A medical image diagnostic report steganography method based on convergent generative adversarial networks

Abstract

The present invention discloses a medical image diagnostic report steganography method based on a convergent generative adversarial network, comprising: inputting a medical image and a corresponding diagnostic report, and steganizing the input diagnostic report into the medical image through a deep generative adversarial network to protect the patient's privacy; extracting private information such as the diagnostic report from the steganographic image through a deep text extraction network to achieve pathological information sharing within and between hospitals; proposing a "0-centered Wasserstein divergence" metric for measuring the manifold distance between the original image and the steganographic image distribution, and ensuring that the entire model can converge to a local optimal point; introducing a local regularization constraint into the generative network to further accelerate convergence while reducing the generalization error; the local regularization constraint satisfies locality and orthogonality, and is sampled from mixed Gaussian noise, which increases information disturbance, thereby improving the information embedding capability and security performance of the model.

Description

Technical Field

[0001] The present invention belongs to the field of information security and combines deep generative adversarial networks, zero-centered Wasserstein divergence and local regularization to achieve secure steganography of medical image diagnostic reports. Background Art

[0002] Hospitals generate a large number of medical images of various modalities and corresponding diagnostic reports every day. How to effectively protect patient pathological information, prevent privacy leaks, and prevent illegal treatments has become a hot research topic in both scientific research and clinical practice. Image steganography attempts to hide sensitive information within images deemed non-suspicious by attackers or eavesdroppers. Medical image steganography plays a crucial role in protecting patient pathological information and preventing privacy leaks. Currently, some researchers have proposed image steganography algorithms to embed secret information into images, thereby protecting image privacy.

[0003] The key to image steganography lies in embedding secret information into an image in a way that is both visually invisible and undetectable. Based on this, researchers have proposed excellent image steganography methods based on deep learning and adversarial learning, which can be applied to medical image steganography. Although existing image steganography methods can be directly applied to medical image steganography, three challenges remain: 1) Almost all image steganography algorithms based on generative adversarial networks fail to converge to a local optimum; 2) Compared to traditional steganography algorithms, although the steganographic capacity has been improved, the steganographic quality of images with high bit rates is relatively low; 3) When embedding a large amount of secret information, the steganographic images generated by current steganography algorithms have a high detectability rate. Summary of the Invention

[0004] The present invention aims to address the above-mentioned problems of the prior art by proposing a method for medical image diagnostic report steganography based on a convergent generative adversarial network, achieving high-capacity and highly secure diagnostic report steganography. During the image steganography process, the "zero-centered Wasserstein divergence" metric is proposed to measure the manifold distance between the original image and the stego-image distribution, ensuring that the entire model converges to a local optimal point, thereby ensuring that the entire model is a convergent deep model. In addition, a local regularization term is proposed to ensure that the generative network has a low generalization error, and it can be shown that the local regularization term can further accelerate model convergence. The local regularization constraint satisfies locality and orthogonality, and is sampled from a mixed Gaussian noise, which increases information perturbation and further improves the model's information embedding capability and security performance.

[0005] This paper establishes a convergent generative adversarial network, uses local regularization to accelerate model convergence, and uses information perturbation to further enhance the model's steganographic capabilities. The technical solutions adopted are as follows:

[0006] A method for steganographically capturing medical image diagnostic reports based on a convergent generative adversarial network (GAN) is proposed. First, a deep generative network is constructed to steganographically capture the input diagnostic report into the corresponding medical image. A deep text extraction network is then used to extract private information, such as the diagnostic report, from the steganographic image, enabling the sharing of pathology information within and between hospitals. During the image steganography process, the "0-centered Wasserstein divergence" metric is proposed to measure the manifold distance between the original image and the steganographic image distribution, ensuring that the entire model converges to a local optimal point, thereby ensuring that the entire model is a convergent deep model. Specifically, the following steps are included:

[0007] 1) During the image steganography process, a deep generative network is used to encode the diagnosis report and the original image respectively to obtain the continuity features of the corresponding modal data. The diagnosis report is then steganographically transferred to the corresponding medical image by fusing the two continuity features, thereby protecting patient privacy and enabling secure information sharing within and between hospitals.

