Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

An introduction to watermarking of medical images

An introduction to watermarking of medical images This paper provides a preliminary investigation on digital ing as an effective technology to protect property rights and limit distribution of multimedia data. First, crucial properties and design requirements of ing schemes are discussed. Then, as ing techniques finds many applications in healthcare industry, aspects of medical image ing are raised. Nowadays, the transmission of digitized medical information has become very easy due to the generality of Internet. However, the digital form of these images can easily be manipulated and degraded. This causes problems of medical security and copyright protection and poses a great challenge to privacy protection using ing techniques. Keywords: DCT; DFT; digital ; DWT; medical images. do not always need to be hidden and they are usually associated with the digital object to be protected (cover image) or to its owner, while steganographic systems just hide any information [2, 5]. Digital image ing The digital ing technique, which is applied to many types of multimedia data, is still a developing area (e.g., [1, 4, 5, 7]). However, this paper focuses on invisible s in still images. The most popular applications of these s are connected with adding annotations, e.g., annotations to medical images related to patients identity or diagnosis. Other common applications are monitoring of images distribution or legitimate rights protection, where the could be a proof of ownership or legal possession. In many cases, also connected with medical applications, s are used to detect malicious modifications in images, as s embedded in the image undergo the same distortions as the cover image (see Figure 1). Figure 2 presents a typical ing system with two key parts: embedding and extraction. The inputs to the system are: cover image I, W and optional secret key k (to enforce the security of the system). The could be the random number sequence (see Figure 3), copyright message, ownership identifier, text, binary or other image (see Figure 4) etc. As a result ed content Iw, for invisible ing schemes should be as close to I as possible, is obtained. Next, in the distribution channel (i.e., the Internet) the ed image Iw is exposed for authorized or unauthorized attacks, which could modify it (Iw*). Then the inputs to the extractor are: modified ed content Iw*, the same secret key k and optionally the or the . Finally the output of the extractor is either the recovered W or just some confirmation if Iw* contain the W. Depending on the different combination of inputs and outputs of the extractor some types of ing schemes can be identified: nonblind ing *Corresponding author: Wioletta Wójtowicz, Institute of Automatics and Biomedical Engineering, EAIiIB Department, AGH University of Science and Technology in Krakow, 30 Mickiewicza Ave, 30-059 Krakow, Poland, E-mail: wwojtowi@agh.edu.pl Introduction Despite the obvious advantages of developing digital technologies (e.g., editing and distributing of multimedia data), there exist important disadvantages: the possibility of unlimited copying of digital content and illegal accessibility to private information. As a solution for the protection of media content different techniques of digital ing are proposed. Digital image ing is a technique that embeds information such as origin, destination or owner identifier, called the , into cover media (host media) itself. ing just like steganography is strongly connected with cryptography [9], both are techniques used to imperceptibly convey information by embedding it into the cover data. But ing has some additional requirements, i.e., robustness against possible attack, that has strong implications in the overall design of ing systems. Furthermore s 10Wójtowicz: An introduction to ing of medical images Noise modulation of with modulated Demodulation of Detected Difference Figure 1Spread spectrum ing algorithm with unauthorized embedding attack, source of code [12] (C) chest image as a cover image, source of image [13] (A) random (B) modulated random , (D) ed image. On the ed image (D) overmarking attack was proceeded ­ logo image (presented on Figure 5) was embedded in the top left corner of (D), (E) extracted demodulated , (F) extracted , (G) difference between (B) and (F) ­ unauthorized embedding was detected. Attacks W Image I Iw Embedder Iw* Extractor W' Distribution channel Secret key k Secret key k Figure 2Typical system model. ed image Extracted (systems require at least the , which is used as a hint to find where the could be in Iw), semiblind ing (does not use the original data for detection), blind ing (it requires neither the I nor the embedded W). For real-world digital ing systems in still images, a few very general properties, shared by all proposed systems, can be identified. One of them is imperceptibility, which is a common requirement and independent of the application purpose. Imperceptibility means that embedded in the image must be invisible to the human eye. In some applications it is often not allowed to alter the image contents by even one bit of information Figure 3Simple ing scheme in DWT domain without any attacks, source of code [12] (A) ­ CT head, source of image [13] (B) as a random series of 1 and ­1 values, (C) ed image ­ was embedded in the LL component of wavelet decomposition (the low frequency region ­ lowpass filters to the columns and to the rows, see Figure 5), then image was reconstructed, (D) extracted ­ as a difference between (A) and (C). Wójtowicz: An introduction to ing of medical images11 ed image ed image ­ only the most significant bits ­ only the most significant bits ed image Gray image and gray LSB=1 LSB=2 LSB=3 LSB=4 LSB=5 LSB=6 LSB=7 ­ only the most significant bits PSNR, dB 1.5 2.0 2.5 3.0 3.5 Amount of bits 5.0 ×106 Figure 4LSB algorithm for the cover image (anterior curvature of the spine), source of image [14] and a as a logo (inserted in the top left corner of the cover image). In all scenarios from the only the most significant bits are selected and embedded in the cover image. Algorithm is elaborated for values of LSB from 1 to 7, (A) value of LSB is equal to 1, (B) value of LSB is equal to 4, (C) value of LSB is equal to 7, (D) PSNR in dB as a function of amount of ed bits in cover image for all tested LSB values. (e.g., medical imaging: [1, 3, 6, 8]). The requirement of imperceptibility can be fulfilled by selecting a reversible ing method which can recover the original cover image by undoing the embedding process at the receiving end after the image verification process is completed. To evaluate the imperceptibility, generally, peak