Ensuring integrity and security of medical image transmission in IoMT using highly imperceptible and robust watermarking approach

Abstract

With the technological revolution, the Internet of medical things (IoMT) has developed to be of immense benefit. In IoMT, medical images and patients’ data are widely transmitted through private/public network. An ideal transmission should not jeopardize the security, confidentiality, authenticity, authorization, or integrity of medical data/images. To ensure effective transmission and address the aforementioned issues, this paper proposes a blind region based medical image watermarking approach where a medical image is partitioned into region of interest (ROI) and region of non-interest (RONI). To ensure ROI intergrity, localized tamper detection and recovery bits (LTDRB) are generated. For precise diagnosis, patient’s electronic health record (EHR) and LTDRB are embedded in RoNI using hybrid DWT-SVD. No embedding is done in RoI to maintain its integrity and high visual quality. To ensure the security and confidentiality of EHR, a novel encryption scheme using Magic Square technique with low computational cost is proposed . Experimental results demonstrates that the proposed scheme provides high imperceptibility (Avg. PSNR>55 dB, SSIM $\approx$ 1 and BER $\approx 0$), robustness, security at low computational cost and high accuracy in tamper detection and recovery. A comparative study with some of the latest related research shows that the proposed scheme provides imperceptibility and robustness at par. However, the proposed scheme shows superior performance by providing higher EHR security at low computational cost and higher accuracy in ROI tamper detection and recovery, which other schemes have overlooked.

Introduction

Internet of Medical Things (IoMT) has revolutionized the healthcare systems by connecting medical equipment via Internet and thus enabling remote monitoring of patients by healthcare personnel¹. It provides a dependable route for the exchange of medical image and patients’ information. A general framework of data transmission in IoMT is shown in Figure 1.

Fig. 1.

Framework of image transmission in IoMT.

Despite the fact that a wide range of medical images and data can be shared in IoMT, there are still concerns about ensuring the authenticity, authorization, confidentiality, security, and integrity of the transmitted data. Medical images shared through a private/public network are vulnerable to intentional/unintentional manipulation resulting in an incorrect diagnosis by the physician². Unauthorized disclosure and tampering of patients’ Electronic Health Record (EHR) is also a serious concern. Thus, ensuring the authenticity, confidentiality, security, and integrity of medical images and EHR is crucial. During medical image transmission, confidentiality is maintained to ensure authorized access. The integrity of a medical image establishes whether the image is intact or tampered, whereas the authenticity determines whether the medical image is from the right source and belongs to the stated patient³.

Data protection approaches in IoMT

Cryptography, steganography, and digital watermarking are commonly used data protection approaches in IoMT. Cryptography transforms original data into an unreadable format, ensuring that only the sender and intended recipient can access the data. While encryption minimizes unauthorized access, illicit copying of encrypted data by an authorized user remains a problem^4,5. In steganography, the original data is secured by concealing it in a cover media in such a way that the presence of the hidden data remains undetectable. The cover media primarily acts as a concealing layer for the original data. Thus, ensuring good visual quality and integrity of the cover media is inconsequential. Also, steganography, does not ensure validity or resistance to change⁶. Digital Image Watermarking (DIW) embeds the watermarks in cover media in undetectable way. It prevents unauthorized copying and redistribution of images^7,8. It is frequently employed for image protection, content authentication, tamper detection, and copyright management. The relationship between various data protection approaches is shown in Fig. 2. Among these, digital watermarking has attracted the most interest since it makes image transmission simple, yet preserves essential characteristics such as authenticity, confidentiality, security, secrecy, and resilience. Cryptography and steganography are improved using digital watermarking⁹.

Fig. 2.

Relationship between cryptography, steganography and digital watermarking.

Medical Image Watermarking (MIW) has emerged as a potential method for resolving security issues in IoMT¹⁰. In MIW, watermarks are interleaved in the medical image to verify integrity and authenticity^11,12. From a diagnostic perspective, the medical image is divided into two regions: the Region of Interest (RoI), which is used for diagnosis, and the Region of Non-Interest (RoNI), which is not significant for diagnosis. RoI pixels are not altered (zero embedding) to ensure their integrity and good visual quality. Minor alteration in RoI might result in a false diagnosis therefore at the reciver end, tamper detection and recovery of RoI is critical¹³. The watermark is embedded in RoNI¹⁴. RoNI watermarks can be extracted and utilized to identify tampering, authorization, and information confidentiality^15–17. A good amount of research has been done in MIW using spatial, spectral and hybrid embedding techniques. However, achieving high security, accuracy in tamper detection and recovery is still a pressing problem. Also, lossless and real time transmission of medical image in IoMT with low computational cost is an important requirement. To address these challenges a region based hybrid MIW scheme is proposed here. In the proposed scheme, cover medical image is partitioned into RoI and RoNI. To ensure RoI integrity, Localized Tamper Detection and Recovery Bits (LTDRB) of RoI is generated. RoNI is utilized for the watermarks embedding process. EHR is encrypted using Magic Square algorithm before embedding in RoNI to ensure its confidentiality. Hybrid Discrete Wavelength Transforms (DWT)- Singular Value Decomposition (SVD) is applied on RoNI for embedding encrypted EHR and LTDRB to ensure high robustness and imperceptibility. The suggested approach assures medical data validity, authorization, integrity, and confidentiality. The proposed scheme can be used in practical applications like IoMT, Telemedicine and general healthcare systems for real time and lossless image transmission or storage. The rest part of this article will go through related work, the proposed work, experimental results followed by conclusion and future work as shown in Fig. 3.

Fig. 3.

Organization of the manuscript.

Related works

Over the past fifteen years, there has been a tremendous advancement in the research and development of watermarking algorithms using techniques from the spatial, frequency, hybrid domains and machine learning tehniques. Spatial domain techniques, such as the Least Significant Bit (LSB) algorithm, simplifies the process by embedding the watermark directly in the image. However, they are more prone to issues such as noise assaults, and compression¹⁸. To address these concerns, several researchers proposed frequency domain techniques for the watermark embedding process. The schemes presented in^19–21 utilized Direct Cosine Transform (DCT), Discrete Wavelet Transform (DWT) for watermark embedding in low frequency band (LL). Frequency domain transform technique increases the watermark’s resistance to the aforementioned issues, but affects the visual quality of the medical cover image. On the other hand, inserting the watermark at a higher frequency band (HH) makes it easier to hide the image, but at the cost of robustness²². The majority of researchers used hybrid transformation approaches to address this challenge, and embedding could be done in combination with low-high (LH, HL) bands. Owing to the enormous success of these hybrid domain approaches where spatial and frequency domain techniques are merged to provide algorithms like, DWT-SVD²³, RDWT-SVD²⁴, IWT-SVD²⁵, FD-DCT²⁶, DWT-Schur²⁷, DWT - DCT - SVD²⁸ and Competitive Algorithm (ICA)- Speed Up Robust Features (SURF) points- DCT²⁹. whereas some of the researchers used deep-learning learning approaches for watermarking.³⁰ utilizes Alexnet to extract visible features from a cover image. In addition, hybrid non-subsampled shearlet transform-SVD zero watermarking is used for embedding the watermark. Similarly,³¹ proposed a hybrid DWT-DCT zero watermarking approach, which uses the MobileNetV2 convolutional neural network to discover embedding places. Even while hybrid transform-based machine learning models are highly resilient, they take longer to run deep learning models. For real-time IoMT communication, the computing time for embedding and extraction is minimal. Therefore, watermarking with deep learning models for IoMT is time-consuming. However, hybrid methods beat individual approaches in terms of resilience without sacrificing robustness and imperceptibility, in less computational cost. They fall short, nevertheless, in protecting the data integrity during transmission. Tamper detection and recovery are examples of features of data integrity. A number of studies have previously addressed the tampering of medical images, but there was not enough information to pinpoint the affected region and locate the original data³². To address this issue, various approaches, such as those presented in^26,33–36, adopted localized tamper detection and recovery methods, in which the location of the tampered area is identified and appropriate recovery methods are used to recover the original medical data. It is noted that various techniques that provide improved accuracy still have room for improvement. Despite improved accuracy being reached by schemes like^37–39, the process of tamper recovery is complicated since the watermark is embedded across the whole image. As the diagnosis portion of a medical image is important, embedding in it might result in an incorrect diagnosis. To tackle this problem, some of the researchers suggested to segment the medical image under analysis into two regions i.e. RoI and RoNI. There is still space to suggest effective methods to ensure RoI integrity.

