You are on page 1of 6

A High Capacity Reversible Multiple Watermarking Scheme for Medical Images

Behrang Mehrbany Irany, Xin Cindy Guo and Dimitrios Hatzinakos

Abstract—Reversible data hiding schemes can be used to embed sensitive personal information in a generic signal without any loss of either the embedded or the host information. Multiple watermarking allows embedding different marks at different stages into the host media. This paper proposes a high capacity reversible multiple watermarking scheme for medical images based on integer-to-integer wavelet transform and histogram shifting. The novelty of the proposed scheme is that it uses a scalable location map and incorporates efficient stopping conditions on both wavelet levels and different frequency subbands of each level to achieve high capacity payload embedding, high perceptual quality, and multiple watermarking capability. Results show that the proposed method attains high perceptual quality in high capacity rates for the medical images. The improvement in the visual quality of the watermarked images along with the possibility of multiple watermark embedding indicate that this algorithm is capable of rendering a wide range of applications that aim at security and privacy protection in the medical field. Index Terms—Reversible watermark, reversible data hiding, multiple watermarks, histogram shifting, integer-to-integer wavelet transform, privacy protection, medical images.

I. I NTRODUCTION Recently various governments have been promoting the Electronic Health Record (EHR) which is collecting the health information of individual patients in the electronic form and sharing it across different health care settings. This approach allows computerizing and simplifying workflow in health care system, increases safety through evidence based decision support, eases quality management, and outcomes reporting; which in turn facilitates diagnosis of patients and education of medical care personnel. All the aforementioned records contain sensitive personal information which should be protected properly. In majority of medical records, personal information is stored in separate files which may be accessed when needed. However, rapid development of tools and software to manipulate and transfer digital multimedia over the internet has made it easier for adversaries and hackers to broaden copy violation, counterfeiting, forgery and fraud into this field. In order to prevent such acts there has been an increase in urgent demands for protecting and authorizing multimedia content against illegal use and distribution. The sensitive nature of individual’s personal data implies additional security to be employed in order to obtain the sufficient privacy protection level. Various techniques have emerged to verify integrity and prevent forgery. Most methods rely on embedding a perceptually
Behrang Mehrbany Irany and Dimitrios Hatzinakos are with the Edward S. Rogers Department of Electrical and Computer Engineering, University of Toronto, 10 King’s College Road, Toronto, ON, Canada, M5S 3G4 e-mail: {behrang,cguo,dimitris}

invisible mark called a digital watermark or signature into the multimedia file [1]. The users are able to verify the content by extracting and comparing the watermark or signature. Coatrieux et al. [2] examined the relevance of watermarking for medical images. They concluded that digital watermarks could be used in addition to the current security tools, such as firewalls and encryption, in order to better protect medical records. Giakoumaki et al. [3, 4] proposed a wavelet transform based watermarking technique for medical data. They were able to address medical data authentication, archiving, and retrieval beside source and data authentication. They could embed multiple watermarks containing patient’s personal and examination data, keywords for information retrieval, physician’s digital signature for authentication, and a reference message for data integrity control in ultrasound images. The main drawback of the method is that the original medical records or images are overwritten and one cannot retrieve the original images from the watermarked ones which may be unacceptable in diagnosis. Reversible watermarking techniques were originally