signal to noise ratio (PSNR) can be used. The PSNR is the least mean square errors between an original and ed image (Figure 4D). Additional requirements, like robustness against modifications and/or malicious attacks, have to be taken into consideration when designing ing techniques as well. However, there are applications where it is less important than for others [2]. security is essential, because in most applications, such as copyright protection, the secrecy of embedded information needs to be assured [10, 11]. Capacity of the ing system is defined as the maximum amount of information that can be embedded in the cover work. The computational cost of ing algorithm should be as little as possible, because a ing method with high complex algorithms will require more software as well as hardware resources. A specific class of s, are fragile s, typically used for authentication of multimedia data. Unlike robust s, any attack on the image invalidates the fragile s present in the image and helps in detecting/identifying any tampering of the image (Figure 1). Therefore a fragile ing scheme should detect tampering with high probability and characterize modifications to the image. s should be perceptually transparent and the system should not require the image at decoding. Typical fragile image ing techniques embed the 12Wójtowicz: An introduction to ing of medical images in the least significant bit planes of an image. An example of this method is presented in the Figure 4. To sum up some issues connected with designing ing systems should be identified. First, if the ed image Iw is unmodified the detected W should be exactly the same as W (see Figure 3). Secondly, for robust ing if Iw is modified, W should still match W well to give a clear judgment of the existence of the . Finally, for fragile ing, W will be totally different from W after even the slight modification to Iw. What is more W indicate the possible tampering to Iw* and give information about degradation of Iw (see Figure 1). cover image as a function of the amount of bits and LSB values. It is observed that the larger the value of LSB is, the more bits of the that can be embedded and the lower the imperceptibility of the system. As a result, even if inserting a in spatial domain enables the fast process of embedding and extracting of the message, this method suffers from low robustness and low imperceptibility. For these reasons transform domains are more popular, especially in security applications (e.g., [9]). Other advantages of transform methods arise from their connection with compression standards (i.e., JPEG and JPEG standards are based on DCT and on DWT methods accordingly). ing domains Although there exist ing methods for almost all types of multimedia data, the number of image ing methods is much larger than for the other types of media. In image ing, the signal is either embedded into the spatial domain representation of the image, or one of many transform domain representations such as discrete Fourier transform [2, 5, 7], discrete cosine transform [2, 8, 11], and Discrete wavelet transform [1, 2, 8, 10, 11]. It is generally argued that embedding s in transform domains provides better robustness against attacks and leads to less perceptibility of an embedded due to the spread of a signal over many spatial frequencies and better modeling of the human visual system (HVS) when using transform coefficients. Transform domains Discrete Fourier transform As an image of size N1×N2 could be treat as two-dimensional signal f(n1, n2), where n1=0, 1, 2, ..., N1-1, n2=0, 1, 2, ..., N2-1, DFT enables to decompose it to some periodic, sinusoidal and cosinusoidal, components. For each pair of frequencies (k1, k2) N1×N2 DFT could be defined as following F ( k1 , k 2 ) = 1 N 1N 2 N 1 -1 N 2 -1 n1 =0 n2 =0 f ( n ,n ) exp - i 2 n1k1 i 2 n2 k2 - N1 N2 Spatial domains Least significant bit (LSB) modification is a common method for embedding the in the spatial domain. The method consists of the manipulation of LSBs of images in a manner which is not detectable and imperceptible to the human eye. The basic idea is that the LSBs of the original 8-bit gray level image are discarded first, then the LSBs are replaced by the permuted binary of the same size as the original one. In the Figure 4A­C three scenarios are presented of embedding the in the medical image, when the value of LSB is equal to 1, 4 and 7, respectively. The least distortion in the cover image is observed when the is embedded only on the last significant bit of the image, but on the other hand, the amount of embedded information is very small. Figure 4D presents changes of quality of the Applying this transformation, the image is considered to be a function of frequencies. In this representation the image has some new advantages in terms of ing applications. First, shifting in spatial domain is cleared like a linear shifting in the DFT phase. Second, scaling in the spatial domain appears as an inverse scaling in the frequency domain. Third, rotation of image just identical to in the spatial domain is caused the rotation of DFT of image at the same extent. As a result robustness against shifting of image in DFT is assured. But other geometrical attacks like rotation and scaling have an effect on DFT magnitude. But it is possible to avoid this problem easily by also using logpolar mapping (LMP). This transformation, also called the Fourier-Mellin transform (see [2, 7]), brings scaling and rotation to translation. This property is very important, especially as s in DFT domain are robust against translations. As a consequence, construction of the transform domain, which could also be insensitive to rotations and scaling is possible. In ing techniques, which are based on DFT transform, DFT is used two times. Firstly to change the domain of the Wójtowicz: An introduction to ing of medical images13 original cover image and secondly to change the coordinates to log-polar ones. Discrete cosine transform DCT is akin to DFT, but it has only real values, as it transforms spatial image values to the sum of cosinusoidal functions for different sequences N N F ( k1 ,k2 )= n1-1 n2-10 C ( n1 )C ( n2 ) f ( n1 ,n2 ) =0 = ( 2 n1+1) k1 ( 2 n2 +1) k2 cos cos , 2N 2N where C(n1)=C(n2)= k1=k2=1, 2, ..., N-1. 