Furthermore, EPR security and confidentiality are essential for IoMT. Most existing watermarking schemes have paid little attention to watermark security. Watermarking schemes presented in^33,40,41 have less focused on watermark security. Chaotic map is applied for security of the watermark in⁴², but chaotic maps suffers from hyper tuning parameter issue. Despite the fact that the schemes presented in^36,42 have achieved high imperceptibility and security but has high computational cost. Since there is room for developing highly secure methods with minimal computational expense. To address the aforementioned concerns, a blind hybrid region-based medical watermarking scheme with symmetric encryption process at low computational cost is proposed to ensure high watermark security. To ensure high tamper detection and recovery accuracy, a novel localized tamper detection approach is suggested. Also, to maintain high ROI imperceptibility, watermark embedding is done at ROINI region. The motivation and contribution of the proposed scheme is elaborated in section .

Contributions of the proposed approach

Literature review highlights extensive research efforts for secure transmission of medical images over IoMT using MIW. While existing approaches have demonstrated promising results, challenges related to achieving higher security, data integrity, robustness, and imperceptibility requires further attention. To address these challenges, a region based blind hybrid medical image watermarking scheme is proposed. Following are the primary contributions of proposed approach:

• High accuracy in RoI tamper detection and recovery: The proposed approach ensures RoI integrity in medical images by performing tamper detection, localization and recovery for RoI pixels at the receiver side. This is achieved by dividing RoI in $3\times 3$ non-overlapping blocks and generating Localized Tamper Detection and Recovery Bits(LTDRB) for each block. Subsequently, LTDRB is embedded in RoNI at the sender side and extracted at the receiver end for tamper detection and recovery. The blockwise LTDRB improves tamper detection and recovery accuracy. The proposed approach is able to detect and localize RoI tampering with accuracy $\ge$ 97%. Further, tampered RoI bits are restored using extracted LTDRB. The visual quality of recovered ROI is excellent as indicated by high PSNR values for different medical image modalities.

• High security with low computational cost for EHR: The proposed approach provides high security for EHR using symmetric encryption based on pseudorandom key and shuffling of EHR pixels. Pseudorandom key is generated by using Polybius and Magic Square technique as well as EHR pixel shuffling based on pseudorandom key has low computational cost. Entropy of encrypted and decrypted EHR affirms that the proposed encryption approach is highly effective.

• High imperceptibility and robustness: The proposed approach ensures high imperceptibility of the medical image (PSNR $\ge 55.5dB$ and $SSIM\approx 1$). Zero embedding is done in RoI keep its pixels unaltered. Hybrid DWT-SVD transform is applied on RoNI for high imperceptibility. The proposed approach is highly robust against geometric, non-geometric and hybrid attacks owing to the implementation of hybrid DWT-SVD. The claim is supported by experimental results.

Proposed work

This work proposes a region based medical image watermarking for secure image transmission in IoMT applications that ensures the integrity, confidentiality, authorization, while being resilient and imperceptible. Workflow of the proposed scheme is shown in Fig. 4.

Fig. 4.

Workflow of the proposed scheme.

The suggested scheme segment the medical cover image ($\mathcal{C}$) of size $P\times Q$ into RoI and RoNI. LTDRB is generated to safeguard RoI integrity. EHR is encrypted using secret pseudorandom key generated through the Polybius and Magic square (PM) technique before embedding to ensure EHR confidentiality and security. For RoNI embedding, a hybrid DWT - SVD transform method is used to achieve higher visual quality and resilience. Combining the RoI and watermarked RoNI yields the final watermarked image $\mathcal{C}_{}^{\prime }$. Watermarks are extracted from the $\mathcal{C}_{}^{\prime }$ at the receiver’s end for validating authorization, authentication, and integrity. The proposed scheme is elaborated in following five modules:

• RoI and RoNI Segmentation

• Encryption and Decryption of Electronic Health Record

• Generation of Localized Tamper Detection and Recovery Bits

• Validating the Medical Image

• RoNI Embedding and Extraction

RoI and RoNI segmentation

The suggested approach segments the cover medical image $\mathcal{C}$ into two regions: RoI and RoNI. RoI is important in diagnosis and RoNI is insignificant in diagnosis decisions. Taking advantage of this aspect, the suggested approach uses the RoNI region for embedding and the RoI region for diagnostic assessments. The majority of researchers use manual and automatic methods for $\mathcal{C}$ segmentation. Manual segmentation is employed in this research since automated segmentation is computationally expensive, has lesser accuracy, and is dependent on visual modality. In manual segmentation approach, a radiologist or physician manually marks the RoI, i.e, an irregular shape ROI is manually indicated by the radiologist/physician in the $\mathcal{C}$ as shown in Fig. 5 (a). In the proposed scheme, the following steps are taken to define a rectangular RoI:

• Draw $P\times Q$ grid lines (P horizontally and Q vertically) to completely cover the image for a $P\times Q$ $\mathcal{C}$, with each pixel occupying one grid cell, as illustrated in Figure 5 (b).

• Find the highest RoI (manually labeled) pixel point (A1, A2, A3, A4) in all four directions. As illustrated in Figure 5 (b), A1, A2, A3, and A4 are the most prominent pixel positions along the north-west, north-east, south-east, and south-west grid lines, respectively.

• subsequently link the points (A1,A2), (A2,A3), (A3,A4), and (A4,A1) to produce a covering rectangle as shown in Figure 5 (c), the resulting rectangular area is the ROI of the $\mathcal{C}$, and the remaining area is the RONI.

Fig. 5.

RoI and RoNI segmentation.

Encryption and decryption of electronic health record

To achieve EHR security and confidentiality, the proposed scheme encrypts EHR prior to embedding using pseudorandom keys. The Polybius and Magic square (PM) approach is used to produce four sets of binary pseudorandom keys of $8\times 8$ size. The PM process is the integeation of Polybius, Magic square and musical notes approaches. Firstly, arrange the decimal integers into the magic square of $N\times N$ matrix, where N is the matrix’s order. It is made up of various integers in the range $(1,N^2)$ that have been randomly mixed so that the sum of the components in each row, column, and diagonal equals one. The term magic constant or magic sum is used to describe this constant sum. The Eq. (1) can be used to get the magic sum.

$$\begin{array}{c}\hfill Sum=N\times (N^2+1)/2\end{array}$$ 1

where Sum stands for the associated matrix’s magic sum and N is the order of the magic square matrix. A magic square of size $6\times 6$ with a magic total of Sum= 111 is taken into consideration in the suggested scheme. The numbers from 1 to 26 are given a character from A to Z in Polybius square, whereas the numerals from 27 to 36 are given a numerical value from 0 to 9. Each row and column of the magic square is given an alphabetical index from the musical notes set $\{A,B,C,E,F,G\}$ in order to preserve indexing. The initial index character is always altered, and a random character is selected from the available musical notes set to preserve randomization. Clockwise steps are done with the remaining indices. In the proposed scheme four pseudorandom keys were produced using a secret 16-character (128-bit) random alphanumeric string (S). Which acts as a secret key between sender and the receiver’s end. Algorithm 1 and Fig. 6 are illustrations of the suggested pseudorandom keys generation method. Also, Figure 7 shows the example of cipher key generation and the equivalent binary value of the generated keys.

Fig. 6.

Cipher key generation.

Fig. 7.

Example of cipher key generation and equivalent binary value.

Algorithm 1.

Pseudorandom keys generation from the randomly chosen secret key.