born to be engaged mainly in situations where the authenticity of a digital image has to be granted and the original content is definitively needed at the decoding side. In reversible watermarking schemes, the original data can be reconstructed from the distorted watermarked image by extracting the watermark. In the recent years several reversible watermarking schemes have been proposed [5–12]. In these approaches watermark is embedded into the original file deliberately so that both the watermark and the file can be recovered completely without losses. Since the pixels of the cover image usually exhibit strong spatial redundancy, it is possible to transform the images so that ”free space” can be created to embed extra information. Like a regular watermark, a reversible watermark should have invisibility, bandwidth efficiency, and security properties. In this paper, a reversible multiple watermarking scheme with high embedding capacity for digital medical images is proposed. The proposed scheme employs histogram shifting and integer-to-integer wavelet transform. The contribution of the proposed scheme is that it uses a scalable location map which size is significantly small comparing to the payload and it incorporates different stopping conditions on both the level of the wavelet transform and the number of coefficients used in each subband to achieve high perceptual quality and multiple watermarking property. The paper is organized as follows. section II formally discusses different components of the proposed algorithm. Section III presents the embedding and the decoding parts

978-1-4577-0274-7/11/$26.00 ©2011 IEEE


Lossless compression does not offer great reduction in message size. respectively. the forward transform of the one level decomposition is defined by the low and high frequency coefficients 𝑠1. The (2.2𝑛 𝑠0. [13] is used in this study. If the to-be-embedded watermark bit is 0. Invertible integer-to-integer Wavelet Transform Conventional wavelet transform cannot be used in the reversible watermarking methods as it doesn’t support full reversibility. 1. such as an image. 10].2𝑛 + 𝑠0. 14]. This map is then compressed with no loss and embedded into the original signal. (b) shifted histogram. the detection procedure follows the reverse pattern exactly.𝑛−1 + 𝑑1.2 .2) integer wavelet transform has one of the best coding performances [13]. The main concept is introduced in this section.𝑖 ’s are integers. in this example 0 is changed to 1. METHODOLOGY A.2𝑛+1 = = 𝑠1. one is added to all coefficients with values greater than 𝑛 as shown in Figure 1(b). receiving a coefficient with value 𝑛 means that bit 0 has been embedded and getting 𝑛 + 1 translates as bit 1. To extract the embedded bits. The original data can be completely recovered by subtracting one from all coefficients greater than the used coefficient value 𝑛. the embedding capacity is limited in a lot of the early algorithms [5. Illustration of histogram shifting (a) original histogram.𝑛 𝑠1. all the coefficients smaller than -1 can be shifted towards negative infinity to create a gap to embed bits at -1. the explained procedure is equally valid.. 𝑑1. it means that the watermarked image has values smaller than 0 or greater than 255 after histogram shifting and inverse wavelet transform.𝑛 ) + 1/2⌋ . (2) Integer-to-integer wavelet transforms are not showing normalized behavior throughout different frequency subbands of the decomposed signal. As a result of such characteristic pattern through experiments and analysis. Histogram shifting method is based on creating a gap.𝑛 ) + 1/2⌋ .2𝑛 + 𝑠0.2) interpolation transform defined by Calerbank et al. Finally. Then the subband is scanned. . Results and discussions are presented in Section IV. where 𝑠0. can be further decomposed into more levels.2𝑛 + ⌊1/4(𝑑1. The histogram after embedding is shown in Figure 1(c). most methods create an location map of all coefficients to show which ones cannot be modified. 𝑠0. Similar to floating point wavelet and wavelet packet transforms. therefore.𝑛 = = 𝑠0. 0 and 1 in the above example. One advantage of the wavelet transform is that it offers good resolution in both spatial and frequency domains.