2 1 for for k1=k2=0 or C(n1)=C(n2)= N N be added to the coefficients of the transformation, that are exposure and frequency functions. The entity of this method is decomposition of the image, as a two-dimensional signal, to the sequence of signals with decreasing resolution. The algorithm in every step, decomposes the image into four subimages (image and with horizontal, vertical and diagonal details of the image) with dimensions equal to quarter of the input image dimension (see Figure 5). The subimages could be next decomposed recursively in the same way and for nth level of decomposition, the following components of the image are obtained: Ln = [Ln1 *[ Ln2 * Ln-1 ]2 ,1 ] 1,2 DnV = [Ln1 *[ Hn2 * Ln-1 ]2 ,1 ] 1,2 Dn H =[Hn1 *[ Ln2 * Ln-1 ]2 ,1 ] 1,2 Dn D =[H n1 *[ H n2 * Ln-1 ]2 ,1 ] 1,2 where: * is the convolution operator, L0 is the , 2,1 are the sampling of rows, 1,2 are the sampling of columns, H n1 , Ln1 are highpass and lowpass filters along the rows, H n2 , Ln2 are highpass and lowpass filters along the columns. In practice to get the wavelet decomposition, highpass filter (H) and lowpass filter (L) are applied to the rows and columns of the image in the spatial domain in four possible combinations: LL, LH, HL, HH, that refer to the components presented in Figure 5. Mathematically wavelet transformation of the image is the convolution operation on image pixels. The lowpass filters remove the high frequency elements from the image. Thus from the image the convulsive changes in intensity of adjacent pixels are removed. Due to this the edges on the image seemed to be blurred. These filters are usually used to eliminate noise or to slightly smooth the image. The highpass filters the high frequency elements from the image. These filters are used to specify the details of the image or specify the The most prominent property of this transform is its low sensitivity to the destruction of values in high frequencies. Most information regarding signal is placed in the low frequency region. In ing practice after computing DCT of image the frequencies are divided into three main categories: low, middle and high frequencies. Due to the fact, that the low frequencies are sensitive to image distortions and high frequencies are eliminated during JPEG compression, the is embedded to the middle frequencies. Embedding the in these coefficients enables robustness against compression (which is impossible in DFT domain) and some other distortions [8, 11]. This method has also some disadvantages such as lack of robustness to geometrical distortions (rotation, scaling and translation). Discrete wavelet transform Opposite to DFT and DCT, DWT is a hierarchical transformation, which enable analysis of the image in the spatialfrequency domain. This transformation provides excellent space for image ing, as the can LL Decomposition HL Reconstructed image LH HH Figure 5One step of wavelet decomposition of the image, source of code [12] (A) the image before decomposition, (B) decomposition after applying one step of DWT, (C) the reconstructed image. 14Wójtowicz: An introduction to ing of medical images edges of objects. It has been shown that details of the image are placed in the low frequency region (LL), therefore inserting the in this area could damage the quality of the image. On the other hand, high frequency coefficients include less information (HH), so they are not robust against compression. For these reasons the is usually inserted in the middle frequency area (LH and HL) (e.g., [7]). Medical image ing Due to recent technological development, the digital world is considered to be effective, convenient and secure. Many businesses, including the medical industry, exploit proposed multimedia solutions and find a lot of new challenges in this area. Telemedicine applications in teleconsulting, and telediagnoisis and telesurgery play a vital role in the evolution of the healthcare domain today. As a result the transmission, storage and sharing of electronic medical data has become common practice. For example the medical images can be given to the patient directly or sent to the patient by online. They can also be maintained as a soft and hard copy in the hospital for diagnosing and for other purposes such as for finding new drugs and scientific research. The following is a simple scenario for an e-diagnosis case. A physician takes an image, and performs his own diagnosis based on the image and then embeds this information as a in the image along with patient's personal information such as name, age, and sex, etc. The ed image is then stored on the database containing the patient historical data in hospital. Another scenario is when the physician sends the image to another physician for second opinion. He or she again makes his/her own diagnosis, embeds the information in the image and sends it back to the sender. This exchange of medical images is usually performed through an unsecure open environment like the Internet. Thus, in these cases medical images can easily be copied or tampered with for illegal purposes, for example, getting a fake health insurance claim from an insurance company. Also tampering of medical images can have serious consequences regarding treatment. For these reasons protecting medical images from forgery has become a very important issue. To guarantee the security, authenticity and management of medical images and information through storage and distribution, the ing techniques are evolving to protect the medical healthcare information [1, 3, 6, 8]. Digital ing can be a solution in resolving these problems: hiding data for the purpose of inserting metadata (to render the image more usable) and information protection (with the applications like copyright protection and content authentication). Constant efforts have been made to provide security solutions to ensure: 1. Authentication ­ images are from the correct sources and go to specified recipients. 2. Integrity ­ received images are not modified during transmission by unauthorized users. 3. Confidentiality ­ medical image transmission cannot be accessed by unauthorized parties. Review of the algorithims There exist many algorithms for inserting s into transform domains. Many of them proceeded from a combination of wavelets and other transforms to improve robustness and the imperceptibility of ing schemes. Manoochehri et al. in [7] propose an algorithm based on the combination of DWT and the Fourier-Mellin transform. In this method after computing the DWT of the cover image and of the , regions of middle frequencies (LH or HL) are selected. Afterwards FMT is applied to both components and finally these subbands are combined. Experimental results proved that DWT and FMT assure robustness against noise and geometrical attacks accordingly. Huai-bin et al. [11] suggest combining DCT and DWT. To ensure the security of the system a secret key based on the Arnold transformation of the image is added. The algorithm involves three level wavelet decomposition of the , and the selection of the LL component and division of this subband into non overlapping blocks. Then DCT is carried out on each block and the is embedded in the middle frequency region. Performance