[0008] 2) During the image steganography process, a local regularization term and the zero-centered Wasserstein divergence are introduced. Compared to other divergence metrics (such as KL distance and Wasserstein distance), the zero-centered Wasserstein divergence ensures that the eigenvalues ​​corresponding to the Jacobian matrix of the model's gradient vector field have negative real parts, thereby ensuring that the constructed model converges to a local optimum. The local regularization term ensures that the generated network has low generalization error and has been shown to further accelerate model convergence. The local regularization constraint satisfies locality and orthogonality, and is sampled from a mixed Gaussian noise, increasing information perturbation and further improving the model's information embedding capability and security performance.

[0009] 3) During the image decoding process, a deep text network is constructed to extract the diagnostic report from the steganographic image, and decoding loss is used to ensure that the extracted information is consistent with the original diagnostic report in both content and semantics.

[0010] 4) During the training process, an alternating training method is first adopted (i.e., when training one module, the parameters of another module are fixed). Then, these three modules are constructed into an end-to-end network for fine-tuning to further improve the steganographic quality of the overall model.

[0011] Furthermore, in step 1), the diagnosis report and original image are encoded separately through a deep generative network to obtain the continuity features of the corresponding modality data. The diagnosis report is then steganographically transferred to the corresponding medical image by fusing the two continuity features. In this process, the convergence of the model is ensured by the "zero-centered Wasserstein divergence" and the local regularization term, which is specifically expressed as:

[0012]

[0013] in, Represents the expectation, G, E and D represent the generation network, extraction network and discriminant network respectively. x, z and t represent the original medical image, Gaussian mixed noise and diagnosis report respectively, and E(G(x,z,t)) represents the text representation decoded by the decoder. k is a positive integer used to constrain the upper bound of the gradient penalty. The first two terms in the above formula represent the maximum and minimum optimization process, the third term is the gradient penalty, the fourth term is the decoding loss, and the last two terms are used for local regularization. α and β are hyperparameters used to adjust the weights between different losses. In order to simplify the expression, the above formula can be divided into three parts: steganalysis loss, decoding loss and local regularization term. The steganalysis loss and local regularization term are expressed as:

[0014]

[0015] in, represents the expectation, z represents Gaussian noise, t represents the diagnosis report. 0 represents the all-zero vector, J x represents the Jacobian matrix, Ι N represents the identity matrix of size N. The first term in the above formula is the steganalysis loss, which is used to drive the stego-image onto the manifold of the real image and maintain the authenticity of the stego-image. The second term in the above formula is the localization term, which is used to maintain locality. The last term is the orthogonal term, which is used to maintain orthogonality and remove redundancy. The decoding loss is used to reduce the difference between the original diagnosis report and the extracted description. During training, the L2 distance is used to construct the decoding loss, which is expressed as:

[0016]

[0017] In order to avoid the computational time-consuming problem caused by processing discrete data, the diagnosis report is encoded to obtain its implicit feature vector, and then the error between the feature vectors is calculated using the decoding loss.

[0018] Furthermore, in step 2), during the image steganography process, the "zero-centered Wasserstein divergence" is proposed. Compared with other divergence metrics (such as KL distance and Wasserstein divergence), the proposed metric ensures that the eigenvalues ​​corresponding to the Jacobian matrix of the model gradient vector field have negative real parts, thereby ensuring that the constructed model can converge to a local optimal point. The zero-centered Wasserstein divergence is defined as:

[0019]

[0020] Among them, k is a positive integer used to constrain the upper bound of the gradient penalty. r and P g represent the data distribution of real images and steganographic images respectively, Pu Represents the mixed distribution of real images and stego images. and They respectively represent the distribution P r , P g and P u expectations, Represents the gradient of the function. It represents the space of all first-order differentiable functions in an open, continuously differentiable space. For the truth function f, the most commonly used constraint f(t) = -log(1+exp(-t)) is added to the continuously differentiable space Ω.

[0021] By defining the non-negativity, symmetry and non-necessity of the distribution triangle inequality of the symmetric divergence, it is proved that the zero-centered Wasserstein divergence is a symmetric divergence measure that can better avoid the problems of gradient vanishing and training instability.