𝑛 and 𝑑𝑗.2𝑛+1 − ⌊1/2(𝑠0. (1) 9000 8000 9000 8000 7000 6000 5000 4000 3000 2000 1000 0 −20 −10 0 10 20 Coefficient Value Coefficient Value (a) 9000 8000 (b) Number of Occurence 7000 6000 5000 4000 3000 2000 1000 0 −20 −10 0 10 20 Coefficient Value (c) Fig.𝑛 + ⌊1/2(𝑠0. At first. used as the free space during data embedding process in the histogram of either the original or transformed data.Number of Occurence 7000 6000 5000 4000 3000 2000 1000 0 −20 −10 0 10 20 Number of Occurence of the proposed reversible multiple watermarking scheme. 12. Compare to other transforms defined in Calderbank’s paper. the low and high frequency coefficients. C. 𝑠0. it is inevitable that the reconstructed signal has values outside of the acceptable range. 2) interpolating transform used in this scheme follows the same pattern and modifications in different frequency subbands lead to diverse effects with different intensities in the spatial domain. Histogram Shifting and Data Embedding Histogram shifting has been applied for reversible watermark embedding in a number of algorithms [8. it is shown that modifications in low frequency coefficients have greater impact in the reconstructed signal[13]. are selected. the (2. The function ⌊∗⌋ rounds ∗ to the nearest integer towards minus infinity. invertible integer-to-integer wavelet transforms based on lifting as proposed in [13] are utilized.𝑛 − ⌊1/4(𝑑1. the aforementioned Equations (1) are applied to both the horizontal and vertical directions. An example of the procedure is explained here to fully demonstrate the technique. The reverse transform of the decomposition is easily calculated as 𝑠0..1 . II. This is defined as underflow or overflow in previous works [10]. B. Since . Overflow and Underflow Problem In most cases. the 𝑛𝑡ℎ low frequency and high frequency wavelet coefficients at the 𝑗 𝑡ℎ wavelet transform level [13].𝑛−1 + 𝑑1. the value of that coefficient is increased by one. all coefficients which values are used in the embedding process (the ones the shifting technique is performed on). Figure 1(a) shows the histogram of the HH subband coefficients of the Lena’s image. To perform the one level decomposition on a two-dimensional (2−𝐷) signal. 𝑠0. To shift the coefficients with negative values.2𝑛+2 ) + 1/2⌋ . For an 8-bit bitmap image. the coefficient is not modified and left intact. Concluding remarks and future directions are drawn in Section V.. starting from -1.2𝑛+2 ) + 1/2⌋ . (c) histogram after watermark embedding where 𝑠𝑗. Hence. generated in the first decomposition level. In order to create a gap to embed bits at a coefficient with value equals 𝑛 (𝑛 > 0). if the to-beembedded bit is 1. To avoid such conditions.𝑛 and 𝑑1. On the other hand. In a one-dimensional (1D) signal 𝑠 = [𝑠0.𝑁 ]. The (2. For instance.𝑛 as 𝑑1.𝑛 are.

to minimize the resultant distortion while maximizing the capacity. If the required space to embed all watermark bits is not acquired when 𝛾1 is reached. As mentioned in II-D. respectively. histogram shifts are started from zero and follow the technique described in II-B. it shows a considerable amount of redundancy as in most images the tone varies smoothly except on the edges and abrupt changes are rare. Therefore. and so on are shifted respectively through the same method. the embedding process starts in the HH1 subband of the first level integer wavelet transform to embed the data. These parameters guide the encoder and decoder through embedding and data extraction processes. In the proposed scheme. the coefficient map employed has the size of the HH3 which is 1/64𝑡ℎ of the original image. it is possible to generate the map for the HH1 from the one for the HH2 subband by upsampling. the number of coefficients to be modified has to be at least the size of the embedded watermark message. . By continuing this pattern. The process starts by shifting all the coefficients in the 𝐻𝐻1 subband according to the value of 𝛾1 in each stage. Hence. To accomplish this goal. 