Pseudorandom keys generation

To generate four psudorandom keys, first, get the magical square numbers from the provided secret cipher text (128- bit), then convert them into the appropriate 8- bit binary representations. An $8\times 8$ matrix $(M_{\mathit{Original}})$ will now be created after two bits were arbitrarily chosen from two randomly chosen positions. Then, as shown in Fig. 8, rotation operations with angles of ${90}^o,{180}^o$, and ${270}^{\mathit{o}}$ are applied to the $M_{\mathit{Original}}$ matrix to get four pseudorandom keys. Further these four keys $(M_{\mathit{Original}}$, $M_{{90}^0}$, $M_{{270}^0}$, $M_{{360}^0}$) were utilized for generation of encrypted EHR. The process of EHR encryption as explained below:

Fig. 8.

Rotation operation performed on the original $8\times 8$ matrix.

EHR encryption

The EHR ($W_i$) of size $C\times D$ is partitioned into two sub-EHR’s ($W_{i}Odd$ and even $W_{i}Even$) which is further split into two partitions each as shown in Fig. 9. EHR partitioning is done by scanning EHR pixels in a raster line fashion. Thus, EHR is divided into odd ($W_{i}Odd$) and even ($W_{i}even$) position pixel parts using Eq. (2) and (3).

$$\begin{array}{cc}\hfill \sum _{i=1}^{C/2}\sum _{j=1}^D{W}_{i}Odd(i,j)=& \left\{\begin{array}{c}{W}_i(k,l),ifmod(l,2)\ne 0\\ ignored,Otherwise\end{array}\right)\hfill \end{array}$$ 2

$$\begin{array}{cc}\hfill \sum _{i=1}^C\sum _{j=1}^D{W}_{i}Even(i,j)=& \left\{\begin{array}{c}{W}_i(k,l),ifmod(l,2)==0\\ ignored,Otherwise\end{array}\right)\hfill \end{array}$$ 3

where, $W_i(k,l)$ is the original EHR. $W_i$Odd(i,j) and $W_i$Even (i,j) (k,l) are $W_i$Odd and $W_i$Even position pixels of the EHR respectively. Further, $W_i$Odd is partitioned into two equal halves vertically using Eq. (4) and Eq. (5).

$$\begin{array}{cc}\hfill \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}Odd1(k,l)=& \sum _{s=1}^{C/2}\sum _{t=1}^{D/2}{W}_{i}Odd(s,t)\hfill \end{array}$$ 4

$$\begin{array}{cc}\hfill \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}Odd2(k,l)=& \sum _{s=1}^{C/2}\sum _{t=(D/2)+1}^D{W}_{i}Odd(s,t)\hfill \end{array}$$ 5

Where C $\times$ D is size of EHR.Similarly, partitioning the $W_i$Even into two parts vertically using following relations Eq. (6) and Eq. (7).

$$\begin{array}{cc}\hfill \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}even1(k,l)=& \sum _{s=1}^{C/2}\sum _{t=1}^{D/2}{W}_{i}Even(s,t)\hfill \end{array}$$ 6

$$\begin{array}{cc}\hfill \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}even2(k,l)=& \sum _{s=1}^{C/2}\sum _{t=(D/2)+1}^D{W}_{i}Even(s,t)\hfill \end{array}$$ 7

Now, divides the each of EHR parts ( $W_i$Odd1, $W_i$Odd2, $W_i$Even1 and $W_i$Even2) into $8\times 8$ non-overlapping blocks. Then the positions of the odd pixels parts ($W_i$Odd1, $W_i$Odd2) were encrypted by performing an XOR operation on the $M_{\mathit{Original}}$, $M_{{90}^0}$ keys. Similarly, encrypt the even parts ($W_i$Even1, $W_i$Even2) using $M_{{180}^0}$, $M_{{270}^0}$ pseudorandom keys respectively. Finally concatenate all the EHR encrypted parts, to get encrypted EHR ($EHR_{}^{\prime }$) using the equations from Eq. (8) to Eq. (12).

$$\begin{array}{cc}\hfill \sum _{s=1}^{C/2}\sum _{t=1}^{D/2}{W}_{i}Odd_{}^{\prime }(s,t)=& \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}Odd1_{}^{\prime }(k,l)\hfill \end{array}$$ 8

$$\begin{array}{cc}\hfill \sum _{s=1}^{C/2}\sum _{t=(D/2)+1}^D{W}_{i}Odd_{}^{\prime }](s,t)=& \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}Odd2_{}^{\prime }(k,l)\hfill \end{array}$$ 9

$$\begin{array}{cc}\hfill \sum _{s=1}^{C/2}\sum _{t=1}^{D/2}{W}_{i}Even_{}^{\prime }(s,t)=& \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}Even1_{}^{\prime }](k,l)\hfill \end{array}$$ 10

$$\begin{array}{cc}\hfill \sum _{s=1}^{C/2}\sum _{t=(D/2)+1}^D{W}_{i}Even_{}^{\prime }(s,t)=& \sum _{k=1}^{C/2}\sum _{l=1}^{D/2}{W}_{i}Even2_{}^{\prime }(k,l)\hfill \end{array}$$ 11

$$\begin{array}{cc}\hfill \sum _{i=1}^C\sum _{j=1}^D{W}_i_{}^{\prime }_{}^{\prime }(i,j)=& \left\{\begin{array}{c}{W}_{i}Odd_{}^{\prime }](k,l),ifmod(j,2)\ne 0\\ W_{i}Even_{}^{\prime }(k,l),Otherwise\end{array}\right)\hfill \end{array}$$ 12

The process of EHR encryption in detailed with numerical example as shown in Fig. 10.

Fig. 9.

Watermark partitioning.

Fig. 10.

EHR encryption process.

EHR decryption

In this section process of decryption of encrypted EHR using secretly received seed value (sum) as elaborated. After extraction of encrypted EHR from watermarked medical image, the decryption process is follows. The process of watermark decryption as explained below: Firstly, partition the extracted EHR ($EHR_{}^{\prime }$ into four parts ($W_{i}Odd1_{}^{\prime }$, $W_{i}Odd2_{}^{\prime }$, $W_{i}Even1_{}^{\prime }$, $W_{i}Even2_{}^{\prime }$), the process of partitioning of $EHR_{}^{\prime }$ into four parts as explained in section . Further, generate four pseudorandom keys $M_{\mathit{original}}_{}^{\prime }$, $M_{{90}^0}_{}^{\prime }$, $M_{{270}^0}_{}^{\prime }$, $M_{{360}^0}$ using secretly received seed value (Sum) from the trusted third party, and followed by process explained in section and Algorithm 1. Then, partition the $W_{i}Odd1_{}^{\prime }$, $W_{i}Odd2_{}^{\prime }$, $W_{i}Even1_{}^{\prime }$, $W_{i}Even2_{}^{\prime }$ parts into $8\times 8$ non-overlapping blocks. Then apply Xor operation between the pixel positions of the odd pixels parts ($W_{i}Odd1_{}^{\prime }$, $W_{i}Odd2_{}^{\prime }$) and two pseudorandom keys $M_{\mathit{Original}}_{}^{\prime }$, and $M_{{90}^0}_{}^{\prime }$. Similarly, decrypt the even parts ($W_{i}Even1_{}^{\prime }$, $W_{i}Even2_{}^{\prime }$) using $M_{{180}^0}_{}^{\prime }$, $M_{{270}^0}_{}^{\prime }$ pseudorandom keys respectively. Finally concatenate all the EHR decryted parts, to get decrypted EHR using the same equations from Eq. (8) to Eq. (12).

Generation of localized tamper detection and recovery bits (LTDRB)

The medical image’s pixel values might be tampered while it is being sent via communication channels. When this happens, a fraudulent medical report is produced. Medical images may undergo intentional/unintentional tampering. The LTDRBs are generated for the medical cover image ROI part in order to address these issues and guarantee the integrity of the received medical image. LTDRB is used for RoI tamper detection and also utilized to recover the tampered RoI in the event that any interference is discovered. To generate LTDRB for RoI, firstly RoI is divided into $3\times 3$ non-overlapping blocks (B). Algorithm 2 and Fig. 11 show the process which has been implemented on each block (B) for the generation of LTDRB. Blockwise recovery bits (RB) can be utilized to retrieve the particular block. In the event that the recovery process is tampered with, the tampered pixel(s) may be easily targeted and retrieved.

