Secure Watermarking Technique for Medical Images with Visual Evaluation

This paper presents a hybrid watermarking technique for medical images. The method uses a combination of three transforms: Discrete Wavelet Transform (DWT), Discrete Cosine Transform (DCT), and singular value decomposition (SVD). Then, the paper discusses the results of applying the combined method on different medical images from eight patients. The images were watermarked with a small watermark image representing the patients' medical data. The visual quality of the watermarked images (before and after attacks) was analyzed using five quality metrics: PSNR, WSNR, PSNR-HVS-M, PSNR-HVS, and MSSIM. The first four metrics' average values of the watermarked medical images before attacks were approximately 32 db, 35 db, 42 db, and 40 db respectively; while the MSSM index indicated a similarity between the original and watermarked images of more than 97%. However, the metric values decreased significantly after attacking the images with various operations even though the watermark image could be<br>retrieved after almost all attacks. In brief, the initial results indicate that watermarking medical images with patients' data does not significantly affect their visual quality and they can still be used by medical staff.


INTRODUCTION
Data hiding has increasingly become an important tool in authentication of images and protection of rightful owners copyright. Also, there is an increasing need to store and transfer patients' medical images over the Internet and other computer networks for sharing among medical staff in medical institutions all over the world. Image watermarking techniques that hides important details inside cover images can be divided into two broad domains: spatial domain and frequency domain [1,2]. Various medical images based watermarking schemes have been proposed in literature [3,4,5]. Three of the most important frequency watermarking methods are the discrete cosine transform (DCT), discrete wavelet transform (DWT) and Singular Value Decomposition (SVD). Many researchers have used a hybrid of two or more transforms in order to compensate for the shortcomings of various transforms.
There are many examples of spatial domain techniques such as LSB substitution, spread spectrum, and patchwork. Lin et al. [6] proposed a spatial watermarking methods where the watermark logo is fused with noise bits first, and then XORed with the feature value of the image by 1/T rate forward error correction (FEC), where T is the times of data redundancy. The watermark bits are extracted by majority voting.
robust adaptive watermarking method based on DCT, SVD and Genetic Algorithm (GA). The host image luminance masking is used and the mask of each sub-band area is transformed into frequency domain. Subsequently, the watermark image is embedded by modifying the singular values of DCT-transformed host image with singular values of mask coefficients of host image and the control parameter of DCT-transformed watermark image using GA. Singh et al. [9] proposed a robust hybrid watermarking technique using DWT, DCT, and SVD. First, the host image into first decomposed by DWT and the Low frequency band (LL) and watermark image are transformed using DCT and SVD. Then the S vector of watermark image is embedded in the S component of the host image and the watermarked image is generated by inverse SVD on modified S vector and original U, V vectors followed by inverse DCT and inverse DWT.

METHODOLOGY
The following sections will give details of the used watermarking algorithm and evaluation metrics.

Watermarking algorithms
The designed and implemented algorithm is a combination of three frequency domain techniques: discrete wavelet transform (DWT), discrete cosine transform (DCT), and singular value decomposition (SVD). DWT decomposes an image into frequency channels of constant bandwidth on a logarithmic scale by separating an image into a set of four non-overlapping multiresolution sub bands denoted as lower resolution approximation image (LL), horizontal (HL), vertical (LH) and diagonal (HH) with the availability of multiple scale wavelet decomposition. The watermark is usually embedded into the high frequency detail sub-bands (HL, LH and HH sub-band) because the human visual system (HVS) is sensitive to the low-frequency LL part of the image. We can usually embed sensitive data such as medical information in higher level subbands since the detail levels carry most of the energy of the image [10]. DWT achieves higher robustness since it has the characteristics of space frequency localization, multi-resolution representation, multi-scale analysis, adaptability and linear complexity [11].
Also, DCT has a very good energy compaction property. It separates the image into different low, high, and middle frequency coefficients [12]. The watermark is embedded in the middle frequency band that gives additional resistance to the lossy compression techniques with less modification of the cover image. The DCT coefficients D(i, j) matrix of an image (N x M) with pixel intensity I(x, y) are obtained as follows: perception of the cover image, which achieves better quality of the watermarked image and better robustness against attacks. Also, singular values represent the intrinsic algebraic image properties [12]. Figure 1 shows the approach taken in embedding the patients' data into a cover medical image; First, DCT is applied on the LL component of the DWT transformed cover image; SVD is applied to the DCT coefficients. Then, the watermark is DCT transformed and the singular values of the SVD transformed coefficients are embedded in the singular values of the DWT transformed coefficients of the cover image. Figure 2 shows the extraction approach of the patient's image data from the watermarked image. The watermarked images is DWT and DCT transformed then SVD is applied to the DCT coefficients; the watermark is extracted from the LL sub band of DWT. For an added security, the watermark image can be encrypted before embedding it in the cover image.

