1 Introduction

Developments in digital and communication technologies has opened up new opportunities in the field of telemedicine where digital medical images and EPR are transmitted over networks for clinical interpretation and diagnosis [11]. However, transmission over public network is prone to infringement of security, confidentiality, copyright, and integrity. We cannot afford any loss or tampering of medical data as it can lead to mis-diagnosis. Thus security, confidentiality, integrity and authenticity are of prime concern during transmission of medical images through public network and must be simultaneously satisfied [4]. Confidentiality of the transmitted medical images is ensured so that only authorized users can access it. Integrity validates whether the digital medical image is intact or tampered. Authenticity verifies whether the medical image is from the correct source and belongs to the claimed patient. In the present scenario the two major methodologies i.e. cryptography and medical image watermarking are used popularly to meet these requirements [1]. Cryptography-based approach has a major limitation that if the medical image is deciphered, or the digital signature is deleted or lost, then it is difficult to verify integrity and authenticity of the image. Therefore, MIW has come up as a promising solution in this scenario [21]. Digital watermarks such as EPR, signature, hospitals logo or trademark or physician ID can be embedded into the medical image without degrading the visual quality of the medical image to provide confidentiality and authenticity [2]. A fragile or cryptographic hash watermark can be embedded to verify the integrity of medical images. Other than ensuring security, confidentiality and integrity MIW also provides other benefits like avoiding detachment, non-repudiation, controlling access, privacy, captioning, indexing, memory and bandwidth saving [26].

Different medical image modalities such as X-ray, ultrasound, MRI, CT-scan, mammogram, etc. are being used for medical imaging [22]. In medical images, information is non-uniformly distributed across the image. Taking advantage of this fact, region based MIW schemes divide medical image into region of interest (ROI) and region of non-interest (RONI) [41]. ROI contains more informative part of the medical image which is used for the diagnosis. RONI has no significant role in diagnosis. ROI and RONI regions have different characteristics and requirements. Therefore, different types of watermarks are embedded to meet the watermarking goal. Watermarking process should keep the ROI intact as any distortions may cause wrong diagnosis. RONI can be used to embed robust watermarks. Usually ROI/RONI selection is done manually by doctors/radiologists.

In general digital image watermarking is defined as process of embedding watermark into the host/cover image [8]. Watermark embedding is done either in spatial/pixel domain or transform domain [5]. In spatial domain watermarking techniques, the watermark is interleaved by directly modifying the intensity/pixel values of the cover image resulting in low computational complexity. LSB embedding, spread spectrum are some of the popularly used spatial techniques. Spatial techniques are fragile. But this fragility can be advantageous in finding out the distorted area and reproduction of original host image from the distorted one [3]. In transform domain schemes, watermark is inserted by modifying transform coefficients of the cover/host image. DFT, DCT, DWT, IWT, SVD are commonly used transforms. Transform domain techniques are comparatively more robust [21]. Each transform has different characteristics and has different representation of the image. One transform may be robust to one set of attacks and the other transform to the other set of attacks. In the past few decades various spatial domain, transform domain and hybrid MIW scheme has been reported in literature [23, 26]. Related MIW schemes are discussed in detail in Section 2.

Organization of the paper is as follows: Related work is discussed in Section 2. The proposed MIW scheme is explained in detail in Section 3. Section 4 provides experimental results and discussions. Finally, Section 5 concludes the paper.

2 Related work

During the last few years various MIW schemes have been proposed for different objectives. These MIW schemes can be classified as irreversible, reversible and region-based schemes [1]. Irreversible watermarking schemes introduce distortions to the original images by irretrievable watermarking operations such as truncation, quantization or bit replacement. On the other hand, reversible MIW schemes are capable of restoring original pixel values in the watermarked images. But, most of them lack tamper localization functionality and cannot verify integrity of medical images. Region based schemes have tamper localization features which offers integrity for medical images. Region-based watermarking schemes divide the digital medical image into ROI and RONI [24]. ROI and RONI have different characteristics; therefore, we can embed different types of watermarks to accomplish different requirements. Some of the related region based MIW schemes are discussed below.

Naseem et al. [27] proposed a fragile watermarking method using chaotic key and residue number system. The original ROI pixel values are replaced by ROI residue. Hash value is embedded in the RONI at the locations selected by using chaotic key. The LSBs of these pixel values are set to zero. In this scheme ROI is reversible whereas RONI is irreversible. Hajjaji et al. [16] proposed a spatial MIW scheme where EPR and digital signature of hospital data generated by using SHA1 are concatenated. Then it is interleaved in the edge pixels using LSB technique. They used turbo algorithm to check error during transmission. Pandey et al. [28] proposed a multiple MIW scheme for tele-ophthalmology applications using fusion of DWT and SVD. Hash value of the iris portion of the digital eye image is generated using SHA-512. They embedded four watermarks i.e. index, reference watermark, caption, and signature simultaneously. Text and image watermarks are inserted in the RONI. A major drawback of hash code watermark is that it is more time consuming and a slight change in hash code leads to false tamper localization at the receivers side. Solanki and Malik [32], proposed DWT based watermarking scheme. Using RSA encryption algorithm the watermark is encrypted and then embedded in the ROI. However, in this scheme pre-processing is required before actual embedding. Pre-processing adds overhead time. Such cryptographic watermarks are computation-intensive and therefore cannot be used in real-time environments. Amit et al. [33] suggested a multiple watermarking scheme based on DWT-DCT-SVD transform. They have applied simplified encryption technique to increase the security of the text watermark and at the same time save execution time.