Fig. 11.

LTDRB bits for RoI: (a) LTDRB Generation from ROI, (b) ROI recovery from LTDRB.

Algorithm 2.

LTDRB generation.

Validating the medical image

The receiver validates the medical image integrity for effective diagnosis. For this, RoI’ of the received medical image is used to create the recovery bits ($RB_{}^{\prime }$). For tamper detection, the produced $RB_{}^{\prime }$ and the extracted RB from RoNI are compared. If they are identical, then it can be concluded that the received image was not altered. If pixel values differs, then the localized tamper recovery process is performed using the extracted LTDRB from RoNI. Algorithm 3 and Fig. 11 depicts the steps involved in the recovery of tampered portion. Algorithm 3 is applied for all the $3\times 3$ non-overlapping blocks to detect localized tamper detection and recovery.

Algorithm 3.

Recovery of tampered RoI.

RoNI embedding and extraction

The suggested approach is made robust and imperceptible by embedding the watermarks in the RONI region of hybrid DWT- SVD transform domain. Embedding in frequency components ( LH, HL sub bands) gives a high level of protection against attacks because it includes salient image components that must be preserved for image integrity. By embedding a watermark in the desired sub-band, DWT offers remarkable imperceptibility⁴³. Because the reflectance component fluctuates fast, the embedded watermark is challenging for the Human Visual System (HVS) to detect. When the DWT coefficients are changed the subband of the image that corresponds to that coefficient alone gets modified and not the entire image. Despite the fact that spatial domain approaches are computationally efficient, the proposed scheme employs a transform domain embedding approach because spatial domain techniques are lacking robustness. Since the authentication of medical images depends on an effective embedding process. The watermark is incorporated into the DWT decomposed cover image frequency bands ( LH, HL) by modifying the pixel intensity values in the frequency domain. After embedding, all of the sub-bands are combined using the inverse process. DWT offers lossless recovery in this inverse process when compared to other transforms like IWT, RDWT, and so forth. Furthermore, DWT is shift invariant and resistant to geometrical attacks, which is a highly desirable property for effective image watermarking. DWT is therefore recommended in this paper. DWT is however susceptible to histogram, filtering, and noise attacks. The proposed plan combines DWT and SVD to address this. In general, SVD is stable and resistant to geometrical attacks like cropping, rotation, and resizing as well as common image manipulations like filtering, histogram equalisation, noise, and so forth^44,45. And also in SVD transformation the changes in the singular values will not affect the quality of the image. Therefore, even though hybrid DWT-SVD is computationally inefficient, this hybrid embedding is suggested in this scheme for the aforementioned reasons.

The key benefit of SVD transformation is that it offers exceptional single-value stability, even when small adjustments are made to the image. The robustness of the method is increased to a higher extent as a result of this SVD feature. In order to edit an image, SVD uses two unique sub-spaces: data and noise sub-spaces. These sub-spaces are often used for noise filtering, but they have also been utilized in watermarking applications.

RoNI embedding

By applying level-1 DWT on RoNI, is subdivided into LL, LH, HL and HH sub-bands. Here, the benefit of being robust to filtering attacks has been utilized. Further LH, HL subbands are selected for embedding. By applying SVD on LH, HL subdands, further each band is decomposed into three matrices. Such as $U_{\mathit{HL}}$, $S_{\mathit{HL}}$, $V_{\mathit{HL}}$, $U_{\mathit{LH}}$, $S_{\mathit{LH}}$, and $V_{\mathit{LH}}$ respectively. A high level of imperceptibility is provided by SVD and DWT together. Additionally, DWT helps separate high-frequency components from low-frequency components. The recovery bits LTDRB generated from RoI and encrypted EHR are embedded in the singular matrices ($S_{\mathit{LH}}$, $S_{\mathit{HL}}$). The singular values are only sensitive to slight modifications in the case of visual distortion. The singular value is furthermore invariant to transformations like translation, zoom, and mirroring. The LTDRB bits and EHR are embedded in $S_{\mathit{LH}}$ and $S_{\mathit{HL}}$, respectively, because singular matrices often carry less relevant image information. To provide great robustness and security, the selected $S_{\mathit{HL}}$ and $S_{\mathit{LH}}$ matrices are further divided into $4\times 4$ non-overlapping blocks. Figure 12 shows the procedure of watermark embedding in RoNI. A random scaling factor, also known as the watermark strengthening parameter, is used to embed the watermarks in diagonal positions as shown in Fig. 13. The final watermarked RoNI ($RoNI_{}^{\prime }$) is then produced by applying the inverse SVD and inverse DWT transformations.

Fig. 12.

Block diagram of the proposed embedding process.

Fig. 13.

Sample of $4\times 4$ SVD block for embedding.

Algorithm 4.

RoNI embedding process.

RoNI extraction

The watermarked RoNI ($RoNI_{}^{\prime }$) is subjected to 1-level DWT, which produces 4 sub-bands ($LL_{}^{\prime }$, $LH_{}^{\prime }$, $HL_{}^{\prime }$, and $HH_{}^{\prime }$) in order to extract the watermarks (LTDRB, $EHR_{}^{\prime }$) from the RoNI. It is necessary to process the sub-bands $LH_{}^{\prime }$ and $HL_{}^{\prime }$ of $RoNI_{}^{\prime }$ in order to extract the $EHR_{}^{\prime }$ and LTDRB bits embedded inside the HL and LH sub-bands of the RoNI. Further, SVD decomposition is then applied to the $LH_{}^{\prime }$ and $HL_{}^{\prime }$ sub-bands. The sub-bands are further divided into $U_{\mathit{HL}}_{}^{\prime }$, $S_{\mathit{HL}}_{}^{\prime }$, $V_{\mathit{HL}}_{}^{\prime }$, $U_{\mathit{LH}}_{}^{\prime }$, $S_{\mathit{LH}}_{}^{\prime }$ and $V_{\mathit{LH}}_{}^{\prime }$. A scaling factor is used to extract the two watermarks from the diagonal locations of the $4\times 4$ non-overlapping blocks in $S_{\mathit{HL}}^{\prime }$ and $S_{\mathit{LH}}^{\prime }$, respectively. Figure 14 shows the block diagram of watermark extraction process.

Fig. 14.

Block diagram of watermark extraction process.

The integrity of RoI is ensured by LTDRB bits, which make it easier to detect image manipulation. In the proposed scheme, the watermarks is retrieved without using the original cover image, thus the suggested method is blind watermarking.

Experimental results

Performance parameters such as imperceptibility, robustness, tamper detection and recovery, security test analysis, payload capacity and computational cost are used to assess the proposed scheme’s performance. Experiments were carried out on a PC with an Intel i7 CPU and 8GB RAM using MATLAB 2019. Test cover images were gathered for the tests from the OPENi⁴⁶, USC-SIPI⁴⁷, Kaggle⁴⁸, Online⁴⁹, and STARE⁵⁰ databases. For analysis , different grayscale and color medical images of various modalities, including CT, MRI, mammography, X-ray, PET, Skin, Retina, and Doppler images of $512\times 512$ and binary EHR of size $64\times 64$has been considered. More than 100 DICOM, JPEG, and TIFF image formats with horizontal and vertical resolution greater than 100 dpi were considered to analyze the performance of the suggested scheme. For performance evaluation, sample test images were taken as horizontal and vertical resolution of 100 dpi as shown in Fig. 15. The proposed scheme also works fine for images with higher resolution (upto 300 dpi). The performance analysis of the suggested strategy is further explained in the following subsections.

Fig. 15.

Test cover images: Gray-scale [1 - 5] and Color images [6 - 10] with EHR [11].

