Hybrid multiple watermarking technique for securing medical images of modalities MRI, CT scan, and X-ray

Received April 17, 2019 Revised Nov 20, 2019 Accepted Dec 6, 2019 In order to contribute to the security of sharing and transferring medical images, we had presented a multiple watermarking technique for multiple protections; it was based on the combination of three transformations: the discrete wavelet transform (DWT), the fast Walsh-Hadamard transform (FWHT) and, the singular value decomposition (SVD). In this paper, three watermark images of sizes 512x 512 were inserted into a single medical image of various modalities such as magnetic resonance imaging (MRI), computed tomography (CT), and X-Radiation (X-ray). After applying DWT up to the third level on the original image, the high-resolution sub-bands were being selected subsequently to apply FWHT and then SVD. The singular values of the three watermark images were inserted into the singular values of the cover medical image. The experimental results showed the effectiveness of the proposed method in terms of quality and robustness compared to other reported techniques cited in the literature.


INTRODUCTION
The evolution of information and communication technologies (ICTs) and the development of telecommunication applications in healthcare networks play an important role in the patient diagnosis. This evolution may raise concerns about the preservation and confidentiality of data transferred between hospitals [1]. In this context, the digital watermarking has emerged as a solution to fight against all types of fraud. The digital watermarking [2,3] is a recent discipline which consists of inserting an invisible watermark into a cover medical image. Currently, the insertion in the frequency domain can ensure a high degree of security of medical data by preserving their integrity and confidentiality.
This contribution is a follow up of our work done in [9], it"s able to insert three watermarks of sizes 512x512 without altering the quality of the watermarked image and its robustness against noise, filtering, and compression attacks. This paper resumes the following sections: section 2 presents the techniques of watermarking used in our method. In section 3, DWT-FWHT-SVD based watermarking algorithm has been 2. The step1 is applied on the first watermark image and then the high frequency sub-band HH 3w1 is selected. 3. Apply the second and the first level of DWT to the second and the third watermark images respectively and then select HH 2w2 and HH 1w3 . 4. Apply FWHT on the selected sub-bands HH 1 , HH 2 , HH 3 and HH 3w1 of the original medical image and the first watermark image respectively and then apply SVD on FWHT coefficients of the selected sub-bands (HH 1 , HH 2 , HH 3 and HH 3w1 ) and on HH 2w2 and HH 1w3 of the second and the third watermark images to get: , i=HH 1 , HH 2 and HH 3 , j=HH 3w1, HH 2w2 and HH 1w3 which α: scaling factor. 6. Apply inverse FWHT and inverse DWT to get the watermarked medical image.

Extracting process
1. Decomposing the cover medical image into the third level DWT transform using Daubechies wavelet and then select HH 1 , HH 2 , and HH 3 . 2. Apply step1 to the first watermark image and select HH 3w1 . 3. Apply the second and the first level of DWT to the second and the third watermark images respectively and then select HH 2w2 and HH 1w3 . 4. Apply FWHT on the selected sub bands HH 1 , HH 2 , HH 3 and HH 3w1 of the original medical image and watermark image1 respectively and then apply SVD on FWHT coefficients of (HH 1 , HH 2 , HH 3 and HH 3w1 ) and on HH 2w2 and HH 1w3 of the second and the third watermark images to get (3) and (4). 5. Apply step 1 and 4 to watermarked medical image to obtain: , i=HH 1 , HH 2 and HH 3 (6) 6. The singular values of the high frequency sub-bands of the three watermarks are obtained by the singular values of HH 3 HH 2 , and HH 1 of the watermarked medical image and the cover medical image respectively by (7): 7. We obtain the extracted watermark image 1 by applying ISVD using (7) and then we apply inverse FWHT and inverse DWT. 8. And finally, we obtain the extracted watermark image 2 and the extracted watermark image 3 by applying ISVD using (7) and then apply inverse DWT.