of the algorithm against common attacks such as Gaussian noise, JPEG compression and cropping is satisfactory, but when the ed image is subject to image scaling attack the algorithm does not perform very well. Lin et al. propose a blind digital image ing algorithm based on multi-strategy [10]. The described method is a combination of DWT decomposition, spread spectrum technique and the Arnold transformation. Applying the distinct dimensional Arnold transformation to the coefficient of four subband image: LH1, HL1 and LH2, HL2 allows the secret key to be obtained. Since intensity of the embedding has difference affecting on the various resolution layers of the wavelet factor, difference embedding strategies were employed. The method presented rises the imperceptibility and the robustness to resist various attacks. As a result many properties of the ing scheme like robustness to JPEG compression, adding noise, etc were improved. Wójtowicz: An introduction to ing of medical images15 ing methods despite their attractiveness in multimedia may encounter limitations in medical images. For example the ing process usually alters the cover image in an irreversible way and may hide some important information necessary for the physician to make the diagnosis. Thus, medical image ing is a special subcategory of image ing in the sense that the images have special requirements. Medical image ing methods try to preserve the image diagnosis quality value by avoiding critical information loss [3]. Requirements of medical image ing ­ Imperceptibility Particularly, as the clinical reading of the images (e.g., for diagnosis) must not be affected, ed medical images should not differ perceptually from cover images. ­ Robustness Medical images are usually ed with the identification codes of the physicians who created the images in order to authenticate them. As previously mentioned, in the telemedicine environment medical images may go through several services and receive different processing and annotations. Thus, s should be robust so that the authenticity of image can be verified in such environments. ­ Capacity In the ing of medical images, all the information necessary for the physician is embedded such as identification of patient, diagnosis report, origin identification (who created the image). This information is further increased when the image is sent to other physician for second opinion. Therefore, capacity for embedding the payload should must be high. ­ Authenticity Only entitled users like patients, personnel and clinicians should have access to the medical data. For this purpose secret keys could be used. For example, a containing the physician's digital signature or identification code might be embedded in the image for identity verification. ­ Reversibility Another assumption is that only the authentic user should be able to reverse the embedding process to extract the original data from the image. This produces the uned to the user which can further use it for making a diagnosis or other research purposes without having an ambiguity about the integrity of image. ­ Intactness of ROI A difficult requirement is that the image may not undergo any degradation that will affect the reading of images. Generally, images are required to remain intact to achieve this, with no visible alteration to their original form. A medical image is comprised of a region of interest (ROI) and a region of non-interest (RONI). The ROI contains the important information on which some decision is made. It is therefore mandatory that the ing process should not affect the ROI adversely. A distorted ROI will lead to a wrong diagnosis. One solution of this problem is to embed the information in the RONI, thereby keeping the ROI intact. Complexity In the telemedicine environment usually images are sent from some remote location to a central hospital for a second opinion from another physician or consultation on the spot. Thus in this case speed becomes an important factor. This demands that the algorithm should be less complex to save execution time. Regarding the applications, medical image ing systems can be broken into three broad categories: robust, fragile and semi-fragile [6]. Robust s are designed to resist attempts to remove or destroy the . They are used primarily for copyright protection and content tracking. Many traditional robust methods are spread-spectrum, whereby the is spread over a wide range of image frequencies. A number of robust medical image ing systems have been developed [8]. On the other hand fragile s are used to determine whether an image has been tampered with or modified. The is destroyed if the image is manipulated in any manner. Fragile s are often capable of localization, and are used to determine where modifications were made to an image. Traditional methods embed checksums or pseudo-random sequences in the LSB plane (see Figures 1 and 4). Semi-fragile s combine the properties of both robust and fragile s. Like robust methods, they can tolerate some degree of change to the ed image. Like fragile methods, they are capable of localizing regions of an image that are authentic and those that have been altered. Conclusion and future directions In this paper some overview of ing techniques as a solution to protect the content of images and assure legitimate use of them was provided. The study discussed underlying requirements and threats to medical image ing and provides some practical examples. This 16Wójtowicz: An introduction to ing of medical images is primarily due to the special nature of medical images, which should not be perceptually altered. This preliminary study has shown that medical image ing is still an open field of research. It appears that a selection of different s for different medical image types is the most appropriate solution to the generic problem. Thus, future work in this area should elaborate many types of medical images, s and algorithms. A lot of possible scenarios must be considered to obtain ed medical images with less degradation and to have recovered with better accuracy. One of the objective is to develop biometricbased ing schemes, in which the content of the is strictly correlated with the image content, for increased robustness, security and accuracy. These algorithms may not affect the quality of the or the recognition performance. Received October 31, 2012; revised December 9, 2012; accepted January 11, 2013; previously published online February 23, 2013 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Bio-Algorithms and Med-Systems de Gruyter

An introduction to watermarking of medical images

Wójtowicz, Wioletta
Bio-Algorithms and Med-Systems , Volume 9 (1) – Mar 1, 2013
Download PDF
8 pages

Loading next page...