Imperceptibility test

Imperceptibility tests ensure that the quality of the medical images remains intact during transmission. It’s crucial to maintain the medical image’s visual qualities during transmission. Compromise with the medical image’s visual quality might result in a misdiagnosis. As a result, imperceptibility is among the top considerations when assessing the effectiveness of the proposed method. High imperceptibility factor is necessary to guarantee effective watermarking of medical images. Two performance metrics, Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), are used to quantify this.The PSNR is calculated as the square of the maximum pixel intensity values to the average error between the original image and the watermarked image in terms of magnitude (MSE). Higher PSNR value suggests that the scheme is more imperceptible. An SSIM statistical metric aids in assessing the decline in image quality depending on a variety of factors, including transmission loss and data compression. Its value ranges between 0 and 1. A high SSIM score indicates that the transmitted and original images are very similar. The mathematical relation for PSNR and SSIM is as provided in⁵¹.

For the grayscale and colour medical images, as shown in Fig. 15 the imperceptibility and robustness of this approach are evaluated in terms of PSNR and SSIM, NC, and BER, with range scaling factor values ($\beta$ = 0.01 and $\beta$ = 0.1). According to Table 1, for both $\beta$ = 0.01 and $\beta$ = 0.1, the PSNR value is higher than 55 dB for test cover images. Although infinite PSNR is ideal value, a PSNR value above 47 dB is regarded as excellent visual quality of the transmitted image^52,53. Also, for all grayscale images SSIM value is 1, and for color images SSIM $\approx$ 1 for different scaling factors ($\beta$ = 0.01 and $\beta$ = 0.1). Grayscale images have an average PSNR of 57.28 dB, while color images have a PSNR of 60.32 dB. This observation indicates that the proposed scheme achieves high imperceptibility for grayscale and color images of various modalities. The proposed MIW scheme guarantees that the imperceptibility criteria are effectively met with varying scaling factor ($\beta$ = 0.01 and $\beta$ = 0.1).

Table 1.

PSNR, SSIM, NC, and BER for gray-scale and color images ($\beta =0.01$, $\beta =0.1$).

	$\beta =0.01$				$\beta =0.1$
	PSNR	SSIM	NC	BER	PSNR	SSIM	NC	BER
Grayscale Images
X-ray	56.31	1	0.9999	0.0002	55.72	1	0.9999	0.0002
CT	57.31	1	0.9998	0.0004	56.92	1	0.9998	0.0004
Ultrasound	55.65	1	0.9999	0.0002	54.37	1	0.9999	0.0002
Mammograph	58.62	1	0.9997	0.0008	57.28	1	0.9997	0.0008
MRI	58.49	1	1	0	57.29	1	1	0
Color Images
Doppler	61.04	1	0.9999	0.0002	60.29	1	0.9999	0.0002
Retina	57.62	0.9998	0.9999	0	55.92	0.9998	0.9999	0
Skin	62.72	1	0.9997	0.0004	61.82	1	0.9997	0.0004
Pet	60.62	0.9999	0.9998	0.0002	59.27	0.9999	0.9998	0.0002
MRI	59.62	0.9999	0.9998	0.0004	58.18	0.9999	0.9998	0.0004

Further, the evaluation of imperceptibility performance is expanded to include 40 images of MRI, CT scan, ultrasound, Doppler, X-ray, Mammograph, skin, retina, and, PET brain images as shown in Figure. 16, and the findings are reported in Table 2. PSNR and SSIM values are greater than 54 dB for all image modalities as presented in Table 2. Tables 1 and 2 show that the proposed approach ensures high visual quality for medical images of various modalities with varying scaling factor.

Fig. 16.

Medical cover images (X-ray, CT scan, ultrasound, MRI, Doppler, retina, skin, and, PET brain images) taken from OPENi⁴⁶, USC-SIPI⁴⁷, Kaggle^48,49, and STARE⁵⁰ datasets.

Table 2.

PSNR, SSIM, NC, and, BER (Under zero attack) for 40 test cover images under zero attack taken from OPENi⁴⁶, USC-SIPI⁴⁷, Kaggle^48,49, and STARE⁵⁰ datasets with $\beta =0.01$.

Images		PSNR	SSIM	NC	BER
X-ray	Img1	56.31	1	0.9999	0.0002
	Img2	54.28	1	0.9999	0.0002
	Img3	53.28	1	1	0
	Img4	55.84	1	1	0
	Img5	54.75	1	0.9998	0.0002
CT	Img1	57.31	1	0.9998	0.0004
	Img2	56.19	1	0.9999	0.0002
	Img3	55.92	1	0.9998	0.0004
	Img4	57.18	1	0.9999	0.0002
	Img5	54.84	0.9999	0.9999	0.0002
Ultrasound	Img1	55.65	1	0.9999	0.0002
	Img2	56.81	1	1	0
	Img3	53.51	1	0.9999	0.0002
	Img4	54.38	1	0.9998	0.0004
	Img5	55.16	0.9999	0.9999	0.0001
MRI	Img1	58.49	1	1	0
	Img2	57.19	1	1	0
	Img3	55.83	1	0.9999	0.0002
	Img4	56.64	1	0.9999	0.0002
	Img5	57.40	1	1	0
Doppler	Img1	61.04	1	0.9999	0.0002
	Img2	59.71	1	0.9999	0.0002
	Img3	58.19	1	1	0
	Img4	57.38	1	0.9998	0.0004
	Img5	60.42	1	0.9999	0.0002
Retina	Img1	57.62	0.9998	0.9999	0
	Img2	55.27	0.9999	0.9999	0
	Img3	56.93	0.9999	0.9999	0.0002
	Img4	54.83	0.9999	0.9999	0
	Img5	56.74	0.9999	0.9998	0.0004
Skin	Img1	61.82	1	0.9997	0.0004
	Img2	59.25	1	0.9999	0.0002
	Img3	58.74	1	1	0
	Img4	57.48	1	0.9998	0.0004
	Img5	60.25	1	0.9999	0.0002
PET	Img1	59.27	0.9999	0.9998	0.0002
	Img2	57.24	0.9999	0.9999	0
	Img3	56.15	0.9998	0.9998	0.0002
	Img4	58.38	0.9999	0.9999	0.0004
	Img5	59.41	0.9999	0.9999	0.0002

Robustness test

One of the important aspect used to assess the performance of the proposed method is robustness. MIW embeds data such as EHR and tamper recovery information in the form of a watermark into medical images. The medical image and associated watermarked data are vulnerable to a variety of attacks during transmission from the transmitter to the recipient. To assure the integrity/resistance of medical image data, it is critical to test the robustness of the proposed MIW system for various types of images under different attacks. Medical images is performed under various attacks and the results were measured in the form of Normalized Correlation (NC) and Bit Error Rate (BER). NC is a metric that compares the similarity of embedded and extracted watermarks. The BER is used to calculate the likelihood of bits being received wrongly owing to noise. It displays the number of bits that separate the embedded and extracted watermarks. The mathematical equations for both of these measures are presented in⁵¹:

Figure 17 shows an extracted watermark from watermarked medical images under zero attacks. As previously indicated, the subjective evaluation of the extracted watermark in Fig. 17 shows that the embedded and extracted watermarks are identical. Table 2 shows the robustness of the proposed approach for 40 images of different modalities under zero attack when $\beta =0.01$. Table 2 shows that for EPR, $NC\simeq 1$and $BER\simeq 0$, i.e. near to the ideal values, are obtained for all images. This observation demonstrates that the proposed approach is robust for many image modalities when $\beta =0.01$. The proposed scheme’s robustness performance for several scaling factors ($\beta$ = 0.01, and $\beta$ = 0.1) under zero attack is further investigated, and NC and BER values are shown in Table 1. According to Table 1, there is no significant difference in NC and BER values for the retrieved EPR with varied scaling factors. Watermarks are successfully extracted (under zero attack) with large scaling factor values. Furthermore, the suggested scheme’s robustness is tested on various medical images using noising, filtering, compression, and geometrical attacks. For the convenience of representation, robustness performance under various attacks for X-Ray, Ultrasound, MRI and Doppler images are only presented. These 4 four images almost reflects the features of other remaining images. Table 3 displays the NC and BER values of EPR for X-ray, Ultrasound, MRI and Doppler images under various attacks. Table 3 solely shows the results of X-ray, Ultrasound, MRI and Doppler images for convenience.

Fig. 17.

Cover image with corresponding watermarked images and extracted watermark under zero attack ($\beta =0.01$): Gray-scale watermarked images (1– 5), Color watermarked images (6–10).

