear spin to map the spatial location and associated properties of specific nuclei or protons, based on the principles of nuclear magnetic resonance.¹ MRI is attracting more attention for brain tumor diagnostics since it provides highly detailed images and information about the localization, the shape, and the size of the tumor, without exposing the patient to high ionization radiation as is the case with other imaging modalities.² In spite of the great advantages of MRI, compared to other imaging technologies, it is still affected by noise from different origins, including the physiological processes’ noise, the stochastic variation, and the artifacts.³ The main noise in MRIs is due to the thermal noise introduced during data acquisition. It is produced by the stochastic motion of free electrons in the radiofrequency (RF) coil and by the eddy current losses in the patient, which are inductively coupled to the RF coil. The variance of thermal noise can be calculated as the sum of noise variances from independent stochastic processes representing the body and the electronics.²

In general, during the MRI acquisition task, there are trade-offs between acquisition speed, temporal resolution, and the signal-to-noise ratio (SNR) which affect the MRI quality. In fact, if the acquisition time is reduced and resolution is increased, both SNR and contrast will degrade. One way to improve MRI quality is to increase acquisition time and the temporal resolution. Such solution is not practical due to patient comfort and technical limitations.⁴ Therefore, preprocessing approaches of denoising and contrast enhancement for MRI data are becoming a point of great interest. Noise and low contrast have a direct impact on the extraction of regions of interest (ROIs) and delineation of different tumor regions, including the enhancing, nonenhancing tumor, necrotic, and edema, and normal brain tissues. Moreover, they not only affect the clinical diagnosis’ accuracy but also disrupt the operation of computer-aided diagnosis in different tasks, such as registration, classification, and especially the ROI extraction. In fact, most of the segmentation and classification algorithms, described in the literature, suffer from sensitivity to intensity inhomogeneities, low contrast, and noise.⁴

Histogram equalization (HE)⁵ is the main classical technique applied for contrast enhancement in the spatial domain. HE is based on better distribution of the pixel’s intensity value over its dynamic range. HE has a tendency to alter the brightness of the processed image causing the generation of nonexisting artifacts. Nevertheless, HE provides good performance for ordinary images; it may provide, in some cases, an over-contrast enhancement of MRIs, especially when the white matter hyperintensity (WMH) signal of the tumor core is high, which could be considered mistakenly as cerebrospinal fluid (CSF). HE-based approaches have gained research interest in the past decade due to their simplicity of use and their performance on almost all types of images. Among them, the adaptive histogram equalization (AHE),⁵ where mapping is applied differently for each pixel, is adaptive to the local distribution of pixel intensities rather than to input image global information. AHE has a tendency to have an excessive contrast enhancement particularly in homogeneous regions. Hence, contrast-limited adaptive histogram equalization (CLAHE) was proposed⁶ to overcome this limitation by providing a clip limit factor. Moreover, various HE-based approaches are proposed in the literature to overcome the aforementioned problem such as brightness-preserving bi-histogram equalization (BBHE)⁷ and brightness preserving dynamic fuzzy histogram equalization (BPDFHE).⁸ In the BBHE approach, the computed image’s histogram is separated into two parts based on the image-mean brightness value. The first part will have the range from minimum gray level to the mean value and the second part from the mean to the maximum gray level. Next, the two histograms of the separated subimages are independently equalized. It has been proved experimentally and mathematically that BBHE can preserve the original brightness to certain limits.⁴ Unlike BBHE, the BPDHE approach partitioned the image histogram, after applying a smoothing Gaussian filter on it, on the basis of its local maximums rather than on its mean. Next, HE mapping is applied independently for each partition. The changes in the dynamic range and the HE process will certainly alter the mean brightness of the output image. Hence, a normalization step is performed on the processed image. Usually, BPDHE could successfully enhance the input image without undesirable effects while maintaining its mean brightness.⁸ In 2017, Isa et al.⁹ proposed a new contrast-enhancement approach, known as “average intensity replacement based on adaptive histogram equalization” (AIR-AHE), for WMH brain MRI using only fluid-attenuated inversion recovery (FLAIR) sequences. This approach combines several established techniques to enhance the contrast of WMH, such as contrast stretching technique followed by CLAHE performed using a prefixed clipping limits value. Finally, a mapping, using a $3\times 3$ sliding window neighborhood operation, is applied for the computation of the new intensity value. A new hybrid contrast-enhancement approach, for computerized tomography (CT-scan) images, was proposed by Ref. 10, which combines contrast stretching and minimum mean brightness error bi-histogram equalization, followed by median filter (MF). This approach ameliorates the contrast of CT-scan images while enhancing the existing impulse noise. Based on only qualitative results, it has shown its good performance when compared to others stand-alone methods. Various contrast stretching-based preprocessing approaches are proposed in the literature, such as ordinary gamma correction¹¹ where the correction factor is set manually. This latter was not efficient for some types of images and does not give the appropriate contrast enhancement for original images. Rahman et al.¹² proposed an adaptive gamma correction for contrast enhancement where the correction parameter was computed automatically based of the original contrast of the image.