𝑚}. (b) corresponding coefficient maps underflow and overflow occurs to pixels that have values close to 0 or 255 in an 8-bit bitmap image. In general. This section introduces the technique to find the most optimum and efficient stopping parameters 𝛾1 and 𝛾2 . These coefficients are ignored in histogram shifting procedure and hence there will not be any changes to the pixels in the spatial domain. one can easily locate the corresponding coefficients in the frequency domain. The number of shifts directly determines the changes in coefficients with high magnitude. The following is the procedure through which the proposed scheme defines the stopping parameters 𝛾1 and 𝛾2 : 1) The 1𝑠𝑡 level integer-to-integer wavelet transform is performed on the host image. as illustrated in Figure 1(a). therefore the start or end shifting coefficients cannot be easily defined. 4. the process moves to either HL1 or LH1 subbands considering which subband results in a better performance and less distortion. -5. 𝛾1 is the stopping parameter used in 𝐻𝐻 subbands and 𝛾2 determines the stopping coefficients of both HL and LH subbands. The advantage of this shifting technique is that fewer shifts are needed if the starting coefficient is zero. . Derivation of Stopping Parameters Stopping parameter determines the last shifted coefficient in a subband by the embedding process. the wavelet coefficients in different subbands have different energies and so are of different importance concerning the resultant marked image and the level of distortion imposed. This means that the size of the general coefficient map can be confined to the map of one subband. Determination of Shifting Coefficients In the aforementioned example in II-B. The stopping parameters are the order of the chosen coefficients in the set {0. 2. It means if there are bits to be embedded left when the coefficient corresponding the stopping parameter is reached in a particular subband. Furthermore. It is shown that coefficients in detailed subbands follow an approximately Laplacian like distribution centered at zero [15]. where 𝑚 is the last available coefficient in the considered subband.. −3.(a) (b) Fig. An example of the coefficient map is shown in Figure 2. then. the most efficient stopping parameter 𝛾2 for these subbands to achieve the least possible distortion in the marked image is found. It is essential to note that the LL subband is never used for data embedding hence never modified in any of the levels for two main reasons. the coefficient distribution does not have a Laplacian shape as the high frequency subbands. 2. -3. In general to reduce the size of the used coefficient map. Examples of coefficient map (a) original images. 2) An iteration process is engaged to evaluate the performance of the marked image at different stages. The original coefficient map is the same size as the original image. First. a stopping parameter 𝛾1 is defined through a well-defined systematic procedure.. D. It should be noted that 𝛾1 is set to be engaged in all HH subbands involved in the embedding process and 𝛾2 is the chosen parameter for all 𝐻𝐿 and 𝐿𝐻 subbands of different wavelet levels. coefficients with values equal to 2 (which are originally coefficients with values equal to 1 but due to the first shift they are relocated to 2). As elucidated and discussed in section II-A. The coefficients corresponding to pixel values between 10 and 245 are used in this case. fewer shifts leads to less distortion once the marked signal is reconstructed. the embedding process switches to the next subband to embed the rest of the bits. considering the performance and efficiency of the process. more than one histogram shifts are usually needed to embed a large amount of data information. Then. However. which play a crucial role in gaining the best visual quality and the least distortion in the marked images in this scheme.. the original map can easily be downsampled. −1. the maximum number of embedded bits is the number of coefficients with values equal 0. Due to this fact. E. changes in the LL subband have the most impact in the reconstructed image quality and if it is tampered with it causes the highest degradation in the visual quality of the watermarked image. The proposed algorithm exploits this property to achieve higher visual quality and less distortion in terms of PSNR at different bit rates in the watermarked image. As an experimental fact. and in order to make the approach more efficient the size of the map can be reduced further. explained in section II-E.