Table 3.

NC and BER values under various attacks with $\beta =0.01$ for X-ray, Ultrasound, MRI and doppler images.

Attacks	X-Ray		Ultrasound		MRI		Doppler
	NC	BER	NC	BER	NC	BER	NC	BER
Gaussian Filter ($3\times 3$)	0.9999	0.0002	0.9997	0.0004	0.9997	0.0008	0.9996	0.0004
Gaussian Filter ($5\times 5$)	0.9998	0.0014	0.9991	0.0014	0.9992	0.0016	0.9994	0.0013
Median Filter ($3\times 3$)	0.8853	0.2962	0.8392	0.3052	0.8782	0.2204	0.8639	0.2051
Median Filter ($5\times 5$)	0.8803	0.3014	0.8136	0.3247	0.8514	0.2015	0.8382	0.2381
Average Filter ($3\times 3$)	0.9991	0.0010	0.9990	0.0014	0.9989	0.0016	0.9994	0.0008
Average Filter ($5\times 5$)	0.9988	0.0056	0.9987	0.0086	0.9894	0.0095	0.9985	0.0068
Weiner Filter ($3\times 3$)	0.9526	0.1127	0.9626	0.1062	0.9382	0.1862	0.9604	0.1726
Weiner Filter ($5\times 5$)	0.9371	0.1206	0.9361	0.1641	0.91062	0.2062	0.9271	0.1941
Butter worth Filter ($3\times 3$)	0.9993	0.0046	0.9995	0.0012	0.9989	0.0101	0.9992	0.0024
Butter worth Filter ($5\times 5$)	0.9990	0.0093	0.9992	0.0021	0.9983	0.0156	0.9989	0.0063
Sharpening	0.9721	0.1894	0.9594	0.2062	0.9802	0.0963	0.9583	0.2004
Histogram Equalization	0.8864	0.2542	0.8752	0.2472	0.8808	0.2362	0.8704	0.1973
JPEG Compression (50%)	0.9947	0.0873	0.9925	0.0904	0.9952	0.0726	0.9873	0.0683
PEG Compression (40%)	0.9902	0.0782	0.9901	0.0891	0.9902	0.0703	0.9816	0.0652
Salt & Pepper Noise (0.002)	0.8873	0.2013	0.8952	0.1995	0.8852	0.2052	0.8832	0.2042
Speckle Noise	0.8253	0.2572	0.8028	0.2162	0.8052	0.2962	0.8173	0.2372
Poission Noise	0.7852	0.3862	0.7968	0.3462	0.8162	0.3072	0.8031	0.2861
Gaussian Noise	0.8528	0.2962	0.8282	0.2862	0.8472	0.2971	0.8462	0.2855
Rotation (10)	0.8604	0.1627	0.8517	0.1371	0.8427	0.1079	0.8183	0.1962
Rotate (45)	0.8316	0.1418	0.8275	0.1162	0.8101	0.1826	0.8283	0.1782
Translate (24.3, 10.1)	0.8826	0.1082	0.8721	0.1008	0.8927	0.1821	0.8824	0.1093
Resize (256)	0.9021	0.0927	0.9182	0.0831	0.9318	0.0284	0.9492	0.0147
Resize (320)	0.8962	0.09026	0.8951	0.08074	0.9073	0.03951	0.9294	0.0273
Cropping (10%)	0.8195	0.2382	0.8371	0.2278	0.8284	0.2086	0.8318	0.2082
Shear (x-shear)	0.8406	0.2428	0.8159	0.2027	0.8318	0.2382	0.8469	0.2085
Gamma Correction (0.25)	0.9027	0.1096	0.9372	0.1721	0.9172	0.1273	0.9283	0.1137
Gamma Correction (0.3)	0.8921	0.1084	0.9228	0.1271	0.9027	0.1025	0.9093	0.0972

The results shown in Table 3, the proposed scheme exhibits excellent robustness and low fluctuations in the related NC, BER for all of the stated images when applied using a Gaussian filter with varying filter sizes ranging from $3\times 3$ to $5\times 5$. Furthermore, for larger filter sizes of $5\times$5, the proposed scheme produces a higher NC for all images. Also, for the Average filter, the proposed scheme kept NC and BER close to ideal values for a variety of filter sizes while exhibiting very low distortions even at larger filter sizes. The proposed scheme is highly resistant to the median and Weiner filters.

The proposed scheme is also resistant to the Butterworth and sharpening filters, and the variation in NC and BER for different filter sizes is minor, with filter sizes ranging from $3\times 3$ to $5\times 5$, and NC ranging from 0.9993 to 0.9990 for a sampled X-ray image. Furthermore, the proposed scheme displays NC, BER greater than the threshold value for histogram Equalisation, and various noise attacks such as Salt & pepper noise, Speckle noise, Poisson noise, and Gaussian noise. This demonstrates that the provided scheme is resistant to noise attacks. Table 3 is also included. NC, BER for the performance of geometrical attacks such as rotation, translation, resize, cropping, shear, and gamma correction. The proposed scheme shows NC$>$ 0.7 (threshold) and BER $<$ 0.4 (threshold) for all of the listed geometrical attacks. This indicates that the proposed scheme is resistant to geometrical attacks. From this discussion it can be conclude that the proposed scheme is highly robust to various common attacks.

Security key test

Confidentiality of EHR is crucial for healthcare applications. To ensure high confidentiality and security, EHR is encrypted using a pseudo-random key. Performance of the proposed encryption scheme has been evaluated by using metrics suc as Correlation Coefficient (CC), entropy⁵⁴, Number of Changing Pixel Rate (NPCR) and Unified Averaged Changed Intensity (UACI)⁵⁵. Table 4 shows the CC values (horizontal, vertical, and diagonal directions) for both the encrypted and decrypted images for 10 test binary images. The correlation between the original watermark image and the encrypted watermark image was less than 0.4, while for the decrypted watermark image, the correlation was 1 for all test images. Table 4 shows information entropy of original and encrypted images. The entropy of all encrypted images stated in Table 4 having entropy higher than the original image. This indicates that the proposed scheme generates strong cipher images. Further, the security performance is validated using NPCR and UACI metrics between to encrypted images with one bit difference in the corresponding seed value in pseudorandom key generation process. Ideally NPCR and UACI confidence values for binary image is 50%⁵⁵. From Table 4, it can be observed that, for all binary test images NPCR and UACI approaching to ideal value of 50%. This confirms that the proposed scheme provides high and effective watermark security.

Table 5.

Entropy, NPCR, UACI for binary images.

Image	Entropy before encryption	Entropy after encryption	NPCR %	UACI %
EHR	0.4275	0.9182	47.23	47.31
EHR2	0.4571	0.9261	48.16	46.36
X - ray	0.7518	0.9825	45.89	44.17
Ultrasound	0.8041	0.9926	47.25	45.27
Pet	0.6717	0.9824	48.37	47.61
Retina	0.7421	0.9926	47.25	44.75
Mammograph	0.7281	0.9782	45.27	46.36
Skin	0.7281	0.9782	45.25	46.83
MRI	0.5823	0.9527	47.36	48.17
CT	0.5921	0.9972	48.62	46.26

Table 4.

CC of encrypted and decrypted EPR and binary image (H- Horizontal, V-Vertical, D-Diagonal).

Image	CC between Original and encrypted Image			CC between Encrypted and decrypted image
	H	D	V	H	D	V
EHR1	0.1036	0.2618	0.1061	1	1	1
EHR2	0.0984	0.1723	0.0086	1	1	1
X-ray	0.2015	0.2164	0.2048	1	1	1
Ultrasound	0.0926	0.1035	0.0863	1	1	1
Pet	0.0871	0.1017	0.0856	1	1	1
Retina	0.0618	0.1082	0.0961	1	1	1
Mammograph	0.0528	0.1102	0.08271	1	1	1
Skin	0.0481	0.0252	0.0381	1	1	1
MRI	0.0722	0.0097	0.0627	1	1	1
CT	0.0962	0.1026	0.0852	1	1	1

Tamper detection and recovery

For an accurate diagnosis, reliable medical images, especially in the RoI, are essential. In this section, the effectiveness of a proposed tamper detection and recovery scheme has been studied. The RoI is used as the tampered area in copy-paste tampering operation is done on 10% of watermarked image. Original watermarked images, tampered images, detected tampered regions, and restored images for different image modalities are shown in Fig. 18. For all images the tampered region detected, localized, and retrieved. Table 6 shows the scheme’s efficacy in terms of accuracy rate and PSNR. The proposed approach has achieved an accuracy rate greater than 97% in tamper detection, and the visual clarity of the retrieved images is excellent. As a result, the suggested technique is able to detect and recover tampering in various image modalities.