Another region based MIW scheme is proposed by Thabit and Khoo [40] for tamper detection, localization, and its recovery. In this scheme ROI feature is extracted using IWT which is further embedded in RONI. This scheme uses Slantlet transform (SLT) for embedding data in ROI and RONI. Authors claim to recover tampered ROI with good visual quality. Memon et al. [25] proposed hybrid watermarking scheme where the medical image is partitioned into ROI and RONI by using vector quantization technique. A fragile watermark is embedded in ROI LSBs. Location map is generated by distributing RONI into blocks of size N × N. A watermark is embedded in the RONI coefficients. High time complexity required for calculations to generate the location map is bottleneck of this scheme. Singh et al. [34] proposed a DWT-DCT-SVD based scheme. In the embedding process, the host image and watermark image is transformed by DCT-SVD transform. This scheme provides higher imperceptibility and robustness. Another DWT, DCT, and SVD based multiple watermarking scheme is presented by Singh [35]. In this scheme for identity authentication, multiple watermarks are embedded into the same medical image or multimedia data simultaneously. For enhanced security of EPR data is encrypted before embedding into the cover image. Singh et al. [36] proposed secure multiple watermarking scheme using spread-spectrum for medical images. Watermark is embedded using selective DWT coefficients by thresholding the coefficient values of column. Error correcting code (ECC) is applied to the text watermark and binary image of doctor’s signature or telemedicine centre name to enhance security. Singh et al. [37] suggested a secure multiple watermarking scheme based on DWT and SVD where the image along with the encrypted and encoded text acts as a watermarks. Guo and Zhuang [14] suggested another MIW scheme based on modified DE for authentication and data hiding. Singh et al. [38] developed a non-blind robust dual MIW scheme based on DWT and SVD using error correcting codes. Text and image watermarks are embedded in radiological image for secure and compact medical data transmission. They advocated use of a hybrid model of two of the ECCs i.e. BCH and repetition code for higher robustness.

Parah et al. [29] presented two different blind MIW algorithms. For embedding the watermark(EPR) relative value of pre-selected DCT coefficients of 8 × 8 block is used in both the algorithms. In the first algorithm the watermark is embedded in the whole medical image whereas in the second approach watermark is embedded in the RONI only. It lacks ROI tamper detection and recovery features. Moreover, embedding capacity is also low. Wu et al. [17] proposed a reversible data hiding method with contrast enhancement for medical images. Another region-based watermarking scheme is proposed by Al-Haj and Amer [1] using both spatial and frequency domains. Robust watermarks are embedded in RONI to provide confidentiality and authenticity using DWT-SVD transform. Fragile watermark is embedded in ROI of medical image by LSB embedding technique to provide integrity. But, this scheme has limitations of low embedding capacity, low accuracy of tamper detection and high computation time. Table 1 illustrates the comparison of different MIW schemes.

Table 1 Literature review of existing MIW schemes (RB-region based, TL-tamper localization, TR-tamper recovery)
Full size table

2.1 Motivation and contribution of the proposed work

From the literature review we have two observations. Firstly, most of the SVD based watermarking techniques in existing literature suffers from false positive detection problem [6]. Secondly, many MIW techniques dealing with tamper detection, localization and recovery have low accuracy in tamper detection and poor visual quality of recovered ROI [1, 19, 42]. Intactness of ROI is very important as it is used for diagnosis. Tampered ROI can lead to wrong diagnosis. Therefore, it is very important to detect ROI tampering accurately. Recovery of tampered ROI with good visual quality is another crucial aspect. We are motivated by these issues. The main contribution of our work is as follows:

  • False positive free: In false positive problem the fabricated watermark is extracted from an arbitrary image where the embedded watermark is completely different from the extracted watermark. To embed the watermark W, SVD is applied on it. Watermark is decomposed into three matrices as given by Eq. 1.

    $$ W = U_{W}S_{W}{V_{W}^{T}} $$
    (1)

    Only the singular values matrix of watermark S W is embedded while the orthogonal matrices U W and V W are not embedded. In extraction procedure, only the diagonal matrix S W is extracted, whereas U W and V W is available at the receiver end. However, the orthogonal matrices U W and V W contain the major information about an image. Thus any one can provide a fake pair of orthogonal matrices and claim that his watermark is present in the watermarked image, which causes false positive detection problem as shown in Fig. 1.