 
/lp/de-gruyter/an-introduction-to-watermarking-of-medical-images-6k4asdQTUp

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher
de Gruyter
Copyright
Copyright © 2013 by the
ISSN
1895-9091
eISSN
1896-530X
DOI
10.1515/bams-2013-0002
Publisher site
See Article on Publisher Site

Abstract

This paper provides a preliminary investigation on digital ing as an effective technology to protect property rights and limit distribution of multimedia data. First, crucial properties and design requirements of ing schemes are discussed. Then, as ing techniques finds many applications in healthcare industry, aspects of medical image ing are raised. Nowadays, the transmission of digitized medical information has become very easy due to the generality of Internet. However, the digital form of these images can easily be manipulated and degraded. This causes problems of medical security and copyright protection and poses a great challenge to privacy protection using ing techniques. Keywords: DCT; DFT; digital ; DWT; medical images. do not always need to be hidden and they are usually associated with the digital object to be protected (cover image) or to its owner, while steganographic systems just hide any information [2, 5]. Digital image ing The digital ing technique, which is applied to many types of multimedia data, is still a developing area (e.g., [1, 4, 5, 7]). However, this paper focuses on invisible s in still images. The most popular applications of these s are connected with adding annotations, e.g., annotations to medical images related to patients identity or diagnosis. Other common applications are monitoring of images distribution or legitimate rights protection, where the could be a proof of ownership or legal possession. In many cases, also connected with medical applications, s are used to detect malicious modifications in images, as s embedded in the image undergo the same distortions as the cover image (see Figure 1). Figure 2 presents a typical ing system with two key parts: embedding and extraction. The inputs to the system are: cover image I, W and optional secret key k (to enforce the security of the system). The could be the random number sequence (see Figure 3), copyright message, ownership identifier, text, binary or other image (see Figure 4) etc. As a result ed content Iw, for invisible ing schemes should be as close to I as possible, is obtained. Next, in the distribution channel (i.e., the Internet) the ed image Iw is exposed for authorized or unauthorized attacks, which could modify it (Iw*). Then the inputs to the extractor are: modified ed content Iw*, the same secret key k and optionally the or the . Finally the output of the extractor is either the recovered W or just some confirmation if Iw* contain the W. Depending on the different combination of inputs and outputs of the extractor some types of ing schemes can be identified: nonblind ing *Corresponding author: Wioletta Wójtowicz, Institute of Automatics and Biomedical Engineering, EAIiIB Department, AGH University of Science and Technology in Krakow, 30 Mickiewicza Ave, 30-059 Krakow, Poland, E-mail: wwojtowi@agh.edu.pl Introduction Despite the obvious advantages of developing digital technologies (e.g., editing and distributing of multimedia data), there exist important disadvantages: the possibility of unlimited copying of digital content and illegal accessibility to private information. As a solution for the protection of media content different techniques of digital ing are proposed. Digital image ing is a technique that embeds information such as origin, destination or owner identifier, called the , into cover media (host media) itself. ing just like steganography is strongly connected with cryptography [9], both are techniques used to imperceptibly convey information by embedding it into the cover data. But ing has some additional requirements, i.e., robustness against possible attack, that has strong implications in the overall design of ing systems. Furthermore s 10Wójtowicz: An introduction to ing of medical images Noise modulation of with modulated Demodulation of Detected Difference Figure 1Spread spectrum ing algorithm with unauthorized embedding attack, source of code [12] (C) chest image as a cover image, source of image [13] (A) random (B) modulated random , (D) ed image. On the ed image (D) overmarking attack was proceeded ­ logo image (presented on Figure 5) was embedded in the top left corner of (D), (E) extracted demodulated , (F) extracted , (G) difference between (B) and (F) ­ unauthorized embedding was detected. Attacks W Image I Iw Embedder Iw* Extractor W' Distribution channel Secret key k Secret key k Figure 2Typical system model. ed image Extracted (systems require at least the , which is used as a hint to find where the could be in Iw), semiblind ing (does not use the original data for detection), blind ing (it requires neither the I nor the embedded W). For real-world digital ing systems in still images, a few very general properties, shared by all proposed systems, can be identified. One of them is imperceptibility, which is a common requirement and independent of the application purpose. Imperceptibility means that embedded in the image must be invisible to the human eye. In some applications it is often not allowed to alter the image contents by even one bit of information Figure 3Simple ing scheme in DWT domain without any attacks, source of code [12] (A) ­ CT head, source of image [13] (B) as a random series of 1 and ­1 values, (C) ed image ­ was embedded in the LL component of wavelet decomposition (the low frequency region ­ lowpass filters to the columns and to the rows, see Figure 5), then image was reconstructed, (D) extracted ­ as a difference between (A) and (C). Wójtowicz: An introduction to ing of medical images11 ed image ed image ­ only the most significant bits ­ only the most significant bits ed image Gray image and gray LSB=1 LSB=2 LSB=3 LSB=4 LSB=5 LSB=6 LSB=7 ­ only the most significant bits PSNR, dB 1.5 2.0 2.5 3.0 3.5 Amount of bits 5.0 ×106 Figure 4LSB algorithm for the cover image (anterior curvature of the spine), source of image [14] and a as a logo (inserted in the top left corner of the cover image). In all scenarios from the only the most significant bits are selected and embedded in the cover image. Algorithm is elaborated for values of LSB from 1 to 7, (A) value of LSB is equal to 1, (B) value of LSB is equal to 4, (C) value of LSB is equal to 7, (D) PSNR in dB as a function of amount of ed bits in cover image for all tested LSB values. (e.g., medical imaging: [1, 3, 6, 8]). The requirement of imperceptibility can be fulfilled by selecting a reversible ing method which can recover the original cover image by undoing