the resultant PSNR of the reconstructed image from the shifted HH1 subband for an increased value of 𝛾1 by 1 is computed and compared to the resultant PSNR of the reconstructed image from each one of the HL1 and LH1 subbands as the first shift at 0 is applied to them. If the PSNR obtained from 𝛾1 is still greater than PSNRs in both HL1 and LH1 subbands this step is repeated and process proceeds with the iteration. PSNR of the reconstructed image resulted from shifts in HL1 and LH1 subbands is computed and compared against the resultant PSNR of the reconstructed image from the first shift in the HH2 subband. these values are computed and set prior to the actual embedding process. 4. following HH1 . multi-pass embedding scheme can be adopted. 4) If there is no other watermark present. denoted by 𝑠𝑢𝑏𝑖𝑑𝑥 . If PSNR from HL1 and LH1 is found to be greater then step is repeated with an increase in 𝛾2 value. Data extracting algorithm 1) The acceptable distortion level. III. Then the process checks if the created free space. as the side information. 4) Again. 3. Figures 3 and 4 illustrate the flowcharts of the embedding and decoding processes respectively for each watermark. which means to perform multiple iterations of the watermarking process using the previously marked image as the input. the one imposing less distortion is taken as the succeeding subband used for defining 𝛾2 and data insertion. is enough to embed the total payload. 2) The watermark message is converted into bits. considering the location map. Data embedding algorithm Marked Image Integer-toInteger Wavelet Transform Overhead Information Key Watermark Payload Extraction Total Watermark Payload Decryption Watermark Message No Inverting Histogram Shifts Inverse Integer Wavelet Transform Watermark Payload Is Complete? Original Host Image Yes Fig. The procedure mentioned above indicates that the stopping parameters 𝛾1 and 𝛾2 are pre-computed in this scheme. and then encrypted. In this case. measured by PSNR. considering the coefficient map. Then it checks. In this case the next two steps are passed and process proceeds from step 6. In cases where the size of payload is greater than the total number of available coefficients in the original image. the iteration process starts by creating the first gap in the histogram of 𝑠𝑢𝑏𝑖𝑑𝑥 subband. 3) PSNR of the reconstructed image is calculated in both HL1 and LH1 subbands. so the first gap is created by shifting all coefficients with values greater or equal to 1.. or in other words. If any overhead is available. else the procedure sets 𝛾1 . is set and given as an input to the system in case of multiple watermarking. the system fails to obtain any stopping parameters for a singlepass embedding procedure. In case the payload is not enough. 2) interpolating wavelet transform. if not the first gap in the 𝑠𝑢𝑏′ (e. the coefficient map is formed to secure the pixel values close to the boundaries of the acceptable range from falling off the . PROPOSED ENCODING AND DECODING ALGORITHMS The proposed scheme can be adopted to guarantee multiple watermark embedding. if the total created hiding space is enough for all the bits to be embedded up to this step.The value of the stopping parameter 𝛾1 equals 1 in the first iteration. If there is more space required. The chosen subband’s name is recorded. which histograms are modified with the first shift at 0.g. Here we propose to use an n-level (2. 3) The process starts by performing integer wavelet transform on the original image. where 𝑛 is defined by the size of the to-be-embedded payload and the values of stopping parameters. Watermark embedding process can be summarized in the following steps: Host Image Coefficient Map Formation Watermark Integer-toInteger Wavelet Transform Find 1 & 2 The Current Pass Watermark Payload Histogram Shift Overhead Information Data Embedding Total Payload Encryption Yes Secret Key Marked Image Another Pass Needed? Final Marked Image Fig. then scans the image to locate any watermarking overhead embedded in the ”host image” from a prior embedding. else the system sets 𝛾2 . then it is extracted and the existing information is used for data embedding of the new watermark. if 𝑠𝑢𝑏𝑖𝑑𝑥 = 𝑖𝑑𝑥 HL1 then 𝑠𝑢𝑏′ =LH1 ) is then created and once more the 𝑖𝑑𝑥 availability of enough space to embed the total payload is checked.