    Matrix U of SVD has a unique property i.e. all the elements in the first column have very close values. In the same elements of first row of V matrix have close values. By analyzing U, we observe that there is a strong correlation between the (U 2,1) and (U 3,1) element [9, 20]. Similarly, (V 1,2) and (V 1,3) elements of V matrix are strongly correlated. We have explored this property of SVD to overcome the false positive problem.

  • SVD has a notable stability and robustness against the common image manipulation (such as filtering, histogram equalization, noise addition, etc.) and geometric attacks (such as cropping, rotation) etc. Whereas, IWT provides tolerance towards compression algorithms and filtering. IWT-SVD has been used for providing robustness to the watermarks embedded in RONI against these attacks although it may increases computational complexity to some extent.

  • Tamper detection, localization and recovery with higher accuracy: pixel-wise positional and relational bits are calculated and embedded in ROI, to detect and localize altered pixels with high precision. Visual quality of recovered pixel is also good.

  • Improved performance: In the comparative study of proposed scheme with the other existing techniques in [1, 29], it is observed that the proposed scheme offers superior performance in terms of imperceptibility of watermarked image, robustness, payload, temper detection accuracy and visual quality of recovered ROI.

  • Multiple watermarks are embedded simultaneously in the medical image [28]. Multiple watermarks i.e. EPR, hospital logo and ROI LSBs are embedded in the same medical image simultaneously for providing confidentiality, authentication and integrity respectively. Reduced storage and less bandwidth requirements are the additional benefit that we get as by-product.

Fig. 1
figure 1

False positive detection

Full size image

3 Proposed technique

The proposed region-based MIW scheme provides confidentiality, authenticity, and integrity for medical images transmitted in telemedicine applications. For embedding the watermarks we firstly divide medical image into ROI and RONI. ROI in medical images can be selected in several ways such as selecting a rectangle or polygon manually by the physician/radiologist, defining a threshold [12], defining seeds [15] or edge detection [23] etc. Mostly, automated ROI detection schemes are specific to a particular image modality and any alterations in the medical image, may yield different ROI boundaries. Keeping the limitations of automated ROI selection in consideration we use manual selection of rectangular ROI by the physician/radiologist. The advantage of representing rectangular ROI is that only two coordinate points i.e. top left and bottom right needed to specified instead of a group of coordinates [23]. The ROI is watermarked in the spatial domain, whereas the RONI is watermarked in the frequency domain using IWT-SVD hybrid transform to achieve different MIW objectives. Spatial techniques are fragile and hence can be used for ROI tamper detection. EPR and hospital logo is embedded in RONI to provide confidentiality and authentication respectively. It is important that these watermarks should be robust. We have used IWT and SVD transform to provide robustness to these watermarks against various signal and geometric attacks. The proposed MIW scheme has four modules i.e. ROI embedding, ROI extraction, RONI embedding and RONI extraction. These modules are discussed in detail in following sub sections.

3.1 ROI embedding

ROI is used for diagnosis. Hence its integrity verification is important. Positional and relational bits are calculated and embedded in ROI pixels using LSB technique for pixel-wise ROI integrity verification as shown in Fig. 2. Original LSBs of ROI pixels are extracted and embedded in RONI to serve as backup for restoring tampered ROI pixels.

Fig. 2
figure 2

Embedding of positional and relational bit in ROI pixels

Full size image

3.1.1 Steps for ROI embedding process

Step 1::

Extract the LSBs of ROI pixels and concatenate them as a single robust watermark for embedding in the RONI.

Step 2::

Calculate the positional bit with respect to its position i.e. row and column. Positional bit is calculated using 6 MSBs of pixel P i represented as b a where a 𝜖 (2...7). P 1 represents bitwise XOR operation between row and pixel value whereas P 2 represents bitwise XOR operation between column and pixel value. P 1 and P 2 is calculated using (2) and (3).

$$ P1=XOR(b_{a-2}^{r},b_{a}) $$
(2)
$$ P2=XOR(b_{a-2}^{c},b_{a}) $$
(3)

Where a =7,6...2 and b r and b c are the binary value of row and column value. Now, calculate positional bit using (4), and embed it in the first LSB of pixel P i .

$$ Positional\;bit=\left\{ \sum\limits_{i=1,2,..6}(P1_{i}\wedge P2_{i}) \right\}mod\;2 $$
(4)
Step 3::

Relational bit shows the relation between 6 MSBs of P i . Calculate relational bit using (5) where b i are the bits of a pixel P i .

$$ Relational\;bit=\left\{ \sum\limits_{i=7,6,..3}(b_{i}\oplus b_{i-1}) \right\}mod\;2 $$
(5)

Embed the relational bit in the second LSB of pixel P i .

Step 4::

Repeat step 2 and step 3 for each ROI pixel to obtain watermarked R O I ′.