Numerous MRI denoising methods have been proposed in the literature. These denoising approaches could be classified into three main groups: filtering-based approaches, transform domain-based approaches, and statistical based approaches.

In filtering approach, the linear and nonlinear filters are used to denoise the MRIs. Linear filtering techniques include spatial and temporal filters. Nonlinear filtering techniques include the anisotropic diffusion filter (ADF),¹³ nonlocal means (NLM) filter,¹⁴ and the combination of domain and range filters (including the bilateral and the trilateral filters).¹⁵ In the transform domain approaches, the transforms such as wavelet, curvelet, and contourlet transforms are employed to denoise MRIs.²

In the statistical approach, mainly the linear minimum mean square error (LMMSE),¹⁶ the estimation of noise based on maximum likelihood approach,¹⁷ Markov random process, phase error estimation,¹⁸ nonparametric estimation, and singular function analysis have been proposed.¹⁹ In Ref. 20, the authors proposed a combination of two filters, the noise-driven ADF and the local LMMSE filter, for denoising MRI corrupted with Rician noise. In this method, the standard deviation of noise as well as the filter parameters is automatically estimated on the basis of a robust estimation of statistical noise parameters. Over the past several years, numerous wavelet-based noise reduction techniques have been proposed for medical imaging; some of these works combine wavelet transform with filtering techniques. In Ref. 21, the authors proposed a denoising approach based on the combination of wavelet transform and ADF, in order to reduce the pseudo-Gibbs artifacts generated due to zeroing some of the transform coefficients. In Ref. 22, the authors proposed a undecimated wavelet transform based on bilateral filtering technique to remove Gaussian noise in MRIs. Better results are obtained confirming that this method could effectively preserve the original image’s relevant features. The authors later extended their work for denoising magnitude MRIs corrupted by Rician noise in Ref. 23. However, the good performance of wavelet-based denoising approaches suffers from some drawbacks related to the estimation of local noise, using the wavelet decomposition, which results in edges blurring and loss of relevant information details.

In relation to brain tumor’s characterization, high-grade glioblastoma tumor could include edema zone, necrotic/cystic zone, and even enhancing and nonenhancing tumor. Such zones could share approximately the same intensity profile with other normal regions.¹

There are different approaches proposed for brain glioblastoma’s MRI denoising. Some of them are adapted from ordinary image denoising approaches, whereas others are dedicated to deal with the nature of a specific type of noise in MRI. Some authors are convinced that MRI should be treated differently^2–4 since noise in MRI does not follow the standard Gaussian assumption but a more complex Rician distribution.⁵

Deep learning-based approaches have been implemented widely in medical image processing due to their flexible design and efficient results.^24,25 However, there are some drawbacks to these approaches. For instance, batch sizes should be chosen carefully,²⁶ and training datasets should include many number of images with variations to obtain a robust method. Moreover, training of a deep neural network is a time-consuming task.

In this paper, we propose a denoising and adaptive contrast-enhancement approach dedicated to MRI glioblastoma brain tumors to improve the relevant image contents through reducing the noise while preserving the actual detailed features. The proposed approach is based on combined denoising technique, using a bilateral filter (BF), and contrast enhancement, using a contrast stretching technique. For the latter, we used the multiplication of input voxels by a contrast stretching factor computed automatically for each processed image depending on the image’s statistical information. The following are the main advantages of the proposed approach:

• -It enhances the contrast of each glioma brain MRI automatically, according to its statistical information.
• -It denoises to carry out the hidden parts of the acquired MRIs while preserving tumor edges and input image brightness and its useful features.

The remainder of this paper is arranged as follows. Section 2 describes in details the proposed approach. Section 3 discusses, evaluates, and compares the performance of the proposed approach with some recent existing approaches reported in the literature. Sec. 4 is dedicated to the results and the performance evaluations. Sec. 5 is devoted to the discussion. Finally, Sec. 6 draws several perspectives on this work.