n n Fig. indices of multiple watermarks which indicate the size of each embedded watermark and the coefficient number its embedding begins at. Figures are zoomed in so the difference can be magnified. 2) Using the provided watermark identification as the input (in case of multiple watermarks). In this algorithm the stopping parameters are chosen to satisfy the PSNR set as the input in step 1. this means the process derives the parameters which result the given PSNR as the condition. and 𝑛 is the total number of pixels in the image. 7) Data including payload and overhead information is embedded in the created gaps. 4) The extracted watermark is decrypted using the decryption key. and the changes done to the histogram during the embedding process due to this single mark are reversed. Figure 7 shows the brain MRI image and the watermarked images at different rates up to 0. CT image of liver and Ultrasound image of a fetus. The proposed algorithms outperforms Lee’s algorithm for capacities greater than 0. as describe in II-E. In these algorithms the overhead information consists of the coefficient map. Figure 8 shows the PSNR versus the embedded payload for the proposed algorithms and Lee’s algorithm [6]. where algorithm 2 performs better than algorithm 1 slightly. The image quality is evaluated by the peak signalto-noise ratio (PSNR) in dB given by PSNR = 10log10 and MSE = 1∑ 2 (Im (i) − Iw (i)) . The overhead information is updated showing the empty space size and its starting coefficient for the future use. can force the weight of the histogram shifts towards the frequency subbands of the lower wavelet levels and hence limited higher image quality can be obtained. CONCLUSION This paper proposes a high capacity reversible multiple data hiding scheme based on integer-to-integer wavelet transform and histogram shifting. EXPERIMENTAL RESULTS range. The embedding capacity is represented by bit per pixel (bpp) and the watermark is generated using the MATLAB function rand(). The medical images used are the MRI output of brain. Finally. 𝑠𝑢𝑏𝑖𝑑𝑥 parameter which indicates the subband (HL or LH) engaged in the embedding process after HH subband in each level.9 bpp. 3) The inverse integer wavelet transform is applied to obtain the host image which contains the rest of the embedded marks. IV. Algorithm 1 uses the same 𝛾1 and 𝛾2 for all the wavelet levels involved as shown in Figure 5(a). after applying integer wavelet transform the overhead information is extracted. the stop subband (since not all subbands This section focuses on the performance of the proposed scheme on grayscale medical images. Test images used in the experiment (a) (b) Fig. the stopping parameters 𝛾1 and 𝛾2 . The total payload of the called mark is extracted using the retrieved overhead information. 5. 5) The stopping parameters 𝛾1 and 𝛾2 are defined considering the acceptable PSNR and the coefficient map. n i=1 n 2552 . Algorithm 2 incorporates the importance of different wavelet levels as shown in Figure 5(b). Extraction of an embedded mark can be done in the following steps: 1) First. the last shifting coefficient at the stop subband. Coefficient masks for (a) proposed algorithm 1 and (b) proposed algorithm 2 might be needed in the last used level). and the number of passes in case of multi-pass embedding process. Three medical images of different variety are selected to test the proposed algorithms as shown in Figure 6. Test results show that the proposed scheme exhibits high image quality (in terms of PSNR) at . All the images have the same size of 512 × 512 pixels.1 bpp. Utilization of weighted stopping parameters as in algorithm 2. Two algorithms are proposed to engage the stopping parameters in different wavelet transform levels. The difference between proposed algorithm 1 and 2 is more in high embedding capacity rates. 6) Histogram shifts are performed in different levels according to the stopping parameters. 6. is implemented for comparison purpose as it gives the best performance in the expansion based scheme proposed. MSE (3) (4) where 𝐼𝑚 and 𝐼𝑤 are the original and watermarked images respectively. using block size 16×16. Lee’s algorithm [6]. the modified signal is converted back to spatial domain. the wanted watermark index and size are extracted from the overhead. V.