3.1.2 Analysis of tampering effects

We analyze the probability of tampering to the watermarked R O I ′. Probability of alteration in positional and relational bits due to altered MSBs is calculated as follows:

Positional bit

Let b r and b c be the binary form of row and column values of R O I ′. The total number of possible alterations for any pixel by MSBs is as follows:

$$ \alpha_{1}=\sum\limits_{i=1,2,..6}{{~}^{6}C_{i}} $$
(6)

Assume that in a pixel P i , k bits of the 6 MSBs are altered, then total number of sets of MSBs that affects the positional bit is calculated as follows:

$$ \alpha_{2}=\sum\limits_{1\leq i\leq k}\binom{k}{i}2^{6-k}\quad \;\;where\; i\; is \;odd $$
(7)

Above equation represents those combinations, for which \(b^{r}_{(a-2)}\) and \(b^{c}_{(a-2)}\) is not same for any bit. Therefore, calculate a multiplicative factor to get the actual number of altered set of MSBs. In the image I of size X × Y, if X ≤ 15 or Y ≤ 15 or both are satisfied then, the multiplicative factor γ is calculated by (8).

$$ \gamma =1-\left[ \frac{\left\{ (X-15)+(Y-15)\right\}}{\left( X\times Y\right)} \right] $$
(8)

Therefore, the number of pixels, that affects the positional bits, can be obtained by (9).

$$ \alpha_{3}=\gamma \times \alpha_{2} $$
(9)

Thus, the probability of alteration in positional bit due to change in pixels is given by (10).

$$ Pr_{(p)}=\frac{\alpha_{3}}{\alpha_{1}} $$
(10)

Relational bit

Relational bit can detect alteration in P i only if the altered pixels contain one altered MSB boundary bit i.e. either 2nd or 7th. Maximum number of possible alterations in MSBs is β 1, given by (11).The total number of combination of MSBs for P i which exactly contains one boundary bit is given by (11) and (12).

$$ \beta_{1}=\sum\limits_{i=1,2,3,..6}{{~}^{6}C_{i}}-\sum\limits_{i=1,2,3,..5}{{~}^{5}C_{i}} $$
(11)
$$ \beta_{2}=\frac {1}{2}\times \sum\limits_{i=1,2,..4}{{~}^{4}C_{i}} $$
(12)

Hence, the number of pixels which actually affects the relational bits β 3, is calculated as follows:

$$ \beta_{3}=\beta_{1}-\beta_{2} $$
(13)

Therefore, the probability of changes in relational bit is given by (14).

$$ Pr_{(r)}=\frac{\beta_{3}}{\alpha_{1}} $$
(14)

Now, the probability for a pixel to get detected can be calculated by (15).

$$ Pr_{(tot)}=Pr_{(p)}+Pr_{(r)} $$
(15)

3.2 ROI extraction

ROI integrity is verified by comparing the calculated and extracted values of positional and relational bits. If the calculated value and extracted values of positional and relational bits matches then the pixels are considered to be intact otherwise tampered. In case of tampering we recover the pixel by using recovery bits extracted from RONI. Steps of ROI extraction is explained below.

Step 1::

For all pixels of watermarked R O I ′, calculate the positional bit using (4). Compare the calculated positional bit with the extracted first LSB of corresponding pixel. If there is match then step 2 is performed. If there is mismatch, mark those pixels as altered one and perform step 3.

Step 2::

Calculate the relational bit by using (5). Compare the calculated relational bit with the extracted second LSB of corresponding pixel. If there is mismatch then those pixels will be marked as altered pixel.

Step 3::

Replace the two LSBs of pixels with the original ROI LSBs extracted from RONI to get recovered ROI.

3.3 RONI embedding

In the proposed scheme three different robust watermarks i.e. EPR, hospital logo (binary image watermark) and ROI LSBs are embedded as robust watermarks as shown in Fig. 3. RONI is watermarked using IWT-SVD hybrid transform to provide higher robustness and imperceptibility. IWT provides robustness, tolerance to various compression algorithms and filtering. It maps integer to integer and thus avoids fractional computations. It allows construction of lossless retrieval. Wavelet transform have excellent spatial localization, frequency spread, and multi-resolution characteristics, are similar to the theoretical models of the human visual system [39]. SVD provides a good stability and robustness against the common image manipulation (such as histogram equalization, filtering, noise addition, etc.). RONI pixels are divided into 8 × 8 blocks and IWT is applied to each block. For each block, four sub-bands CA, CV, CH and CD of 4 × 4 is obtained. SVD is applied on CV, CH and CD sub-bands yielding their corresponding U, S and V matrix. Two watermark bits are embedded in a 4 × 4 sub block. First bit is embedded by adjusting second and third element of first column (U 2,1 and U 3,1) of U matrix. Whereas, second bit is embedded by adjusting second and third element of first row (V 1,2 and V 1,3) of V matrix. There is a strong correlation between (U 2,1 and U 3,1) of U matrix and (V 1,2 and V 1,3) of V matrix. This property has been explored for embedding watermark bits. We embed watermark bits of EPR, hospital logo and ROI LSBs in CV, CH and CD respectively. For additional security and robustness of EPR we use BCH error correction code. Detail description of RONI embedding process is provided below.