EXPERIMENTAL RESULTS
We have implemented our proposed multiple watermarking technique DWT-FWHT-SVD in MATLAB R2013b. Several experimental tests were performed on the original medical images of various modalities of size 512x512 shown in Figure 1. 'Elaine', 'Hill', and 'Fingerprint' are considered as watermarks of the same size of the cover image shown in Figure 2. To evaluate the performance of our proposed algorithm in terms of invisibility and robustness, various evaluation tools are taken into consideration such as: the Mean Square Error (MSE), Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) and Normalized Correlation Coefficient (NC). The PSNR is described as follows [15]:  [18] represents the mean square error which measures the perceptual distance between original and watermarked image: A and A" represent the original and the watermarked medical image of size MxN respectively. The second metric used for evaluating the quality of the image is Structural Similar Index Measure (SSIM) [24,25] measures the similarity between two images: the original image and the watermarked image. It varies between two values 1 and -1. The 1 value indicates that the two images are identical. (10) I and Iw are the original image and watermarked image respectively.
is luminance comparison function, is contrast comparison and is structural comparison. The robustness is measured by normalized correlation coefficient (NC) [13]: W (i, j) and W" (i, j) are the original and extracted watermark respectively. The optimal NC value must be superior to 0.97. The invisibility of inserting three watermarks into a single cover medical image is evaluated by calculating PSNR and SSIM. Figure 3 shows the cover medical images 'MRI', 'CT scan', and 'X_ray', the watermarks 'Elaine', 'Hill' and 'Fingerprint' and, the watermarked images. This figure shows that our proposed algorithm gives a perfect invisibility.
In Tables 1 and 2 the performance of our approach has been evaluated without any noise attacks. In Table 1, we have calculated the PSNR and the SSIM values of the cover medical images 'MRI image', 'CT image', 'Thorax X_Ray' and 'Shoulder X_Ray' at different gain factors. The maximum PSNR and SSIM values obtained against the three inserted watermark images 'Elaine', 'Hill' and 'Fingerprint' are respectively INF and 0.9995 at gain factor 0.01. And the minimum PSNR and SSIM values are respectively 58.2494 dB and 0.8956 at gain factor 0.08. To say that a PSNR is good, it must be greater than 30 dB and also for SSIM it must tend to 1. So from Table 1, we can say that the results are satisfying. In order to highlight our approach in term of invisibility, a comparison has been made with other reported techniques [4,6,7,8] shown in Table 1. Also at gain factor 0.9 and 0.003 the PSNR values obtained respectively are 45.1898 dB and INF against 22.6116 dB of [5] and 41.3448 dB of [22]. From the results obtained, we can conclude that our algorithm gives a good invisibility. Moreover, the NC values of the cover medical images without applying any noise attack are 1 in all gain factors as shown in Table 2.
To check and improve the robustness of our watermarking system, we have applied a various attacks such as JPEG compression, Gaussian noise, Salt & pepper, median filtering, and histogram attacks at scaling factor 0.07. The NC performance of our approach against various attacks compared to existing methods [4,5,6,8,22] is shown in Table 3. The maximum NC value obtained at scaling factor 0.07 is 1 against JPEG compression (QF-100) compared to 0.9939 of [6] and 0.9886 of [8], 1 against Salt & pepper (density=0.001), compared to 0.8227 of [4], 0.9908 of [6], 0.9658 of [8], and 0.8227 of [22], and 1 against median filtering compared to 0.9993 of [4], 0.9939 of [5], 0.9861 of [6], and 0.0123 of [8]. Moreover, the NC values obtained against all attacks are superior to other reported techniques [4,5,6,8,22]. The results obtained in terms of imperceptibility, robustness, and capacity of insertion approve that our contribution is effective.

CONCLUSION
In this article, we have presented a robust multiple watermarking methods based on the combination of three techniques DWT, FWT, and SVD. Three watermark images 'Elaine', 'Hill' and 'Fingerprint' are embedded into a single cover medical image of modalities: magnetic resonance imaging (MRI), computed tomography (CT), and X_Ray. The insertion of these three watermarks was made in the high-frequency sub bands HH1, HH2 and HH3 of the cover medical image. However, the insertion into theses sub bands can augment the robustness of watermarking scheme and can give a good quality of resulting image. The obtained results approve that the quality of our proposed scheme is perfect compared to other methods in terms of invisibility, robustness, and capacity. In the future works, we will optimize the performance of our contribution against geometric attacks.