the embedding process at the receiving end after the image verification process is completed. To evaluate the imperceptibility, generally, peak signal to noise ratio (PSNR) can be used. The PSNR is the least mean square errors between an original and ed image (Figure 4D). Additional requirements, like robustness against modifications and/or malicious attacks, have to be taken into consideration when designing ing techniques as well. However, there are applications where it is less important than for others [2]. security is essential, because in most applications, such as copyright protection, the secrecy of embedded information needs to be assured [10, 11]. Capacity of the ing system is defined as the maximum amount of information that can be embedded in the cover work. The computational cost of ing algorithm should be as little as possible, because a ing method with high complex algorithms will require more software as well as hardware resources. A specific class of s, are fragile s, typically used for authentication of multimedia data. Unlike robust s, any attack on the image invalidates the fragile s present in the image and helps in detecting/identifying any tampering of the image (Figure 1). Therefore a fragile ing scheme should detect tampering with high probability and characterize modifications to the image. s should be perceptually transparent and the system should not require the image at decoding. Typical fragile image ing techniques embed the 12Wójtowicz: An introduction to ing of medical images in the least significant bit planes of an image. An example of this method is presented in the Figure 4. To sum up some issues connected with designing ing systems should be identified. First, if the ed image Iw is unmodified the detected W should be exactly the same as W (see Figure 3). Secondly, for robust ing if Iw is modified, W should still match W well to give a clear judgment of the existence of the . Finally, for fragile ing, W will be totally different from W after even the slight modification to Iw. What is more W indicate the possible tampering to Iw* and give information about degradation of Iw (see Figure 1). cover image as a function of the amount of bits and LSB values. It is observed that the larger the value of LSB is, the more bits of the that can be embedded and the lower the imperceptibility of the system. As a result, even if inserting a in spatial domain enables the fast process of embedding and extracting of the message, this method suffers from low robustness and low imperceptibility. For these reasons transform domains are more popular, especially in security applications (e.g., [9]). Other advantages of transform methods arise from their connection with compression standards (i.e., JPEG and JPEG standards are based on DCT and on DWT methods accordingly). ing domains Although there exist ing methods for almost all types of multimedia data, the number of image ing methods is much larger than for the other types of media. In image ing, the signal is either embedded into the spatial domain representation of the image, or one of many transform domain representations such as discrete Fourier transform [2, 5, 7], discrete cosine transform [2, 8, 11], and Discrete wavelet transform [1, 2, 8, 10, 11]. It is generally argued that embedding s in transform domains provides better robustness against attacks and leads to less perceptibility of an embedded due to the spread of a signal over many spatial frequencies and better modeling of the human visual system (HVS) when using transform coefficients. Transform domains Discrete Fourier transform As an image of size N1×N2 could be treat as two-dimensional signal f(n1, n2), where n1=0, 1, 2, ..., N1-1, n2=0, 1, 2, ..., N2-1, DFT enables to decompose it to some periodic, sinusoidal and cosinusoidal, components. For each pair of frequencies (k1, k2) N1×N2 DFT could be defined as following F ( k1 , k 2 ) = 1 N 1N 2 N 1 -1 N 2 -1 n1 =0 n2 =0 f ( n ,n ) exp - i 2 n1k1 i 2 n2 k2 - N1 N2 Spatial domains Least significant bit (LSB) modification is a common method for embedding the in the spatial domain. The method consists of the manipulation of LSBs of images in a manner which is not detectable and imperceptible to the human eye. The basic idea is that the LSBs of the original 8-bit gray level image are discarded first, then the LSBs are replaced by the permuted binary of the same size as the original one. In the Figure 4A­C three scenarios are presented of embedding the in the medical image, when the value of LSB is equal to 1, 4 and 7, respectively. The least distortion in the cover image is observed when the is embedded only on the last significant bit of the image, but on the other hand, the amount of embedded information is very small. Figure 4D presents changes of quality of the Applying this transformation, the image is considered to be a function of frequencies. In this representation the image has some new advantages in terms of ing applications. First, shifting in spatial domain is cleared like a linear shifting in the DFT phase. Second, scaling in the spatial domain appears as an inverse scaling in the frequency domain. Third, rotation of image just identical to in the spatial domain is caused the rotation of DFT of image at the same extent. As a result robustness against shifting of image in DFT is assured. But other geometrical attacks like rotation and scaling have an effect on DFT magnitude. But it is possible to avoid this problem easily by also using logpolar mapping (LMP). This transformation, also called the Fourier-Mellin transform (see [2, 7]), brings scaling and rotation to translation. This property is very important, especially as s in DFT domain are robust against translations. As a consequence, construction of the transform domain, which could also be insensitive to rotations and scaling is possible. In ing techniques, which are based on DFT transform, DFT is used two times. Firstly to change the domain of the Wójtowicz: An introduction to ing of medical images13 original cover image and secondly to change the coordinates to log-polar ones. Discrete cosine transform DCT is akin to DFT, but it has only real values, as it transforms spatial image values to the sum of cosinusoidal functions for different sequences N N F ( k1 ,k2 )= n1-1 n2-10 C ( n1 )C ( n2 ) f ( n1 ,n2 ) =0 = ( 2 n1+1) k1 ( 2 n2 +1) k2 cos cos , 2N 2N where C(n1)=C(n2)= k1=k2=1, 2, ..., N-1. 