Fig. 3
figure 3

Block diagram of RONI embedding

Full size image

3.3.1 Steps of RONI embedding process

Step 1::

Perform BCH encoding (error correction codes) to the 7 bit ASCII representation of text watermark (EPR) bits as follows

$$ W_{b}=BCH(watermark\; bits) $$
(16)
Step 2::

Convert each of the three robust watermarks i.e. W b , hospital logo and ROI LSBs in the form of 1-D bit patterns.

Step 3::

Divide RONI into 8 × 8 non-overlapping blocks.

Step 4::

Apply the 1-level IWT on the 8 × 8 RONI block. Four sub-bands CA, CV, CH and CD of size 4 × 4 are obtained.

Step 5::

Apply SVD on CV, CH and CD sub-bands

$$ X = U_{X}S_{X}{V^{T}_{X}} \quad\quad\quad\quad where \; X \epsilon (CV,CH, CD) $$
(17)
Step 6::

Embed two watermark bits each of encoded EPR (W b ), hospital logo and ROI LSBs simultaneously in CV, CH and CD sub-band respectively. One bit each is embedded by adjusting (U 2,1 and U 3,1) and (V 1,2,V 1,3) coefficients according to the adjustment process explained in algorithm 1.

Step 7::

Perform inverse SVD on watermarked U, S X , and watermarked V matrices for each IWT sub-band.

Step 8::

Apply inverse IWT on the sub-bands CA and watermarked CV, CH, CD.

Step 9::

Repeat through step 4 to step 8 until all watermark bits are embedded in their respective sub-bands to construct the watermarked RONI.

3.3.2 Algorithm for SVD coefficient adjustment

Watermark bits are embedded by adjusting relations between (U 2,1 and U 3,1), (V 1,2 and V 1,3) coefficients of U and V matrices. For embedding watermark bit 1, the difference between U 2,1 and U 3,1 should be negative and greater than threshold T. To embed watermark bit 0, the difference between U 2,1 and U 3,1 should be positive and greater than the threshold T. When these two conditions are violated, SVD coefficients are adjusted according to the Algorithm 1. Similarly, V 1,2 and V 1,3 coefficients of V matrix are adjusted for embedding watermark bit. The process of SVD coefficient adjustment is shown in shown in Fig. 4. The flowchart indicating the detailed steps for SVD coefficient adjustment is shown in Fig. 5. Value of threshold T is determined empirically according to the trade off between imperceptibility and robustness. At high threshold, robustness is high but imperceptibility is low.

figure c
Fig. 4
figure 4

SVD coefficient adjustment

Full size image
Fig. 5
figure 5

Flowchart for SVD coefficient adjustment

Full size image

3.4 Extraction from RONI

Extraction from RONI is just reverse process of embedding in RONI. Figure 6 shows the block diagram of RONI extraction process. Steps for extraction from RONI is explained below.

Step 1::

Divide watermarked RONI into 8 × 8 blocks.

Step 2::

Apply 1-level IWT on 8 × 8 RONI block. Four sub-bands CA, CV, CH and CD of size 4 × 4 are obtained.

Step 3::

Apply SVD on CV, CH and CD sub-bands as following

$$ X = U_{X}S_{X}{V^{T}_{X}} \quad\quad\quad\quad where \; X\epsilon(CV,CH, CD) $$
(18)
Step 4::

From each sub-band the first watermark bit \(W^{\prime }_{1}\) is extracted from (U 2,1,U 3,1) coefficients of the U matrix using following relation.

$$ W_{1}^{\prime}=\left\{\begin{array}{lll} 0& \text{ if}\,\, U_{2,1}>U_{3,1} \\ 1& \text{ if}\,\, U_{2,1}\leq U_{3,1} \end{array}\right. $$
(19)

The second watermark bit \(W^{\prime }_{2}\) is extracted by applying the following relation between (V 1,2,V 1,3).

$$ W_{2}^{\prime}=\left\{\begin{array}{lll} 0& \text{ if}\,\, V_{1,2}>V_{1,3} \\ 1& \text{ if}\,\, V_{1,2}\leq V_{1,3} \end{array}\right. $$
(20)
Step 5::

Repeat step 2 to step 4, until all watermarked RONI blocks are processed.

Step 6::

Reconstruct the three watermarks by cascading relevant watermark bits extracted from the respective sub-bands of all blocks.

Step 7::

Perform BCH decoding to the watermark bits extracted from CD sub-band to get the extracted EPR(text watermark).

Fig. 6
figure 6

Block diagram of RONI extraction

Full size image

In the proposed scheme, experimentation is validated on wide range of gray scale medical images of different modalities. However, proposed scheme can be generalized for the color medical images also. We have applied proposed scheme for color medical images using RGB color model. In RGB color model red, green and blue channel can be extracted and treated as a grayscale image. Watermarks are embedded in the blue channel as explained in Sections 3.1 and 3.3. Then the original blue channel in the RGB image is replaced by the watermarked blue channel to obtain the watermarked color image. Watermarks are extracted from watermarked blue channel according to the extraction process explained in Sections 3.2 and 3.4.