2. Proposed Approach for Denoising and Contrast Enhancement

The proposed approach has mainly two objectives: the noise removal and the contrast enhancement when applied for brain glioblastoma tumors through MRI. Such process could be considered as a challenging task. Frequently, denoising process could lead to loss of some details and edges in images that are very relevant and very crucial for the performance of the remaining different postprocessing methods applied to MR data, such as segmentation and registration. On the other hand, it is also challenging to avoid over-contrast enhancement while reducing noise in uniform regions and preserving the actual image’s detailed features. To overcome such drawbacks, the proposed approach combines several established techniques for denoising and contrast enhancement based on optimal parameter selection and optimization. Since the glioblastoma regions are hyperintense in T1G modality, we choose this latter modality as an input image to test the performance of the proposed approach. Skull stripping operation^27,28 is absolutely required as a preprocessing step, in order to avoid interference during the detection of similar intensity between tumoral zone and nontumoral zone and to provide better results in contrast enhancement. The overall flowchart of the proposed approach is illustrated in Fig. 1.

Fig. 1.

The flowchart of the proposed algorithm for contrast enhancement.

The basic steps of the proposed approach are as follows:

• 1.load the skull-removed MRIs;
• 2.apply BF to the original skull-removed MRIs;
• 3.compute the statistical parameters (mean and variance);
• 4.compute the contrast stretching parameter (“$C$”) on the basis of the filtered image mean value;
• 5.apply the contrast stretching technique using the filtered image;
• 6.get output preprocessed image.

2.1. Bilateral Filter

BF is an edge-preserving filter widely used for denoising medical image since it removes noise without affecting edges and textural features.²⁹ It effectively filters noise in uniform intensity’s areas (white matter, gray matter, and CSF) while preserving the different glioblastoma zones (tumor, necrosis, and even edema), contours, and structures.³⁰ It has proven its efficiency and its simple formulation contributes to its popularity. Such filter does not involve the solution of partial differential equation and could be implemented in a single iteration. The BF could be considered as a combination of two Gaussian filters: spatial filter $g_{{\sigma }_s}$ and range filter $g_{{\sigma }_r}$. The filtered image $I_B$ is obtained by replacing the intensity value of each pixel with an average weighted value using the geometric and the photometric similarities between the pixels within a spatial window $\mathrm{\Omega }$. Filtered image $I_B$ of an original image called $I_{\text{in}}(x,y)$, is given as³¹

$$I_B=\frac{1}{W_n}\times \sum _{x_i\in \mathrm{\Omega }}\sum _{y_j\in \mathrm{\Omega }}{I}_{\text{in}}[(x-x_i),(y-y_j)]\times g_{{\sigma }_s}(x_i,y_j)\times g_{{\sigma }_r}[I_{\text{in}}(x_i),I_{\text{in}}(y_j)],$$ (1)

where

• -$W_n=\sum _{x_i\in \mathrm{\Omega }}\sum _{y_j\in \mathrm{\Omega }}{g}_{{\sigma }_s}(x_i,y_j)\times g_{{\sigma }_r}[I_{\text{in}}(x_i),I_{\text{in}}(y_j)]$ represents the normalization factor whose objective is to preserve image energy.
• -$g_{{\sigma }_s}=1/2\pi {\sigma }_s^2\times \exp [-(x^2+y^2)/2{\sigma }_s^2]$ represents the spatial Gaussian filter and ${\sigma }_s$ corresponds to its standard deviation.
• -$g_{{\sigma }_r}=1/\sqrt{2\pi }{\sigma }_r\times \exp \{-{[f(x_i,y_j)-f(x,y)]}^2/2{\sigma }_r^2\}$ represents the range Gaussian filter and ${\sigma }_r$ corresponds to its standard deviation.

2.2. Contrast Stretching

The contrast transformation proposed in this research is based on the contrast stretching process given in Eq. 2, where $I_B$ and $I_{\text{out}}$ represent, respectively, the filtered input image using the BF and the output intensity image value:

$$I_{\text{out}}=I_B\times C.$$ (2)

The automatic computation of the parameter $C$ could be considered as a challenging task due to the corrupted-noise-level variability of MRI scans. For the proposed preprocessing approach, we proposed a new automatic technique to compute the scalar $C$ on the basis of the filtered image’s statistical information, mainly the mean (${\mu }_{I_B}$) and the variance ($V_{IB}$).