Fig. 18.

Watermarked images (1-4), Copy-paste tampered images (a-d), Tamper detected area marked with white pixels (i - iv), Recovered images (I - IV).

Table 6.

Recovered images tamper detection accuracy rate (%) and PSNR.

Image	Accuracy	PSNR
X - ray	98.27	54.04
Ultrasound	98.68	53.82
Pet	97.92	58.72
Skin	98.85	60.34
Retina	97.85	55.82
Mammograph	98.05	55.02
Skin	98.28	56.38
MRI	99.01	56.93
CT	97.26	53.58

Payload

This section analyzes the payload capacity associated with the proposed method. The number of watermark bits that can be embedded in cover image is termed as payload of watermarking scheme. The payload of the proposed scheme is calculated using following relation:

$$\begin{array}{cc}& Payload=\frac{Totalnumberofwateramrkbits}{Totalnumberofcoverimagepixels}bpp\hfill \end{array}$$ 13

$$\begin{array}{cc}& \therefore Payloadoftheproposedscheme=\frac{A\times B}{M\times N}bpp\hfill \end{array}$$ 14

For example:

$$\begin{array}{cc}& Totalnumberofcoverimagepixels=512\times 512=2,62,144pixels\hfill \\ \hfill & Totalnumberofwatermarkbits=64\times 64=4096bits\hfill \\ \hfill & Embeddingcapacity(Payload)=\frac{4096}{2,62,144}=0.015625bpp\hfill \end{array}$$

Proposed scheme having payload of 0.015625 bpp.

Computational cost

This section presents an analysis of the computational cost associated with the proposed scheme. Table 7 displays the computational cost associated with the elementary expensive steps for a medical cover image of dimensions $M\times N$ and watermark’s dimensions $A\times B$. The computational cost of the proposed scheme is determined solely based on the expensive steps. Table 7 states that the suggested MIW system has a computational cost of $O(M{N}^2)$. The suggested scheme’s computing cost is evaluated in three scenarios. In scenario 1, estimate the cover image size as $512\times 512$ and the watermark as $64\times 64$. In scenario 2, consider high resolution images with cover image size of $1024\times 1024$ and watermark as $128\times 128$. In scenario 3, consider high resolution images with cover image size of $1024\times 1024$ and watermark as $64\times 64$ The encryption and decryption have a low computational cost of O(AB), which is solely dependent on the size of the watermark. Encryption and decryption costs rise with increasing watermark size, as shown in Table 8 for two watermark sizes of $64\times 64$ and $128\times 128$, respectively. In accordance with Table 8, the encryption and decryption timings for $64\times 64$ and $128\times 128$ watermark sizes are 0.06 and 0.13 seconds (approximately), respectively. This suggests that the cost of the encryption and decryption processes rises with watermark size. In addition, Table 9 displays the extraction times for color and grayscale images for the three scenarios mentioned above. While the run-time for watermark embedding in grayscale and color images is less than 0.11 seconds and 0.39 seconds, respectively, the extraction time for grayscale and color images for scenarios 1 and 3 is approximately 0.08 and 0.12 seconds. This suggests that there is no effect on computing costs when the size of the cover image (high resolution images) is raised while keeping the watermark size fixed. The use of high-resolution images with larger watermark sizes also results in a longer embedding and extraction process, as demonstrated by scenario 2. Therefore, it can be seen that a high-resolution cover image has no influence on computation costs as long as the watermark size is increased. Similarly scenario 2 demonstrates that utilizing high-resolution images with larger watermark sizes causes the embedding and extraction process to take longer. Therefore, it can be seen that a high-resolution cover image has no influence on computation costs as long as the watermark size is increased. The computational cost of the suggested model rises in proportion to the watermark size. The study leads to the conclusion that the suggested scheme has an affordable computational cost.

Table 7.

Computational cost of the major expensive steps in the proposed scheme.

Operation	Computational cost
EPR encryption/ decryption	O(AB)
1-level 2D DWT	O(2MN)
1-level 2D inverse DWT	O(2MN)
SVD decomposition	$O(2M{N}^2+2n^3)$
SVD re-composition	O(2min[M, N], MN)
BR, NBR partitioning	O(MN)
Embedding process/Extraction process	O(MN)
LTDRB generation	O(MN)
Overall complexity	$O(M{N}^2)$

Table 8.

Encryption and decryption time in seconds. (EN-Encryption, DC- Decryption).

Binary image	Scenario 1: Watermark Size ($64\times 64$)		Scenario 2: Watermark Size ($128\times 128$)
	EN time (Secs)	DC time (Secs)	EN time (Secs)	DC time (Secs)
EHR	0.086417	0.078452	0.164897	0.149524
X – ray	0.068532	0.055327	0.122910	0.108513
CT	0.085286	0.075291	0.165681	0.145318
Mammograph	0.075271	0.064284	0.150815	0.139652
Ultrasound	0.052874	0.050731	0.107532	0.098632

Table 9.

Embedding (ET) and extraction time (EXT) in seconds.

	Scenario 1 ($512\times 512$, $64\times 64$)		Scenario 2 ($1024\times 1024$,$128\times 128$)		Scenario 3 ($1024\times 1024$,$64\times 64$)
	ET	EXT	ET	EXT	ET	EXT
Grayscale image
X – ray	0.1079	0.0728	0.1973	0.1381	0.1083	0.0618
CT	0.1286	0.0864	0.2091	0.1681	0.1186	0.0792
Ultrasound	0.1107	0.0853	0.2091	0.1781	0.1117	0.0971
Mammograph	0.1009	0.0753	0.1981	0.1371	0.1017	0.0917
MRI	0.1377	0.0954	0.2518	0.1781	0.1491	0.0982
Color image
Doppler	0.3976	0.1055	0.6824	0.2151	0.4028	0.1098
Retina	0.3874	0.1175	0.6961	0.2218	0.3892	0.1098
Skin	0.4074	0.1376	0.7912	0.2718	0.3987	0.1402
PET	0.3996	0.1475	0.8012	0.2718	0.4001	0.1398
MRI	0.3863	0.1387	0.7123	0.1781	0.3893	0.1397

Comparative analysis

A comparative evaluation of the suggested approaches has been done with the most recent and widespread MIW schemes put forward by Alshanbari³³, Singh et al.⁵¹, Zermi et al.⁵⁶, Pallaw et al.⁵⁷ and Chowdary et al.⁵⁸ in order to validate the relevance of the suggested approach within the current state of knowledge.