4 Experimental Results and Discussions

4.1 Experimental configuration

Experiments have been conducted to evaluate the performance of the proposed scheme in terms of imperceptibility, robustness, time performance, tamper detection and recovery. Experimentation is done using MATLAB R2013a on a PC having Intel core2 Duo, 3.00 GHz, 2GB RAM. Gray scale and color medical images of different modalities have been used as test cover image as shown in Figs. 7 and 8 respectively. All the cover images are of size 512 × 512. Binary watermark of hospital logo having size 64 × 32 and EPR of 2048 bits are embedded in the medical images for authenticity and confidentiality respectively as shown in Fig. 9. For all experimentation we have taken threshold T as 0.05.

Fig. 7
figure 7

Grayscale medical cover images of different modalities

Full size image
Fig. 8
figure 8

Color medical cover images of different modalities

Full size image
Fig. 9
figure 9

Watermarks: Hospital logo, EPR

Full size image

4.2 Performance measures

The performance of the MIW schemes can be assessed on the basis of imperceptibility, robustness and payload. For objective evaluation of imperceptibility between original and watermarked medical images peak signal-to-noise ratio (PSNR) and structure similarity measure index (SSIM) are used. Normalized correlation (NC) and bit error rate (BER) are used to measure the similarity and differences between original watermark and extracted watermark. Evaluation metrics used to evaluate imperceptibility and robustness of proposed scheme are given in Table 2.

Table 2 Evaluation metric used for analysis and comparisons
Full size table

4.3 Imperceptibility test

Imperceptibility is one of the important requirments of MIW scheme. Watermarked image should not have any perceivable distortion as it will be used for diagnosis. Peak signal to-noise ratio (PSNR) and structure similarity measure index (SSIM) are used for objective evaluation of imperceptibility between original and watermarked medical images. Higher the PSNR value, higher is the imperceptibility. Empirically tested threshold value of PSNR is 35dB. From Tables 3 and 4 we can observe that average PSNR value of gray scale and color medical images is about 45dB which is more than the empirical threshold. SSIM value is about 0.97 which is near to the ideal value indicates that the watermarked images have no significant distortions. PSNR and SSIM values are consistently good for all the tested image modalities therefore it is anticipated that it will yield good imperceptibility with other modalities also. Subjective evaluation of visual quality of watermarked images show no considerable perceivable difference between original and watermarked image as shown in Fig. 10 for gray scale images and Fig. 11 for color images with corresponding PSNR values. Objective and subjective evaluations, reveals that the proposed scheme introduced no distortions, thus we can claim that imperceptibility requirement is achieved.

Fig. 10
figure 10

Watermarked grayscale images with their corresponding PSNR values(dB)

Full size image
Fig. 11
figure 11

Watermarked color images with their corresponding PSNR values(dB)

Full size image
Table 3 PSNR and SSIM of gray scale medical images
Full size table
Table 4 PSNR and SSIM of color medical images
Full size table

4.4 Robustness test

Robustness is another important requirement of MIW. It is necessary that the watermarks should survive intentional/unintentional attacks with minimum distortion as they will be used for authentication of ownership and originating source and integrity verification. Robustness of the proposed scheme has been tested against various attacks like salt and pepper noise, histogram equalization, JPEG compression, average filtering and cropping attacks.

4.4.1 Robustness against salt and pepper attack

Watermarked medical images have been subjected to salt and pepper noise of varying densities (0.001 and 0.01). NC and BER of image watermark and EPR extracted from different medical images under various intensities of salt and pepper noise is provided in Table 5. In all cases NC values are more than 0.9870 (more than the threshold). Thus, extracted watermarks are clearly visible. BER is 0 in all cases for extracted EPR when noise density is 0.001 and in all other cases it is close to 0. NC and BER values do not vary significantly even at higher noise density which indicates that the proposed scheme is effective at high noise density also. Therefore, from the results of Table 5 it is evident that the proposed scheme effectively resists the salt and pepper attack.

Table 5 NC and BER of image watermark and EPR extracted from different medical images under various intensities of salt and pepper noise
Full size table

4.4.2 Robustness against JPEG compression

Watermarked images may undergo JPEG compression to reduce media storage and transmission costs. Medical images of different modalities have been tested for JPEG compression under different quality factors. Higher the quality factor lower is the compression and vice versa. Thus, the quality of extracted watermark increases with the increase in quality factor as shown in Table 6. It can be observed from Table 6 that the proposed scheme can resist JPEG compression at low quality factor also. NC values are more than the empirical threshold indicating that extracted watermarks are easily distinguishable. BER value decreases with the increase in JPEG quality factor, and approaches to zero when JPEG quality factor is higher than 50. Proposed scheme shows good resistance against JPEG compression.

Table 6 NC and BER of image watermark and EPR extracted from different medical images under various quality factors of JPEG compression
Full size table

4.4.3 Robustness against histogram equalization

Histogram equalization attack changes most of the pixel values of image. Under this attack, performance of proposed algorithm is given in Table 7. NC values indicate that the extracted watermarks are distinctly visible. The results in Table 7 shows that the proposed algorithm is effective for the histogram equalization as BER is less than 1.