Depending on the MRI brain glioblastoma tumors’ image contrasts, we could divide it into two principle classes: the high contrast class (K1), which regroups the high-grade (HG) tumors, and the moderate or dark contrast class (K2), which contains the low-grade (LG) tumors. In the (K1) class, images have a good quality and the tumors appear on hyperintense signal due to the absorption of the contrast product. Then, the main target is to preserve the image quality. The transformation curves for different values of the mean (${\mu }_{I_B}$) and the variance ($V_{IB}$) are very close to a line ensuring a little change in contrast and offer a little intensity changes. For this reason, we could define the contrast stretching parameter ($C$) as

$$C=\frac{{\mu }_{I_B}+2\times V_{IB}}{{\mu }_{I_B}-2V_{IB}}\text{if}{I}_{\text{in}}\in (K1).$$ (3)

In the (K2) class, the tumors appear on moderate or dark contrast. For this reason, our target is to increase the contrast in order to better distinguish the tumor. Hence, the contrast stretching parameter ($C$) could be defined as

$$C=2\times \frac{{\mu }_{I_B}-2\times V_{IB}}{{\mu }_{I_B}+2\times V_{IB}}\text{if}{I}_{\mathrm{in}}\in (K2).$$ (4)

The tumor intensity in this class is clustered in a small range of dark gray levels around the brain tissues’ mean. To increase the contrast, the transformation should spread out the dark or moderate intensities to higher intensities. We desire, then, to distribute the intensities over wider ranges.

3. Materials and Evaluation Metrics

The experimental results were assessed on the basis of approved evaluation metrics using the BraTS 2015 Benchmark.³² Such database offers the opportunity to researchers to compare their results using the same data. In this section, we present the used datasets as well as the evaluation metrics used to evaluate the performance of the proposed preprocessing approach.

3.1. Multimodal Brain Tumor Segmentation Benchmark

The BraTS 2015 data have been organized by Menze, Reyes, Farahani, and Kalpathy-Cramer, in conjunction with the MICCAI 2015 conference. This publicly available training dataset has been usually used to compare the existing segmentation approaches. In the present study, we propose to use this database to compare preprocessing approaches. The coregistered, the skull-stripped, and the annotated training datasets and the testing dataset are available via the Virtual Skeleton Database.³³ The dataset is represented as signed 16-bit integers, but only positive values are used. For each case, four MRI modalities have been proposed: T1, T2, T1c, and FLAIR. In this work, only the T1c modality is used as an input for the proposed preprocessing approach. For this purpose, we select 20 high grade and 10 low grade glioblastomas to validate the proposed approach.

3.2. Evaluation Metrics

Image enhancement and visual qualities improvement could be considered as a subjective problem. For this reason, evaluation metrics are established to illustrate the impact of enhancement methods on the original image. For further performance evaluation and to measure the effectiveness of our proposed approach when compared to other existing ones, different image quality measurement metrics are used: mean square error (MSE), peak signal-to-noise ratio (PSNR), edge preservation index (EPI), absolute mean brightness error (AMBE), structural similarity index measurement (SSIM), and Dice metric (DM) to study the impact of the proposed approach on the segmentation results.

3.2.1. Mean square error

MSE measures the cumulative squared error between the preprocessed image $I_{\text{out}}$ and the acquired one $I_{\text{in}}$. MSE is calculated as

$$\mathrm{MSE}=\frac{1}{M\times N}\times \sum _{i=0}^{M-1}\sum _{j=0}^{N-1}{\Vert I_{\text{in}}(i,j)-I_{\text{out}}(i,j)\Vert }^2,$$ (5)

where $M\times N$ represents the size of the images. A lower MSE value means lesser errors.

3.2.2. Peak signal-to-noise ratio

PSNR is an assessment of image quality that measures how much the enhanced image has been degraded when referred to the acquired image. A higher PSNR value means less noise, indicating higher image quality. PSNR is calculated according to the equation below, where $I_{\text{out}}^{\max }$ is the maximum intensity value in the output image ($I_{\text{out}}$) and MSE is the mean square error explained in the previous subsection.

$$\mathrm{PSNR}=10\times {\mathrm{Log}}_{10}\left(\frac{I_{\text{out}}^{{\max }^2}}{\mathrm{MSE}}\right).$$ (6)