2 1 for for k1=k2=0 or C(n1)=C(n2)= N N be added to the coefficients of the transformation, that are exposure and frequency functions. The entity of this method is decomposition of the image, as a two-dimensional signal, to the sequence of signals with decreasing resolution. The algorithm in every step, decomposes the image into four subimages (image and with horizontal, vertical and diagonal details of the image) with dimensions equal to quarter of the input image dimension (see Figure 5). The subimages could be next decomposed recursively in the same way and for nth level of decomposition, the following components of the image are obtained: Ln = [Ln1 *[ Ln2 * Ln-1 ]2 ,1 ] 1,2 DnV = [Ln1 *[ Hn2 * Ln-1 ]2 ,1 ] 1,2 Dn H =[Hn1 *[ Ln2 * Ln-1 ]2 ,1 ] 1,2 Dn D =[H n1 *[ H n2 * Ln-1 ]2 ,1 ] 1,2 where: * is the convolution operator, L0 is the , 2,1 are the sampling of rows, 1,2 are the sampling of columns, H n1 , Ln1 are highpass and lowpass filters along the rows, H n2 , Ln2 are highpass and lowpass filters along the columns. In practice to get the wavelet decomposition, highpass filter (H) and lowpass filter (L) are applied to the rows and columns of the image in the spatial domain in four possible combinations: LL, LH, HL, HH, that refer to the components presented in Figure 5. Mathematically wavelet transformation of the image is the convolution operation on image pixels. The lowpass filters remove the high frequency elements from the image. Thus from the image the convulsive changes in intensity of adjacent pixels are removed. Due to this the edges on the image seemed to be blurred. These filters are usually used to eliminate noise or to slightly smooth the image. The highpass filters the high frequency elements from the image. These filters are used to specify the details of the image or specify the The most prominent property of this transform is its low sensitivity to the destruction of values in high frequencies. Most information regarding signal is placed in the low frequency region. In ing practice after computing DCT of image the frequencies are divided into three main categories: low, middle and high frequencies. Due to the fact, that the low frequencies are sensitive to image distortions and high frequencies are eliminated during JPEG compression, the is embedded to the middle frequencies. Embedding the in these coefficients enables robustness against compression (which is impossible in DFT domain) and some other distortions [8, 11]. This method has also some disadvantages such as lack of robustness to geometrical distortions (rotation, scaling and translation). Discrete wavelet transform Opposite to DFT and DCT, DWT is a hierarchical transformation, which enable analysis of the image in the spatialfrequency domain. This transformation provides excellent space for image ing, as the can LL Decomposition HL Reconstructed image LH HH Figure 5One step of wavelet decomposition of the image, source of code [12] (A) the image before decomposition, (B) decomposition after applying one step of DWT, (C) the reconstructed image. 14Wójtowicz: An introduction to ing of medical images edges of objects. It has been shown that details of the image are placed in the low frequency region (LL), therefore inserting the in this area could damage the quality of the image. On the other hand, high frequency coefficients include less information (HH), so they are not robust against compression. For these reasons the is usually inserted in the middle frequency area (LH and HL) (e.g., [7]). Medical image ing Due to recent technological development, the digital world is considered to be effective, convenient and secure. Many businesses, including the medical industry, exploit proposed multimedia solutions and find a lot of new challenges in this area. Telemedicine applications in teleconsulting, and telediagnoisis and telesurgery play a vital role in the evolution of the healthcare domain today. As a result the transmission, storage and sharing of electronic medical data has become common practice. For example the medical images can be given to the patient directly or sent to the patient by online. They can also be maintained as a soft and hard copy in the hospital for diagnosing and for other purposes such as for finding new drugs and scientific research. The following is a simple scenario for an e-diagnosis case. A physician takes an image, and performs his own diagnosis based on the image and then embeds this information as a in the image along with patient's personal information such as name, age, and sex, etc. The ed image is then stored on the database containing the patient historical data in hospital. Another scenario is when the physician sends the image to another physician for second opinion. He or she again makes his/her own diagnosis, embeds the information in the image and sends it back to the sender. This exchange of medical images is usually performed through an unsecure open environment like the Internet. Thus, in these cases medical images can easily be copied or tampered with for illegal purposes, for example, getting a fake health insurance claim from an insurance company. Also tampering of medical images can have serious consequences regarding treatment. For these reasons protecting medical images from forgery has become a very important issue. To guarantee the security, authenticity and management of medical images and information through storage and distribution, the ing techniques are evolving to protect the medical healthcare information [1, 3, 6, 8]. Digital ing can be a solution in resolving these problems: hiding data for the purpose of inserting metadata (to render the image more usable) and information protection (with the applications like copyright protection and content authentication). Constant efforts have been made to provide security solutions to ensure: 1. Authentication ­ images are from the correct sources and go to specified recipients. 2. Integrity ­ received images are not modified during transmission by unauthorized users. 3. Confidentiality ­ medical image transmission cannot be accessed by unauthorized parties. Review of the algorithims There exist many algorithms for inserting s into transform domains. Many of them proceeded from a combination of wavelets and other transforms to improve robustness and the imperceptibility of ing schemes. Manoochehri et al. in [7] propose an algorithm based on the combination of DWT and the Fourier-Mellin transform. In this method after computing the DWT of the cover image and of the , regions of middle frequencies (LH or HL) are selected. Afterwards FMT is applied to both components and finally these subbands are combined. Experimental results proved that DWT and FMT assure robustness against noise and geometrical attacks accordingly. Huai-bin et al. [11] suggest combining DCT and DWT. To ensure the security of the system a secret key based on the Arnold transformation of the image is added. The algorithm involves three level wavelet decomposition of the , and the selection of the LL component and division