Table 7 NC and BER of image watermark and EPR extracted from different medical images under histogram Equalization
Full size table

4.4.4 Average filtering

The watermarked images were subjected to average filtering attack with mask size 3 × 3 and the results are tabulated in Table 8. The NC value offered by the scheme is quite acceptable and the BER value is closer to 0 as shown in Table 8.

Table 8 NC and BER of image watermark and EPR extracted from different images under average filtering
Full size table

4.4.5 Cropping

A cropping attack from top left corner and center is applied on the watermarked medical image using mask size 25 % of dimension of the image. The results obtained after cropping 25 % from top left corner and 25 % center in terms of NC and BER is given in Table 9. NC and BER values shows that the proposed scheme sustains both of the cropping attacks.

Table 9 NC and BER of image watermark and EPR extracted from different images under cropping attack
Full size table

Most of the popular attacks in image processing has been applied extensively to authenticate the robustness of the proposed scheme in this section. Also, payload of the proposed algorithm is analyzed and experimentally proved that with this payload; all the attacks are sustained. Whereas, other attacks included in the standard watermark benchmarking tools can be studied in future.

4.5 Tamper detection, localization and recovery

Erasing and copy-paste tampering are imposed on the watermarked image to evaluate tamper detection, localization and recovery as shown in Fig. 12. Accuracy of tamper detection is upto 98 % as given in Table 10. It is also evident from the table that PSNR values of recovered image are very close to the PSNR of untampered watermarked image. Subjective evaluation from Fig. 12 further affirms that recovered images have good visual quality. Thus, the proposed scheme successfully meets the ROI integrity requirement with high accuracy.

Fig. 12
figure 12

Tamper localization and recovery. (a) Watermarked image (b) Localization of erase tampering (c) Recovery of erase tampering (d) Copy-paste tampering (e) Localization of copy & paste tampering (f) Recovery of copy-paste tampering

Full size image
Table 10 Performance of tamper detection and PSNR value of recovered medical images
Full size table

4.6 Time performance

Time performance of the proposed technique is evaluated for gray scale and color medical images of different modalities. From Figs. 13 and 14 it can be observed that the relative time requirements for the watermarking is moderate and regardless of image modality. The prime objective of our scheme is to obtain good imperceptibility, higher robustness, tamper detection and recovery with higher accuracy. Therefore, modest increase in computation time due to SVD can be seen as trade off to achieve these objectives.

Fig. 13
figure 13

Embedding and extraction time for gray-scale medical images

Full size image
Fig. 14
figure 14

Embedding and extraction time for color medical images

Full size image

4.7 Comparative analysis

To further validate the performance of the proposed scheme, it is compared with existing schemes i.e. Al-Haj and Amer’s scheme [1] and Parah et al. scheme [29]. In Al-Haj and Amer’s scheme [1] the author had proposed a region based blind scheme using DWT-SVD transform. Local fragile watermark is embedded in ROI of the image using a reversible scheme in the spatial domain to provide integrity. Three robust watermarks are embedded in RONI. Parah et al. [29] proposed two different MIW algorithms using DCT. In both of the algorithms two coefficients of 8 × 8 DCT block are selected and their magnitudes are compared to embed the watermark/EPR bits. In the first algorithm watermark is embedded in the entire image whereas in the second algorithm watermark is embedded in RONI only.

4.7.1 Imperceptibility

The average PSNR acheieved in Al-Haj and Amer’s scheme [1] is about 33 dB and in case of Parah et al. [29] it is about 40 dB and 54 dB in the first and second algorithm respectively. Average PSNR of our scheme is higher than Al-Haj and Amer’s scheme [1] and the first algorithm of Parah et al. [29]. In the second algorithm of Parah et al. [29] average PSNR is more than our scheme but at the cost of low payload. So, it is concluded that our scheme has better imperceptibility than these existing schemes without sacrificing payload.

4.7.2 Robustness

NC and BER values of extracted image watermark and EPR of the proposed scheme is compared with and Al-Haj and Amer [1] in Table 11 under various attacks. It can be observed from Table 11 that NC and BER values of proposed scheme is better than the compared scheme in all attacks. NC value decreases significantly with increase in noise density in Al-Haj and Amer [1] scheme. Also, the quality of extracted watermark deteriorates significantly with decrease in JPEG quality factor. Therefore Al-Haj and Amer [1] scheme cannot be used effectively at high compression ratio and noise densities. The proposed scheme outperforms it in terms of robustness. Besides this tamper detection of Al-Haj and Amer [1] can be easily forged. Whereas, tamper detection and localization is comparatively more accurate in our scheme.