3.2.3. Edge preservation index

EPI is used to calculate the edge-preserving ability.³⁴ It is calculated as

$$\mathrm{EPI}=\frac{\sum (\mathrm{\Delta }{I}_{\text{in}}-\overline{{\mathrm{\Delta }I}_{\text{in}}})*\sum (\mathrm{\Delta }{I}_{\text{out}}-\overline{\mathrm{\Delta }{I}_{\text{out}}})}{\sqrt{\sum {(\mathrm{\Delta }{I}_{\text{in}}-\overline{\mathrm{\Delta }{I}_{\text{in}}})}^2*\sum {(\mathrm{\Delta }{I}_{\text{out}}-\overline{\mathrm{\Delta }{I}_{\text{out}}})}^2}},$$ (7)

where $\mathrm{\Delta }{I}_{\text{in}}$ and $\mathrm{\Delta }{I}_{\text{out}}$ are, respectively, high-pass filtered versions of the image $I_{\mathrm{in}}$ and $I_{\text{out}}$, obtained with a $3\times 3\text{pixel}$ standard approximation of the Laplacian operator. The larger value of EPI means more ability to preserve edges.

3.2.4. Absolute mean brightness error

AMBE is an assessment of brightness preservation. It evaluates the ability of the contrast enhancement methods in maintaining the mean brightness of the original image. AMBE computes the absolute mean brightness difference between the acquired image $I_{\text{in}}$ and the preprocessed one $I_{\text{out}}$ as

$$\mathrm{AMBE}=|I_{\text{in}}-I_{\text{out}}|.$$ (8)

3.2.5. Structural similarity index measurement

SSIM is a measure indicating the ability to preserve details and structures of interest between acquired and preprocessed images.³⁵ The result of SSIM lies between $-1$ and 1. SSIM is used to compare the luminance, the contrast, and the structure of two different images. Given an acquired image $I_{\text{in}}$ and a preprocessed image $I_{\text{out}}$, both of size $M\times N$, the SSIM is defined as

$$\mathrm{SSIM}(I_{\text{in}},I_{\text{out}})=l(I_{\text{in}},I_{\text{out}})\times c(I_{\text{in}},I_{\text{out}})\times s(I_{\text{in}},I_{\text{out}}),$$ (9)

where

$$l(I_{\text{in}},I_{\text{out}})=\frac{2\times {\mu }_{I_{\text{in}}}\times {\mu }_{I_{\text{out}}}+C1}{{\mu }_{I_{\text{out}}}^2+{\mu }_{I_{\text{in}}}^2+C1},$$ (10)

$$c(I_{\text{in}},I_{\text{out}})=\frac{2\times {\delta }_{I_{\text{in}}}\times {\delta }_{I_{\text{out}}}+C2}{{\delta }_{I_{\text{out}}}^2+{\delta }_{I_{\text{in}}}^2+C2},$$ (11)

$$s(I_{\text{in}},I_{\text{out}})=\frac{{\sigma }_{I_{\text{in}}{I}_{\text{out}}}+C3}{{\sigma }_{I_{\text{in}}}\times {\sigma }_{I_{\text{out}}}+C3},$$ (12)

• -$l(I_{\text{in}},I_{\text{out}})$ represents the luminance comparison function. It measures the closeness of the input and the output images’ mean luminance (${\mu }_{I_{\text{in}}}$ and ${\mu }_{I_{\text{out}}}$). This factor is equal to 1 only if (${\mu }_{I_{\text{in}}}={\mu }_{I_{\text{out}}}$);
• -$\mathrm{c}(I_{\text{in}},I_{\text{out}})$ represents the contrast comparison function. It measures the contrast closeness of the input and the output images. Here, the contrast is measured by the standard deviation ${\sigma }_{I_{\text{out}}}$ and ${\sigma }_{I_{\text{in}}}$. This term is maximal and equal to 1 only if (${\sigma }_{I_{\text{out}}}={\sigma }_{I_{\text{in}}}$);
• -$s(I_{\text{in}},I_{\text{out}})$ represents the structure comparison function. It measures the correlation coefficient between the original and the enhanced images. Here ${\sigma }_{I_{\text{in}}{I}_{\text{out}}}$ is the covariance between the input image and the output image.

3.2.6. Dice metric

The DM computation is usually used for quantitative segmentation evaluation and comparison. This metric is used in order to compute the similarity between the obtained segmentation result and the expert segmentation result.³⁶ DM could be given as

$$\mathrm{DM}=\frac{2\times V_{se}}{V_s+V_e},$$ (13)

where $V_s$, $V_e$, and $V_{se}$ are, respectively, the areas/volumes of the obtained segmented region, the expert segmentation region, and the intersection between them. DM is always in [0, 1], with 1 indicating a perfect match between the two segmentations. The higher DM could attest the performance of the proposed approach.