13. 10. pp.6 PSNR=46.A. “Circular interpretation of histogram for reversible watermarking. vol. no. Macq. 2001.R.D.M.Q. 8.” Circuits and Systems for Video Technology. 3. Rodriguez. Calderbank. 890 – 896. “Reversible watermark using the difference expansion of a generalized integer transform. “A multiple watermarking scheme applied to medical image management.E. 16. pp. aug. 2004. vol. A. Thodi and J. “Multimedia data-embedding and watermarking technologies. IEEE Transactions on. 2001 IEEE Fourth Workshop on. IEMBS ’04.3 PSNR=55.5 0. Collorec.2. 3. pp.1 0. 2007.3 0. Lotspiech. 1147 –1156.” Information Technology in Biomedicine. Kalker.3 0. .7 0.9 30 0 0. [2] G. 16. Yun-Qing Shi. pp. Alattar. M. K. IEEE Transactions on. (c) bpp=0. vol.2 0.g. 1997. feb 1995. for medical images. 250 –255. 8. Coatrieux. 2000. sept. no. ISCAS ’04. H. Kobayashi. Ultrasound 60 Proposed−1 Proposed−2 Lee et al. 2. and A. 3241 – 3244.7 0. 2000 IEEE EMBS International Conference on.M. and (e) Ultrasound of the fetus [4] [5] (a) (b) [6] [7] [8] (c) (d) [9] Fig. Birney and T. 1064 –1087. 2003. S. II – 29–32 Vol. and Wei Su.3 0. Z. pp. MRI of the brain image: (a) original images. Yoo. Sankur. Fischer.4 0. (b) bpp=0.” Image Processing. “On the modeling of dct and subband image data for compression. pp. vol. Thodi and J. IEEE Transactions on. “Reversible image watermarking based on integer-to-integer wavelet transform.” in D-Lib Mag.” in Information Technology Applications in Biomedicine.7 0. 4. Sweldons. “Relevance of watermarking in medical imaging.23dB.8 0. Pavlopoulos. 26th Annual International Conference of the IEEE. C. “Reversible watermarking by prediction-error expansion.M. 2004. C. pp. and N.” in Circuits and Systems. no. 0. Giakoumaki.J. J. 2006. IEEE Transactions on.2 0. march 2007. 2. “Wavelet transforms that map integers to integers. Shi. pp. and (d) bpp=0. and D. 2. “Reversible data embedding using a difference expansion.2 0. S. 21 – 25.” in Engineering in Medicine and Biology Society. 6th IEEE Southwest Symposium on. vol. and D. 2004. vol. Proceedings of the 2004 International Symposium on.9 30 0 0. A. 28-30 2004. 8. 345 –350. 13. 7. L. De Vleeschouwer. W. 321 –330. Morimoto. Delaigle.6 0. 1998. no. It will eliminate the need to store private data separately. 6. Mintzer. 186 –193. Koutsouris. vol.D. 65 Liver−CT Proposed−1 Proposed−2 Lee et al. 4. Yeo. Rodriguez. Sunil Lee. “High capacity lossless data hiding based on integer wavelet transform. march 2006..9 Payload(b/pixel) Payload(b/pixel) Payload(b/pixel) (a) (b) (c) Fig.” Proceedings of the IEEE.1 0. and R.6 0. 354 – 362. and B. “Multiple [12] [13] [14] [15] image watermarking applied to health information management. no. 332 – 369. vol.” Image Processing.32dB [10] [11] high embedding rates. 23-26 2004.” Image Processing. Guorong Xuan. [3] A. R EFERENCES [1] M.8 0. pp. 3.H. vol. and B.” in Image Analysis and Interpretation. Tewfik. 60 60 55 55 55 50 PSNR(dB) 50 PSNR(dB) 50 PSNR(dB) 45 45 45 40 40 40 35 35 35 30 0 0. Dec. pp. Future work includes designing a robust reversible multiple watermarking scheme which tolerates slight geometric modifications and lossy compressions such JPEG. no. vol. and T. pp. Giakoumaki. 721 –730. pp.5 0.8 0. pp. Ansari.4 0. jun 1998. Yizhan Zhen. F. J. N. “Reversible data hiding. IEEE Transactions on.C. 2004. B. Jidong Chen.” Information Forensics and Security. Ni. Distortion versus capacity graphs of the tested methods for different medical images (a) MRI of the brain. 1-5 2004.6 0. Pavlopoulos. Y.9 bpp. Daubechies.” in Multimedia Signal Processing. Maitre.9 PSNR=39. pp.4 0. Swanson. IEEE Transactions on. vol. I. “Expansion embedding techniques for reversible watermarking. 722 –732.96dB. D.” Applied and Computational Harmonic Analysis. 5. 86.Brain−MRI 65 Proposed−1 Proposed−2 Lee et al.5 0. Jun Tian. Y. Chengyun Yang. oct. e. Koutsouris. Proceedings. The proposed algorithm can be used to embed sensitive personal information into generic files such as health records and medical images. no. (b) CT scan of the liver. R. 2000. Zhicheng Ni. 2. Rolland. A. no. and Junxiang Zheng.1 0. aug. IEEE Transactions on. D.” Circuits and Systems for Video Technology. “Reversible image watermarking. 3. no.J.