[0022] Let P r and P g is a probability distribution in an open, bounded, and continuous set Ω. For continuous distributions P and Q, if the distribution metric D(·) satisfies D(P,Q)≥0, and D(P,Q)=D(Q,P), then the metric is called a symmetric divergence.

[0023] Let x≡0, we have

[0024]

[0025] Since k is a positive integer, D(p(x),q(x))≥0 holds.

[0026] In the generative adversarial system, when p(x) = q(x), the generator and the discriminator achieve Nash equilibrium. At this time, E x～p(x) [f(x)] and E x～q(x) The contribution of [f(x)] cancels out, and the gradient of the discriminator cannot be updated, i.e. It can be seen that D(p(x),q(x))=0 when and only when p(x)=q(x). Select f(t)=-log(1+exp(-t)), and we can get:

[0027]

[0028] It can be seen that when p(x)≠q(x), D(p(x),q(x))>0. Therefore, it can be concluded that: When p(x)=q(x), E x～p(x) [f(x)] and E x～q(x)The contributions of [f(x)] cancel each other out, and the gradient penalty only acts on the discriminator. Therefore, it is obvious that D(P,Q)=D(Q,P).

[0029] In summary, the 0-centered Wasserstein divergence is a symmetric divergence measure.

[0030] By proving that the zero-centered Wasserstein divergence is a symmetric divergence measure, it can be further proved that this divergence can enable the deep generative adversarial network model to converge to the local optimal point, which has a positive effect on the convergence of the generative adversarial network model.

[0031] In the Dirac-GAN scenario, the gradient vector field Defined as in, represents the objective function of Dirac-GAN, θ and represent the parameters of the generator and discriminator respectively. and Respectively Regarding the parameters θ and Assume that the Dirac-GAN parameters converge to the equilibrium point is (0,0), we can calculate the gradient vector field At the balance point The Jacobian matrix of is: in, f'(0) is the first-order derivative of the equilibrium point. R is the 0-centered Wasserstein divergence regularization term, expressed as:

[0032] The calculation process is as follows:

[0033] Gradient vector field At the balance point The objective function is expressed as The corresponding Jacobian matrix is ​​specifically expressed as:

[0034]

[0035] And there

[0036]

[0037]

[0038] In the Dirac-GAN scenario, it can be concluded that Therefore, for any local neighbor point set P D For x, we can get:

[0039]

[0040]

[0041]

[0042]

[0043] Combining the above four equations, we can get the gradient vector field At the balance point The Jacobian matrix of

[0044] Calculate the Jacobian matrix The characteristic roots of are:

[0045]

[0046] It can be seen that all eigenvalues ​​have a negative real part. Therefore, when updating the network using the gradient descent algorithm, giving a sufficiently small learning rate can ensure that the network model is locally converged.

[0047] Furthermore, in step 2), by proposing a local regularization term, it can be proved that the local regularization term can further accelerate the convergence of the model while ensuring that the generated network has a lower generalization error.

[0048] When the Wasserstein-1 distance is used to estimate the distance from the empirical distribution to the true distribution, in a finite set of n samples, for a continuous distribution P in a compact space X n and The k-order derivative of Wasserstein distance approaches 0, that is, Scholars have shown that when the data dimension d>2, the upper bound of the convergence rate can be expressed as:

[0049]

[0050] Where n represents the number of training samples. Let P r and P g are the true distribution and the generated distribution respectively, and have and where d r and d g Represent the dimensions of the true distribution and the generated distribution respectively. For the empirical distribution P r ′ and P g ′, for n training samples, the convergence ratio product is:

[0051]

[0052] In the medical image diagnosis report steganography scenario, the dimensions of the true distribution and the generated distribution are equal, i.e., d r =d g =d. Therefore, the above formula can be expressed as:

[0053] When the local regularization term is introduced, the convergence ratio product is:

[0054] It is not difficult to see that Further we can get: In other words, the introduction of local regularization terms can lead to faster convergence.

[0055] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the medical image diagnostic report steganography method based on a convergent generative adversarial network as described above is implemented.

[0056] The advantages and beneficial effects of the present invention are as follows:

[0057] 1) A method for steganographic analysis of medical image diagnostic reports based on a convergent generative adversarial network is proposed. This method exhibits good convergence, high steganographic capacity, and high security. The method comprises a local generative network, an information extraction network, and a discriminative network. The local generative network outputs high-quality stego-images, the information extraction network extracts diagnostic reports from the stego-images, and the discriminative network diligently distinguishes between genuine and stego-images.