4. Results and Performance Evaluation

In this section, we will evaluate the performance of the proposed preprocessing approach for denoising and contrast enhancement when applied for MRI glioblastoma brain tumors. A comparative study has been performed using four recent state-of-the-art contrast enhancement methods. We first evaluate the preprocessing metric evaluation (MSE, PSNR, EPI, AMBE, and the SSIM) and then evaluate the DM parameter, in order to illustrate the impact of the preprocessing on the segmentation process. All compared methods have been performed on T1G modality sequence since it is the best MRI modality to observe the hypointense part of the glioblastoma tumor core.

4.1. Bilateral Filter Validation

In this subsection, we will test the efficiency of the BF, when applied to the processed datasets, through quantitative and qualitative evaluations. A comparative study has been performed between the BF and three usual denoising filters mostly used for MRI data: ADF, guided filter (GF), and MF.

For the validation results, we used the BF with the following setting parameters: the spatial domain standard deviation ${\sigma }_s=3$, the intensity domain standard deviation ${\sigma }_r=0.1$, and the spatial window $\mathrm{\Omega }=3$. Table 1 provides the obtained performance of applied BF when compared to the above-mentioned three denoising filters using the metrics MSE, PSNR, and EPI. One could notice that the BF performs better, giving the lowest MSE and the highest PSNR and EPI values compared to tested MRI denoising filters. The obtained performance confirmed that the BF preserves edges better than ADF and does not degrade the quality of the filtered output image through edges blurring.

Table 1.

Computed measurement metric for ADF, GF, MF, and BFs.

	MSE	PSNR	EPI
BF	$\mathbf{1.0}\times {\mathbf{10}}^{-\mathbf{4}}$	39.81	0.75
ADF	$1.2\times {10}^{-3}$	29.13	0.50
GF	$1.7\times {10}^{-4}$	37.74	0.72
MF	$1.4\times {10}^{-4}$	38.66	0.70

4.2. Preprocessing Methodology Evaluation

In this section, we report the qualitative and quantitative results of the proposed preprocessing approach over both HG and LG cases from BraTS2015 datasets. We also perform a comparative study with the four preprocessing methods: BBHE, CLAHE, BPDFHE, and AIR-AHE using the same database. Furthermore, this section is supported by several visual illustrations showing the obtained results on typical example. All compared methods have been implemented on T1G modality sequence. Moreover, we provide plots that detail the obtained performances. We tested the performance of all the studied approaches:

• •for a set of 20 subject cases of LG glioblastoma brain tumor,
• •for a set of 20 subject cases of HG glioblastomas.

Figure 2 illustrates the comparison results of enhanced MRIs using BBHE, CLAHE, BPDFHE, AIR-AHE, and the proposed approach when applied to an HG case from the BraTS 2015 database.

Fig. 2.

Comparison results of enhanced MRIs tested on an HG case from the BraTS training dataset: (a) original image; (b) BBHE; (c) CLAHE; (d) BPDFHE; (e) AIR-AHE; and (f) proposed approach.

One could notice that the image enhanced with the proposed approach [Fig. 2(f)] has a moderate contrast enhancement on the tumor region, contrary to the enhanced images obtained with compared approaches that show an extreme or over-contrast enhancement on the entire MRI. Figure 3 illustrates the comparison results of enhanced MRIs using BBHE, CLAHE, BPDFHE, AIR-AHE, and the proposed approach when applied to an LG case from the BraTS 2015 database.

Fig. 3.

Comparison results of enhanced MRI tested on an LG case from the BraTS training dataset: (a) original image; (b) BBHE; (c) CLAHE; (d) BPDFHE; (e) AIR-AHE; and (f) proposed approach.

Again, the image enhanced with the proposed approach [Fig. 3(f)] has a moderate contrast enhancement on the tumor region, contrary to the enhanced images obtained with compared approaches that show an extreme or over-contrast enhancement on the entire MRI.

For quantitative evaluation, we used the five above-mentioned image quality metrics, namely MSE, PSNR, EPI, AMBE, and SSIM. Using the studied preprocessing approaches, the average of each metric has been computed for all the tested HG and LG cases.

Figure 4 represents the MSE evaluation. It is obvious that the proposed approach provides the lowest mean value of MSE for both HG and LG cases [MSE (HG) = 0.0004 and MSE (LG) = 0.0005], confirming the good performance of the proposed approach in comparison to other studied approaches.

Fig. 4.

MSE evaluation.