Table 11 Comparison of NC and BER of image watermark and EPR extracted from MRI image under various attacks using the proposed scheme and Al-haj and Amer’s scheme [1]
Full size table

NC and BER values of extracted image watermark and EPR of the proposed scheme is compared with Parah et al. [29] in Table 12. It can be concluded from Table 12 that proposed scheme is more robust than Parah et al. [29] except in histogram equalization which can be seen as future improvement work. Parah et al. [29] does not verify integrity of medical images as it lacks feature for tamper detection, localization and recovery.

Table 12 Comparison of NC and BER of image watermark and EPR extracted from CT scan image under various attacks using the proposed scheme and Parah et al. scheme [29]
Full size table

4.7.3 Payload

The payload of the proposed scheme is compared with the schemes under comparison. In both the algorithms of Parah et al. [29], EPR and authentication watermarks of pre-specified size are embedded. Hence, payload in their scheme is constant i.e. 3072 and 2209 bits in the first and second algorithm respectively. Whereas, the proposed and Al-Haj and Amer scheme [1] embeds ROI LSBs in RONI as recovery bits. But, the size of ROI LSBs are dependent on the manual selection of ROI. Comparison of payload of the proposed scheme and other existing schemes is provided in Table 13. It can be seen that embedding capacity of our scheme is comparatively higher than other schemes.

Table 13 Payload comparison
Full size table

Example:-

For illustration, consider that a rectangular ROI region of size 64 × 32 be selected manually in both of the algorithms for fair comparison.

Total selected ROI bits = 64 × 32 = 2048 bits

According to Ali and Amers scheme[3]:

Cover image size = 2048 × 2048 = 4194304 pixels

Size of EPR = 204 × 96 = 19584 bits

Size of authentication watermark = 81 × 50 = 4050 bits

Selected ROI bits = 64 × 32 = 2048 bits

Total payload = 25682 bits

$$\begin{array}{@{}rcl@{}} Embedding\; capacity &=& \frac{(Total \; number \; of \; watermark \; bits)}{(Total \; number \; of \; pixels)}\\ &=&25682\; bits/4194304\; pixels \\ &=& 0.0061\; bpp \end{array} $$

Thus the embedding capacity of Al-Haj and Amers scheme [1] is 0.0061 bpp.

Now, according to our proposed scheme:

Cover image size = 512 × 512 = 262144 pixels

Size of EPR = 2048 bits

Size of authentication watermark = 64 × 32 = 2048 bits

Selected ROI bits = 64 × 32 = 2048 bits

Total payload = 6144 bits

$$\begin{array}{@{}rcl@{}} Embedding\; capacity &=& \frac{(Total \; number \; of \; watermark \; bits)}{(Total \; number \; of \; pixels)}\\ &=&25682\; bits/4194304\; pixels \\ &=& 0.0234\; bpp \end{array} $$

The embedding capacity of proposed scheme is 0.0234 bpp.

4.7.4 Time complexity

Time complexity of an algorithm signifies the total time taken by an algorithm to run as a function of the input size to the problem. Time complexity of the proposed scheme is compared with Bao and Ma [7], Su et al. [31], Chang et al. [10], Hu and Hsu [18] and Patra et al. [30]. Computational complexity of the transformations used popularly in the watermarking schemes is listed in Table 14. Computations involving a single element operation are not taken into account, as it is negligible in the comparison to that of matrix operations. According to the asymptotic estimates shown in Table 14, the transforms and matrix operations required for the schemes under comparison are summed up and tabulated in Table 15. The total number of operations and the sizes of input data for all the schemes under comparison is also specified in Table 15. Most of the computations of the proposed scheme are confined to a matrix of size 4 × 4, thus the overall computational requirement is alleviated. IWT is lifting based transform therefore requires half the number of computations as compared to DWT [13]. Due to these obvious reasons the proposed scheme demands relatively lesser amount of computation as compared to the schemes under comparison developed from the DWT-SVD or DCT framework.

Table 14 Computational complexity of matrix operation/transforms
Full size table
Table 15 Computational complexity comparison
Full size table

5 Conclusion

This paper presents a new region based MIW scheme using IWT-SVD transform to address the issues of authentication, confidentiality and integrity in tele-medicine. IWT-SVD hybrid transform is applied to include the good stability of SVD and the capability of IWT to preserve a perfect reconstruction, by integer to integer mapping. It also offers high resistance against geometrical and non-geometrical attacks. Proposed scheme is blind and free from false positive problem. From the experimental analysis, it is evident that the proposed scheme has good imperceptibility and higher robustness against most of the attacks such as salt and pepper, histogram equalization, JPEG compression, average filtering and cropping.

Proposed scheme successfully detects and locates tampering with high accuracy. Also, the visual quality of recovered images is very close to the watermarked image. It is tested on gray-scale as well as color medical images and shows consistent results with different medical image modalities. Therefore, it can be applied to 8-bit depth and 24-bit depth medical images of different modalities. Comparative analysis shows that the proposed scheme outperforms the existing schemes and may find potential solution to the telemedicine applications.

Although, most of the common image processing attacks have been applied extensively to authenticate the robustness of the proposed scheme but study of robustness under other attacks included in the standard watermark benchmarking tools can be seen as future work.