[0058] 2) The “0-centered Wasserstein divergence” is proposed to measure the manifold distance between the original image and the stego-image distribution, ensuring that the entire model can converge to the local optimal point, thus ensuring that the entire model is a convergent deep model.

[0059] 3) A local regularization term is proposed. While ensuring that the generative network has low generalization error, it can be shown to further accelerate model convergence. Local regularization constraints satisfy locality and orthogonality. Sampling from a mixture of Gaussian noise increases information perturbation, further improving the model's information embedding capability and security performance.

[0060] 4) This paper first employs an alternating training approach (i.e., while training one network, the others remain fixed) to train a local generative network, an information extraction network, and a discriminative network from scratch using a medical image dataset. These three modules are then constructed into an end-to-end network for fine-tuning, further improving the overall network's retrieval accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Provides an implementation algorithm framework diagram for the present invention;

[0062] Figure 2 Figure 1 is the medical image steganography experimental result of the present invention;

[0063] Figure 3 This is the second figure of the medical image steganography experimental results of the present invention;

[0064] Figure 4 This is a comparison diagram of the convergence results of the present invention. DETAILED DESCRIPTION

[0065] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0066] A deep generative adversarial network (GAN) is used to stego-embed diagnostic reports into medical images, protecting patient pathology information and preventing privacy leaks. During the image steganography process, a generative network is used to generate an image similar to the original image, preventing attackers from stealing private information. During the stego-image decoding process, an information extraction network is used to extract the diagnostic report from the stego-image, thus achieving secure image steganography. To ensure the convergence of the GAN and ultimately obtain an optimal model, a metric called "zero-centered Wasserstein divergence" is proposed. This metric measures the manifold distance between the original image and the stego-image distribution. This ensures that the eigenvalues ​​corresponding to the Jacobian matrix of the model's gradient vector field have negative real parts, thus ensuring that the constructed model converges to a local optimum. Furthermore, a local regularization term is proposed to ensure that the GAN has low generalization error and is proven to further accelerate model convergence. The local regularization constraints satisfy locality and orthogonality. The local regularization is sampled from a mixture of Gaussian noise, increasing information perturbation and further improving the model's information embedding capability and security performance. During the training process, an alternating training method is first adopted (i.e., when training one module, the parameters of another module are fixed), and then the three modules are constructed into an end-to-end network for fine-tuning to further improve the steganographic quality of the overall model.

[0067] The technical solution of the present invention will be described in detail below: a medical image diagnostic report steganography method based on a convergent generative adversarial network, comprising:

[0068] To protect patient pathology information and prevent privacy leaks and illegal treatment, a deep generative network is used to encode the diagnostic report and original image separately, obtaining the continuous features of the corresponding modal data. The diagnostic report is then steganographically embedded into the corresponding medical image by fusing the two continuous features. A deep text network is then constructed to extract the diagnostic report from the steganographic image, using a decoding loss to ensure that the extracted information is consistent with the original diagnostic report in both content and semantics.

[0069] In the image steganography process, the "zero-centered Wasserstein divergence" is proposed. Compared with other metrics (such as KL distance and Wasserstein distance), the proposed metric makes the eigenvalues ​​corresponding to the Jacobian matrix of the model gradient vector field have negative real parts, thereby ensuring that the constructed model can converge to the local optimal point.

[0070] In the image steganography process, a local regularization term is proposed to ensure that the generated network has a low generalization error. It can also be shown that the local regularization term can further accelerate the convergence of the model. The local regularization constraint satisfies locality and orthogonality. The sample is sampled from mixed Gaussian noise, which increases the information perturbation and further improves the information embedding ability and security performance of the model. Specifically expressed as:

[0071]

[0072] in, Represents the expectation, G, E and D represent the generation network, extraction network and discriminant network respectively. x, z and t represent the original medical image, Gaussian mixed noise and diagnosis report respectively. E(G(x,z,t)) represents the text representation decoded by the decoder. k is a positive integer used to constrain the upper bound of the gradient penalty. The first two terms in the above formula represent the maximum and minimum optimization process, the third term is the gradient penalty, the fourth term is the decoding loss, and the last two terms are used for local regularization. α and β are hyperparameters used to adjust the weights between different losses. J x represents the Jacobian matrix, Ι N represents the identity matrix of size N. To simplify the expression, the above formula can be divided into three parts: steganalysis loss, decoding loss and local regularization term.