As depicted in Fig. 5, the PSNR evaluation parameter also showed the higher performance of our proposed approach in comparison to other studied approaches. Indeed, the average PSNR value for HG glioblastomas was about 36.39 for the proposed approach, whereas an average PSNR values of 16, 20, 18, and 15, respectively, are obtained for BBHE, CLAHE, BPDFHE, and AIR-AHE approaches. For LG glioblastoma subject cases, the proposed approach provides the highest PSNR values in comparison to other studied approaches as shown in Fig. 5. Hence, the obtained results confirm that the proposed approach could enhance the contrast without degrading the original image quality.

Fig. 5.

PSNR evaluation.

EPI is an important criterion for evaluating image denoising and contrast-enhancement approaches, especially for the case of MRI brain tumors. The value of EPI is in the range of [0,1]; a highest value of EPI means better egdes preservation.

Figure 6 illustrates the EPI quality measurement comparison results. The obtained performance showed that the proposed approach provides the highest average EPI values for both LG and HG glioblastoma tumors. In fact, the average EPI value for HG glioblastomas was about 0.91 for the proposed approach, whereas an average EPI values of 0.37, 0.47, 0.45, and 0.39, respectively, for BBHE, CLAHE, BPDFHE, and AIR-AHE approaches, are obtained. Similar performance has been observed for LG glioblastomas with an average value of 0.91 for the proposed approach and average values of 0.36, 0.51, 0.37, and 0.34 for BBHE, CLAHE, BPDFHE, and AIR-AHE, respectively.

Fig. 6.

EPI evaluation.

Brightness preservation (AMBE) is the major measure indicator of contrast enhancement method’s efficiency. Ideally, the AMBE value should be zero, meaning that the mean input and output brightness are equals. Hence a small AMBE value is usually desired. Figure 7 illustarates AMBE quality measurement comparison results.

Fig. 7.

AMBE evaluation.

The obtained performance showed that the proposed approach provides the lower-average AMBE values for both LG and HG glioblastoma tumors. In fact, the average AMBE value for HG glioblastomas was about 0.01 for the proposed approach, whereas an average AMBE values of 0.05, 0.05, 0.07, and 0.12, respectively, for BBHE, CLAHE, BPDFHE and AIR-AHE approaches, are obtained. Similar performance has been observed for LG glioblastomas with an average value of 0.01 for the proposed approach, and average values of 0.041, 0.039, 0.085, and 0.13, are obtained for BBHE, CLAHE, BPDFHE, and AIR-AHE, respectively. Hence the obtained results confirm that the proposed approach outperforms the usually used state-of-the art brightness pereservation approaches (mainly BBHE approach).

For the SSIM metric, the obtained performance, as shown in Fig. 8, confirms the superiority of the proposed approach in comparison to other tested approaches for both HG and LG cases. This demonstrates that the enhanced image obtained with the proposed approach of denoising and contrast enhancement conserves very well the structure and features of original image.

Fig. 8.

SSIM evaluation.

The above results attest that the proposed approach performed better than stand-alone contrast enhancement techniques. This observation emphasizes that denoising is important, given further improvement on the processed image using only contrast enhancement methods. In addition, one could notice the importance of an ideal choice of contrast enhancement parameter, which is calculated, in the proposed approach, according to the nature of original MRI (original image statistical information, i.e., contrast, variance, standard deviation, etc.). This may explain the superiority of our proposed approach, in addition to the appropriate filters that we used specifically for the MR brain glioblastoma images.

5. Discussion

As illustrated in the previous section, the proposed preprocessing algorithm present a good performance compared to four recent state-of-the art approaches (BBHE, CLAHE, BPDFHE, and AIR-AHE). In fact, it presents the lowest MSE value and the lowest AMBE value, attesting the image detail conservation and the brightness preservation. It also gives the highest PSNR value, the highest SSIM value and the highest EPI value, indicating that the proposed algorithm enhances the contrast without degrading the image quality, conserves very well the structure and even features, and preserves the edge.

To evaluate the performance of the proposed method on the segmentation process, we compute the DM between the obtained result and the ground truth, given by the BraTS 2015 database. A comparative study has been made in order to illustrate the performance of the proposed preprocessing approach

Figure 9 illustrates the segmentation result obtained by applying the expectation maximization (EM) segmentation method³⁷ on a first real case “case 1” from the BraTS training dataset.

Fig. 9.

Segmentation results on a real case 1 from the BraTS training dataset: (a) original image; (b) ground truth; (c) BBHE; (d) CLAHE; (e) BPDFHE; (e) AIR-AHE; and (f) proposed approach.