An overview of the schemes in comparison is presented in In Table 10. The method put forth by Alshanbari³³ uses LZW compression for data compression and DWT-SVD transformation for embedding watermarks. The cover image is $480\times 680$ pixels in size, and the watermark is $60\times 60$ pixels. The SHA-256 cryptographic hashing technique ensures security. The approach, however, lacks tamper detection features. The second method Singh et al.⁵¹ suggests makes use of a larger cover image of $512\times 512$ and a watermark of $64\times 64$ pixels. The scheme includes tamper detection methods in addition to data compression using DAC and Huffman encoding. However, the scheme’s security is not specifically specified, and it is not obvious if it makes use of hashing or encryption technologies. The third method put out by Zermi et al.⁵⁶ also use DWT-SVD transformation but leaves the watermark size unspecified. The lack of compression or encryption techniques in the scheme makes it susceptible to attacks. Furthermore, the method lacks any mechanisms to detect tampering, which makes it easier to alter the watermark. The proposed scheme also employs DWT-SVD transformation with $512\times 512$ for the cover image and $64\times 64$ for the watermark. It features tamper detection methods and a MP strategy for data encryption. It is more safe than the previous technique because it specifically states employing a hashing algorithm (SHA-256) for security. Similarly, Pallaw et al.⁵⁷ proposed a SLT-RSVD scheme for medical image transmission. This scheme uses firefly optimization technique for scaling factor optimization. This scheme lacks in security and integrity checking. Furthermore, Chowdary et al.⁵⁸ suggested an hybrid DWT - Hessenberg Decomposition (HMD)- SVD for medical image transmission. This scheme uses Arnold map for watermark security and lacks in integrity checking. The computational cost of all the scheme shown in Table 10 in seconds including embedding, encryption operations. The computations cost of the suggested scheme is low when compared to state-of-the-art schemes that can be seen in Table 10. Also the proposed scheme, which incorporates data encryption to protect the watermark, looks to be the most secure and reliable overall. Further, the performance of the proposed scheme is analysed with watermarking features like imperceptibility, robustness and embedding capacity.

Table 10.

Comparison of different watermarking schemes. (TD- Tamper detection).

Scheme	Method	Cover image size	Watermark size	Watermark Security	TD	Payload
Alshanbari33 (2021)	DWT-SVD	$480\times 680$	$60\times 60$	LZW	SHA	0.01102
Singh et al.51 (2022)	DWT-SVD	$512\times 512$	$64\times 64$	Huffman & DAC	Yes	0.01562
Zermi et al.56(2022)	DWT-SVD	$256\times 256$	$64\times 64$	No	No	0.06251
Pallaw et al.57 (2023)	SLT- RSVD	$700\times 600$	$64\times 64$	XOR	No	0.00975
Chowdary et al.58 (2025)	DWT-HMD-SVD	$512\times 512$	$64\times 64$	Arnold map	No	0.01562
Proposed	DWT-SVD	$512\times 512$	$64\times 64$	MP	Yes	0.01562

PSNR and NC, two performance indicators, are used in the Table 11 to compare various watermarking methods. While NC gauges how comparable the original and watermarked images are, PSNR gauges the caliber of the watermarked image. On X-ray and CT images, Alshanbari³³ embeds watermarks via the DWT-SVD transformation. For X-ray and CT images, the technique obtains a PSNR of 47.86 and 48.00, respectively. For X-ray and CT scans, the NC values are 0.9216 and 0.8847, respectively. In addition, Zermi et al.⁵⁶ embed watermarks on CT and X-ray images using the DWT-SVD transformation. In comparison to the first scheme, the second method obtains a higher PSNR of 57.04 for X-ray pictures and 55.85 for CT images. Additionally, it successfully generates a perfect NC value of 1 for both CT and X-ray pictures. Singh et al.⁵¹ insert watermarks using the DWT-SVD transformation on CT and X-ray images. In contrast to the preceding two techniques, it produces a PSNR of 41.96 and 43.23 for X-ray and CT images, respectively. For both X-ray and CT pictures, the NC values are ideal (1). Also Pallaw et al.⁵⁷ proposed a slant transform based hybrid scheme, the PSNR and NC for the X - ray , CT images are 58.61 dB, 58.04 dB, 0.9998 and 0.9987 respectively. The proposed method also embeds a watermark on CT and X-ray images using the DWT-SVD transformation. For X-ray and CT images, it obtains a PSNR of 56.31 and 57.31, respectively. For both X-ray and CT images , the NC values are ideal (1). Overall it can be conclude that the⁵⁶ and proposed strategies outperform the other schemes in terms of higher NC values and greater PSNR values under zero attacks. This shows that the watermarked images created by these techniques are of higher quality and more resemblance of the original images. It is crucial to remember that each scheme’s effectiveness may vary depending on the particular application and specifications. Table 12 provides a comparison of the robustness of diverse watermarking techniques when subjected to different types of attacks. The comparative attacks employed in this study include sharpening, Gaussian filtering, JPEG compression, median filtering, and salt and pepper noise. The evaluation of the resilience of each approach is conducted through the computation of the Normalized Correlation (NC) coefficients between the unaltered and tampered watermarked images. Alshanbari’s³³ initial proposal exhibits a higher degree of resilience towards sharpening, median filtering, and salt and pepper noise assaults in comparison to alternative schemes. Nevertheless, it exhibits lower resilience when subjected to JPEG compression. Zermi et al.⁵⁶ have proposed a second scheme that exhibits a high degree of robustness against attacks involving Gaussian filtering and median filtering. Nevertheless, the system is susceptible to sharpening and salt and pepper noise assaults. Schemes proposed in^51,57 exhibits a high degree of robustness against attacks from JPEG compression and salt and pepper noise. Nevertheless, it is susceptible to median filtering and sharpening assaults.

Table 11.

Performance comparison of proposed scheme with respect to^{33,51,56–58}.

Image	X – ray		CT
Scheme	PSNR	NC	PSNR	NC
33	47.86	0.9216	48.00	0.8847
51	41.96	1	43.23	1
56	57.04	1	55.85	1
57	58.61	0.9988	58.04	0.9987
58	52.67	1	49.86	1
Proposed	56.31	1	57.31	1

Table 12.

Robustness comparison (NC) of X - ray for^33,51,56, and the proposed scheme under different attacks.

Attack	Schemes
	33	51	56	57	Proposed
Sharpening	0.8819	0.9625	0.8119	0.9572	0.9721
Gaussian filter	0.9191	0.9999	0.9989	0.9991	0.9999
JPEG compression (QF = 70)	0.8118	0.9932	0.9631	0.9986	0.9982
Median Filtering ($3\times 3$)	0.9103	0.8743	0. 7952	0.8154	0.8853
Salt and pepper	0.8168	0.8042	0.9461	0.9879	0.8873

The proposed scheme exhibits a high degree of robustness against all attacks listed in the table, as evidenced by its attainment of the highest NC values across all attacks.

Further, computational cost of the proposed scheme is analyzed with comparative scheme by considering embedding and extraction time in seconds. Average embedding and extraction time for 100 medical images taken by the schemes in comparison is shown in Fig. 19. It can be observed from Fig. 19 that the proposed scheme shows less computational time for embedding and extraction time than the current state-of-the-art schemes. This can be attributed to low encryption and decryption cost.

Fig. 19.

Average embedding and extraction time for 100 medical images in seconds.

The embedding capacity of the proposed scheme and comparative schemes listed in Table 10. The embedding capacity of the proposed scheme is higher than that of the schemes proposed in^33,57 and equal capacity of the scheme proposed in⁵¹ with less computational cost. The scheme proposed in⁵⁶ has higher embedding capacity than the suggested scheme with the lack of security and integrity features. Therefore, this discussion suggests that the proposed methodology exhibits greater resilience against a range of attacks, with high security and high embedding capacity than the comparative methodologies. Therefore it can be claim that the proposed scheme is performed superior performance than the existing schemes.

Conclusion and future work

In order facilitate effective medical image transmission via IoMT and to guarantee high imperceptibility, integrity, authenticity, secrecy, and robustness, a secured region-based MIW approach is suggested in this study. The proposed approach provides high imperceptibility (PSNR>54db and SSIM>0.9997) and robustness (NC>0.9997) for all imagine modalities under zero attack. The performance of the proposed technique is evaluated under various image processing assaults, including filtering, noise, and geometrical assaults. The findings of the experiment indicate a higher robustness performance by the proposed scheme. Future studies can explore automated RoI and RoNI segmentation schemes with low computational cost and independent of the image modality. In addition, multiple ROI detection and handling can be seen as a future research area.

Author contributions

P.S., K.J.D., T.K.N., C.S.P. and H.K.T. designed the method, conducted experiments. P.S., K.J.D., T.K.N. wrote the main manuscript text. S.A.H., and S.M. reviewed and edited the manuscript. All authors reviewed the manuscript. The authors read and approved the final manuscript.

Data availibility

The datasets for the current study are available from the publicly available data set,^46–50.

Declarations

Competing interests

The authors declare no competing interests.