[0073] like Figure 1 As shown, the specific steps are as follows:

[0074] Step 1: Build a local generation network

[0075] During the image steganography process, a deep generative network is used to encode the diagnosis report and the original image separately, obtaining the continuity features of the corresponding modal data. The diagnosis report is then steganographically transferred to the corresponding medical image by fusing the two continuity features. To accelerate convergence and reduce generalization error, a local regularization term is introduced to ensure locality and orthogonality. Specifically, it is expressed as:

[0076]

[0077] in, represents the expectation, z represents Gaussian mixed noise, t represents the diagnosis report. 0 represents the matrix of all zeros, J x represents the Jacobian matrix, Ι N Represents the identity matrix of size N.

[0078] Step 2: Build an information extraction network

[0079] During the image decoding process, a deep text network is constructed to extract the diagnostic report from the steganographic image. The decoding loss is used to ensure that the extracted information is consistent with the original diagnostic report in terms of both content and semantics. The specific expression is as follows:

[0080]

[0081] Step 3: Construct the “0-centered Wasserstein divergence”

[0082] The present invention proposes the "0-centered Wasserstein divergence" to make the characteristic roots corresponding to the Jacobian matrix of the model gradient vector field have negative real parts, thereby ensuring that the constructed model can converge to the local optimal point. 0-centered Wasserstein divergence:

[0083]

[0084] Among them, k is a positive integer used to constrain the upper bound of the gradient penalty. r and P g represent the data distribution of real images and steganographic images respectively, P u Represents the mixed distribution of real images and stego images. Denotes the space of all first-order differentiable functions in the open, continuously differentiable space Ω. For a truth function f, the most commonly used constraint f(t) = -log(1 + exp(-t)) is added to the continuously differentiable space Ω.

[0085] Step 4: Construct the objective function

[0086] Introducing hyperparameters α and β, combined with the “0-centered Wasserstein divergence” and local regularization term, constructs the overall maximization-minimization optimization, which is expressed as:

[0087]

[0088] In summary, the innovations and advantages of the present invention are:

[0089] The medical image diagnostic report steganography method proposed in the present invention utilizes a local generation network and an information extraction network to realize the steganography and decoding of diagnostic reports, thereby achieving the protection of patient privacy and the secure sharing of information within and between hospitals.

[0090] In response to the problem that the current steganography algorithms based on deep neural networks cannot converge, the present invention proposes a medical image diagnostic report steganography method based on convergent generative adversarial networks to ensure that the model has provable convergence.

[0091] The "0-centered Wasserstein divergence" is proposed, so that the characteristic roots corresponding to the Jacobian matrix of the model gradient vector field have negative real parts, thereby ensuring that the constructed model can converge to the local optimal point.

[0092] A local regularization term is proposed to ensure that the generated network has a lower generalization error. At the same time, it can be proved that the local regularization term can further accelerate the convergence of the model.

[0093] The local regularization constraints satisfy locality and orthogonality, and are sampled from mixed Gaussian noise, which increases information disturbance and further improves the information embedding capability and security performance of the model.

[0094] The medical image steganography experimental results obtained by the method of the present invention are as follows: Figure 2 and Figure 3 shown.

[0095] In order to verify the convergence of the algorithm, SGAN[4], HidingGAN[5], GRDH[6] and SteganoGAN[7] were selected as comparison methods for comparative experiments. Figure 4 The convergence of GAN-based steganography methods is demonstrated, all based on the assumptions of Dirac-GAN[8]. Figure 4 As shown in Figure 2, SGAN[4], HidingGAN[5], GRDH[6] and SteganoGAN[7] cannot converge to local equilibrium. In contrast, the algorithm proposed in this paper can achieve local convergence, as shown in Figure 2. Figure 4 (e). Thanks to the local regularization, the proposed algorithm converges faster, e.g. Figure 4 (f) shown.