Figure 10 illustrates the segmentation result obtained by applying the EM segmentation method on a second real case “case 2” from the BraTS training dataset.

Fig. 10.

Segmentation results on a real case 2 from the BraTS training dataset: (a) original image; (b) ground truth; (c) BBHE; (d) CLAHE; (e) BPDFHE; (e) AIR-AHE; and (f) proposed approach.

To evaluate the preprocessing approach performance, we compute the DM between the obtained result and the ground truth. Table 2 exhibits the DM evaluation on two cases from the BraTS 2015 databases. We could notice that the proposed method presents the highest DM value, illustrating thus its outperformance compared to other methods.

Table 2.

DM evaluation over two cases from the BraTS 2015 dataset.

	BBHE	CLAHE	BPDFHE	AIR-AHE	Proposed method
Case 1	0.4443	0.3369	0.5074	0.4035	0.9392
Case 2	0.3627	0.4254	0.3731	0.4672	0.9179

6. Conclusion

Denoising and contrast enhancement are still difficult tasks in brain MRI preprocessing, especially in the case of glioblastoma brain tumors, due the trade-off between the noise reduction and the contrast enhancement.

In this paper, we proposed a new efficient and simple preprocessing approach based on combined denoising and contrast enhancement techniques for MRI glioblastoma brain tumors. The denoising has been performed through nonlinear filtering techniques mainly the BF and contrast enhancement through contrast stretching techniques, where the contrast is enhanced differently for each image based on its statistical information. The performance of the proposed approach was compared to four recent state-of-the art approaches (BBHE, CLAHE, BPDFHE, and AIR-AHE). Different image quality metrics were used in the comparative study. The quantitative results confirmed the higher performance of our proposed method in tumor edge preservation, which is a very important criterion for clinical diagnostics. We have also evaluated the impact of the preprocessing on the segmentation process by evaluating the DM parameter between the obtained results and the ground truth. Furthermore, the obtained results reveal that the proposed approach can successfully enhance the contrast of MRI glioblastoma brain tumors while preserving image quality and image significant features more efficiently than the other compared techniques.

Biographies

Hiba Mzoughi is a PhD student in electrical engineering at the National Engineering School of Gabès, Tunisia. She is attached to a research laboratory specializing in medical imaging and signal processing the Advanced Technologies for Medicine and Signal (ATMS). Recently, she authored a conference paper published at the Fourth International Conference of IEEE, 2018 (ATSIP). Her current research interests include magnetic resonance imaging (MRI), brain tumors, and deep learning based-methods for segmentation and classification tasks.

Ines Njeh is an assistant professor at the Higher Institute of Computer Science and Multimedia, Gabès. She received her MS degree in electronic and new technologies from the National Engineering School of Sfax in 2010 and her PhD in electrical engineering in 2014. She is the author of many journal articles on MRI processing and computer-aided diagnosis for brain glioma exploration. Her current research interests include deep learning-based methods for MRI exploration.

Mohamed Ben Slima received his PhD in telecommunications from the National Institute of Scientific Research, QC, Canada, in 1996. Currently, he is a professor at the National School of Electronics and Telecommunications of Sfax, Sfax University. He was the vice president R&D of Measurement Microsystems, Canada. His expertise has been also enhanced by his career (started in 1995) as an assistant professor at the University of Quebec at Trois-Rivieres, Canada.

Ahmed Ben Hamida graduated in electrical engineering in 1988. He obtained his PhD in 1998. Currently, he is a professor at ENIS School of Engineering of Sfax, Tunisia, and he is the head of the ATMS Research Unit, a novel creation of the year 2006, gathering university researchers in technological sciences and from the Faculty of Medicine.

Chokri Mhiri is professor of neurology in the Faculty of Medicine, University of Sfax, Tunisia. He received his medical doctor degree from the University of Sfax in 1987. He obtained a research master’s degree in neurosciences from the Pierre Marie Curie University in Paris, France, in 1990. He graduated as neurologist in 1991. In 1996, he was nominated as head of the Department of Neurology in the Sfax Faculty of Medicine. He assumed several responsibilities in his hospital (president of the medical board of Habib Bourguiba University Hospital (2008–2013); director of Clinical Investigation Center since 2015). He is member of the scientific board of Sfax Faculty of Medicine since 2001. Currently, he is the president of the Tunisian Society of Neurology. He has participated in the publication of more than one hundred peer-reviewed scientific papers.

Disclosures

Dr. Mzoughi has nothing to disclose.