of this subband into non overlapping blocks. Then DCT is carried out on each block and the is embedded in the middle frequency region. Performance of the algorithm against common attacks such as Gaussian noise, JPEG compression and cropping is satisfactory, but when the ed image is subject to image scaling attack the algorithm does not perform very well. Lin et al. propose a blind digital image ing algorithm based on multi-strategy [10]. The described method is a combination of DWT decomposition, spread spectrum technique and the Arnold transformation. Applying the distinct dimensional Arnold transformation to the coefficient of four subband image: LH1, HL1 and LH2, HL2 allows the secret key to be obtained. Since intensity of the embedding has difference affecting on the various resolution layers of the wavelet factor, difference embedding strategies were employed. The method presented rises the imperceptibility and the robustness to resist various attacks. As a result many properties of the ing scheme like robustness to JPEG compression, adding noise, etc were improved. Wójtowicz: An introduction to ing of medical images15 ing methods despite their attractiveness in multimedia may encounter limitations in medical images. For example the ing process usually alters the cover image in an irreversible way and may hide some important information necessary for the physician to make the diagnosis. Thus, medical image ing is a special subcategory of image ing in the sense that the images have special requirements. Medical image ing methods try to preserve the image diagnosis quality value by avoiding critical information loss [3]. Requirements of medical image ing ­ Imperceptibility Particularly, as the clinical reading of the images (e.g., for diagnosis) must not be affected, ed medical images should not differ perceptually from cover images. ­ Robustness Medical images are usually ed with the identification codes of the physicians who created the images in order to authenticate them. As previously mentioned, in the telemedicine environment medical images may go through several services and receive different processing and annotations. Thus, s should be robust so that the authenticity of image can be verified in such environments. ­ Capacity In the ing of medical images, all the information necessary for the physician is embedded such as identification of patient, diagnosis report, origin identification (who created the image). This information is further increased when the image is sent to other physician for second opinion. Therefore, capacity for embedding the payload should must be high. ­ Authenticity Only entitled users like patients, personnel and clinicians should have access to the medical data. For this purpose secret keys could be used. For example, a containing the physician's digital signature or identification code might be embedded in the image for identity verification. ­ Reversibility Another assumption is that only the authentic user should be able to reverse the embedding process to extract the original data from the image. This produces the uned to the user which can further use it for making a diagnosis or other research purposes without having an ambiguity about the integrity of image. ­ Intactness of ROI A difficult requirement is that the image may not undergo any degradation that will affect the reading of images. Generally, images are required to remain intact to achieve this, with no visible alteration to their original form. A medical image is comprised of a region of interest (ROI) and a region of non-interest (RONI). The ROI contains the important information on which some decision is made. It is therefore mandatory that the ing process should not affect the ROI adversely. A distorted ROI will lead to a wrong diagnosis. One solution of this problem is to embed the information in the RONI, thereby keeping the ROI intact. Complexity In the telemedicine environment usually images are sent from some remote location to a central hospital for a second opinion from another physician or consultation on the spot. Thus in this case speed becomes an important factor. This demands that the algorithm should be less complex to save execution time. Regarding the applications, medical image ing systems can be broken into three broad categories: robust, fragile and semi-fragile [6]. Robust s are designed to resist attempts to remove or destroy the . They are used primarily for copyright protection and content tracking. Many traditional robust methods are spread-spectrum, whereby the is spread over a wide range of image frequencies. A number of robust medical image ing systems have been developed [8]. On the other hand fragile s are used to determine whether an image has been tampered with or modified. The is destroyed if the image is manipulated in any manner. Fragile s are often capable of localization, and are used to determine where modifications were made to an image. Traditional methods embed checksums or pseudo-random sequences in the LSB plane (see Figures 1 and 4). Semi-fragile s combine the properties of both robust and fragile s. Like robust methods, they can tolerate some degree of change to the ed image. Like fragile methods, they are capable of localizing regions of an image that are authentic and those that have been altered. Conclusion and future directions In this paper some overview of ing techniques as a solution to protect the content of images and assure legitimate use of them was provided. The study discussed underlying requirements and threats to medical image ing and provides some practical examples. This 16Wójtowicz: An introduction to ing of medical images is primarily due to the special nature of medical images, which should not be perceptually altered. This preliminary study has shown that medical image ing is still an open field of research. It appears that a selection of different s for different medical image types is the most appropriate solution to the generic problem. Thus, future work in this area should elaborate many types of medical images, s and algorithms. A lot of possible scenarios must be considered to obtain ed medical images with less degradation and to have recovered with better accuracy. One of the objective is to develop biometricbased ing schemes, in which the content of the is strictly correlated with the image content, for increased robustness, security and accuracy. These algorithms may not affect the quality of the or the recognition performance. Received October 31, 2012; revised December 9, 2012; accepted January 11, 2013; previously published online February 23, 2013

Journal

Bio-Algorithms and Med-Systems – de Gruyter

Published: Mar 1, 2013

There are no references for this article.