Evaluation metrics
Four visual metrics (WSNR, MSSIM, PSNR-HVS-M, and PSNR-HVS) described by Ponomarenko et. al. [13] are used for comparing the watermarked images with their originals. Traditionally, the efficiency of an image processing operation ; i.e. lossy compression is usually analyzed in terms of rate-distortion curves. These curves represent dependencies of PSNR (or MSE) on bits per pixel (bpp) or compression ratio (CR) where PSNR and MSE are calculated for some original image and the corresponding processed image.
where denote the values of the original and processed pixels and N, M denote an image size [14]. In order to obtain a high imperceptibility of the watermarked image, it is desirable to have a high value of PSNR; meaning a lesser value of MSE.
Also, usually the similarity and differences between an original image and a processed image is measured by the Normalized Correlation (NC). Its value is generally 0 to 1. Ideally it should be 1 but a value 0.7 or higher is usually acceptable [15].
where denote the values of the original and processed pixels and X, Y denote an image size.
Two different distorted images with the same PSNR value with respect to the same original image may give significantly different visual impact. It is well known that conventional quality metrics, such as MSE, SNR and PSNR do not always correlate with image visual quality [17,18]. Therefore, the choice of a proper visual quality metric for analysis and comparisons is always problematic since the human visual system (HVS) is nonlinear and it is very sensitive to contrast changes and to noise [19]. Many studies have confirmed that the HVS is more sensitive to low frequency distortions rather than high frequency components. The best performance was achieved by the metrics PSNR-HVS-M, PSNR-HVS, and WSNR [14] especially if there is noise or the images are to be compressed. HVS-based models are the result of trade-off between computational feasibility and accuracy of the model. HVS-based models can be classified into two categories: neurobiological models and models based on the psychophysical properties of human vision. Psychophysical HVS-based models are implemented in a sequential process that includes luminance masking, colour perception analysis, frequency selection, and contrast sensitivity [19].
Recently, processing of images is done using perceptual image quality assessment methods, which attempt to simulate the functionality of the relevant early human visual system (HVS) components. These methods usually involve a pre-processing process that may include image alignment, point-wise nonlinear transform, low-pass filtering that simulates eye optics, and color space transformation, a channel decomposition process that transforms the image signals into different spatial frequency as well as orientation selective subbands, an error normalization process that weights the error signal in each subband by incorporating the variation of visual sensitivity in different subbands, and the variation of visual error sensitivity caused by intra-or inter-channel neighbouring transform coefficients, and an error pooling process that combines the error signals in different subbands into a single quality/distortion value [20].
PSNR-HVS takes into account the HVS properties such as sensitivity to contrast change and sensitivity to low frequency distortions; while the PSNR-HVSM takes into account the contrast sensitivity function (CSF). Similar to PSNR and MSE, the visual quality metrics PSNR-HVS and PSNR-HVSM can be determined: where I,J denote image size, K=1 [(I-7)(J-7)64] , are DCT coefficients of 8x8 image block for which the coordinates of its left upper corner are equal to i and j, Xij e are the DCT coefficients of the corresponding block in the original image, and is the matrix of correcting factors [21].

The Weighted Signal to Noise Ratio (WSNR) is a noise metric where the difference (residual) between the original and the processed images must be noise. (WSNR) uses a Contrast Sensitivity Function (CSF) given by the following:
where is a radial angular frequency The WSNR between an original image (x) and a processed image (y) is: The structural similarity index (SSIM) measures the similarity between two images [19]. SSIM compares two images using information about luminous, contrast and structure. SSIM metric is calculated on various windows of an image. The measure between two windows x and y of common size N×N is given as follows: MSSIM (Multi-Scale Structural Similarity) is a multi-scale extension of a SSIM metric. MSSIM [22] is introduced to incorporate the variations of viewing conditions to the previous single-scale SSIM measure. MSSIM is known as mean structural similarity index metric [22] and it is given by: where M is the correlation between two images x, y Correlation is a similarity measure between two functions. The correlation measure between two functions x(x,y) and s(x,y) in discrete form is defined as: Where is the complex conjugate, x=0, 1,…….., M-1 and y=0, 1,……, N-1 The algorithm was evaluated using five quality metrics. Table 1 shows the PSNR, P-HVS, P-HVS-M, WSNR, and MSSIM metrics among all the watermarked images before any attacks. It can be observed that the PSNR average value is about 32 db, P-HVS average value is around 35 db, P-HVS-M average value is about 42 db, and the WSNR average value varies from 35 db to 47 db. The MSSIM metric shows that the watermarked images are highly visually similar to the original images with similarity index values between the original and the watermarked images of more than 0.97%. Also, it can be observed that there is no significant difference between the average metric values among the various images; only the WSNR value of the of the Head image varies from one image to another with approximately 15 db difference between the Fingerprints image and the Head image; that is mainly due to the characteristics of the two images.  The limitation of this research is that the algorithms cannot determine how much of medical information is lost after watermarking medical images or even after attacking the images. Only medical doctors can decide the important segments of a medical image that are affected by watermarking or by attacking. Also, the effects can vary from one image to another. Finally, recovering the watermark after some attacks does not necessarily indicate that all medical information is preserved.

CONCLUSIONS
The results of this limited research show that watermarking medical images with a watermark of patients' personal details does not significantly affect the visual quality of the original medical images; and they can be utilized for their medical purpose. It was experimentally quantitatively demonstrated using Human Visual System (HVS) metrics that the watermarked medical images were similar to their originals. Also, choosing the appropriate watermarking algorithm is essential to obtain the robustness, imperceptivity and security needed to protect the patients' personal data inside a medical image and there are many transform domain algorithms that are available and can be utilized to preserve the characteristics of the original images. Artificial intelligence methods will be used in the future to classify the effectiveness of new algorithms.