[0096] [1]Li Q,Wang X,Wang X,et al.An encrypted coverless information hidingmethod based on generative models[J].Information Sciences,2021,553(3):19–30.

[0097] [2] Tancik M, Mildenhall B, Ng R. StegaStamp: Invisible hyperlinks in physical photographs [C] / / IEEE Conference on Compupter Vision and Pattern Recognition, 2021: 2117–2126.

[0098] [3]Jing J, Deng X, Xu M, et al. HiNet: Deep image hiding by invertiblenetwork[C] / / International Conference on Computer Vision, 2021:4733–4742.

[0099] [4]Jamie H, Danezis G. Generating steganographic images via adversarialtraining[C] / / Advances in Neural Information Processing Systems, 2017:1954–1963.

[0100] [5] Wang Z, Gao N, Wang X, et al. HidingGAN: High capacity information hiding with generative adversarial network [J]. Computer Graphics Forum, 2019, 38(7): 393–401.

[0101] [6]Zhang Z, Fu G, Di F, et al. Generative reversible data hiding by image to image translation via GANs[J]. Security and Communication Networks, 2019, 13(4): 21–31.

[0102] [7] Zhang K, Alfredo C, Xu L, et al. SteganoGAN: high capacity imagesteganography with GANs[Online]. 2019, https: / / arxiv.org / abs / 1901.03892.

[0103] The above embodiments should be understood as merely illustrating the present invention and not as limiting the scope of protection of the present invention. After reading the contents of the present invention, technicians may make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A medical image diagnostic report steganography method based on convergent generative adversarial networks, characterized in that: The following steps are involved: 1) During the image steganography process, a deep generative network is used to encode the diagnosis report and the original image separately to obtain the continuity features of the corresponding modality data. The diagnosis report is then steganographically transferred to the corresponding medical image by fusing the two continuity features. 2) In the image steganography process, a local regularization term and zero-centered Wasserstein divergence are introduced; The local regularization term is Among them, x and t represent the original medical image and diagnosis report respectively, 0 represents the all-zero vector, and J x represents the Jacobian matrix, Ι N represents the identity matrix of size N, represents the expectation, z is Gaussian noise, α and β are hyperparameters, P r represents the original medical image distribution, P z represents the sampled Gaussian distribution, D and G represent the discriminant network and the generative network respectively. The first term in the above formula is the steganalysis loss, which is used to drive the stego image onto the manifold of the real image to maintain the authenticity of the stego image. The second term is the localization term, which is used to maintain locality. The zero-centered Wasserstein divergence is defined as: Where k is a positive integer, P r and P g represent the data distribution of real images and steganographic images respectively, P u represents the mixed distribution of real images and steganographic images, and They respectively represent the distribution P r , P g and P u expectations, represents the gradient of the function, represents the space of all first-order differentiable functions in an open, continuously differentiable space. For a truth function f, the most commonly used constraint f(t) = -log(1+exp(-t)) is added to the continuously differentiable space Ω. 3) During the image decoding process, a deep text network is constructed to extract the diagnostic report from the steganographic image, and decoding loss is used to ensure that the extracted information is consistent with the original diagnostic report in both content and semantics.

2. The medical image diagnostic report steganography method based on a convergent generative adversarial network according to claim 1, characterized in that: Local regularization constraints satisfy two conditions: locality and orthogonality. The locality constraint is expressed as: The orthogonality constraint is expressed as: G represents the generative network, x and t represent the original medical image and diagnosis report respectively, 0 represents the all-zero vector, and J x represents the Jacobian matrix, Ι N represents the identity matrix of size N, Expresses expectation, P r represents the original medical image distribution.

3. The medical image diagnostic report steganography method based on a convergent generative adversarial network according to claim 1, characterized in that: The decoding loss is 4. A medical image diagnostic report steganography method based on a convergent generative adversarial network according to any one of claims 1 to 3, characterized in that: The constructed objective function is:

5. The medical image diagnostic report steganography method based on a convergent generative adversarial network according to claim 4, characterized in that: During the training process, an alternating training method is first adopted, and then the network is constructed as an end-to-end model for fine-tuning.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the medical image diagnostic report steganography method based on a convergent generative adversarial network as described in any one of claims 1 to 5.

Images