A Medical Image Steganography Method and System Based on Reversible Deep Networks

Abstract

This invention discloses a medical image steganography method and system based on reversible deep networks. To address the limitations of existing technologies, such as limited steganography capacity and difficulty in adapting to multimodal medical image scenarios, the method of this invention includes: receiving the original medical image and the secret information to be embedded, and normalizing and spatially aligning the medical image; inputting the image into a multi-scale feature extraction network to extract low-level texture features, mid-level edge features, and high-level semantic features, and concatenating them along the channel dimension to form a fused feature; concatenating the fused feature with the encrypted secret information and inputting it into a reversible neural network encoder to generate a steganographic image; reversibly extracting the fused feature and encrypted information through a reversible neural network decoder to reconstruct the original medical image and secret information; and optimizing network performance using a structure-aware joint loss function. This method can be applied to secure storage of medical images, multi-center collaboration, and medical blockchain evidence storage.

Description

Technical Field

[0001] This invention relates to the field of information security technology, specifically to a medical image steganography method based on reversible deep networks, and also to a medical image steganography system based on reversible deep networks. Background Technology

[0002] With the rapid development of medical informatization, medical images (such as CT, MRI, and PET) have become core data resources in clinical diagnosis, remote consultation, and scientific research analysis. In practical applications, medical images not only carry important lesion information and anatomical structures, but may also contain sensitive content such as patient identity and medical record tags. Therefore, ensuring the security, privacy, and traceability of images during storage, transmission, and sharing has become a critical issue that urgently needs to be addressed in the field of medical image processing.

[0003] To enhance the information security of medical images, researchers have proposed various image protection methods. Among them, traditional encryption methods (such as AES and RSA) can effectively prevent information leakage by encrypting image data. However, encrypted images cannot be used directly for diagnosis and must be decrypted before viewing, which affects clinical efficiency and the convenience of processing procedures.

[0004] Anonymization methods mainly rely on deleting DICOM header information or image tags to mask identity identifiers, but the image itself may still be able to identify the patient's identity through structural features, resulting in limited privacy protection.

[0005] In addition, traditional image steganography techniques (such as LSB least significant bit embedding and DCT domain steganography) have been widely studied and applied in natural images. However, since medical images are extremely sensitive to grayscale distribution and structural information, directly applying the above methods often introduces artifacts or structural distortions, thereby affecting the reliability of diagnosis.

[0006] In recent years, image steganography techniques based on deep learning have gradually developed. For example, steganography methods based on convolutional neural networks (CNNs) or generative adversarial networks (GANs) have higher concealment and information embedding capacity. However, most of these methods lack complete reversibility, and the embedding process cannot recover the original image, failing to meet the dual requirements of "diagnostic usability + data traceability" in medical data. In addition, existing research has mostly focused on single modalities (such as MRI or CT), while in clinical practice, multimodal fusion applications (such as PET / CT, MRI / CT) are becoming increasingly common, and existing methods are difficult to adapt to the cross-modal steganography needs between multi-source images.

[0007] In summary, existing technologies suffer from limitations such as limited steganography capacity, susceptibility to image structure distortion, lack of bidirectional reversibility, and difficulty in adapting to multimodal medical image scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide a medical image steganography method and system based on reversible deep networks, in order to solve the problems of limited steganography capacity, easy distortion of image structure, lack of bidirectional reversibility, and difficulty in adapting to multimodal medical image scenarios in the existing technology.

[0009] To achieve the above functions, this invention designs a medical image steganography method based on reversible deep networks, executing the following steps S1-S6 to construct a reversible neural network, including a reversible neural network encoder and a reversible neural network decoder, to complete the encryption and decryption of medical images:

[0010] Step S1: Receive the original medical image and the secret information to be embedded, and normalize the original medical image while performing spatial alignment.

[0011] Step S2: Construct a multi-scale feature extraction network. Input the normalized medical image into the multi-scale feature extraction network to extract the low-level texture features, mid-level edge features and high-level semantic features of the medical image, and then stitch them together in the channel dimension to form a fused feature.

[0012] Step S3: Encrypt the secret information to be embedded based on its data format;

[0013] Step S4: Construct a reversible neural network encoder, concatenate the fused features and the encrypted secret information in the channel dimension to form a joint representation, and input the joint representation into the reversible neural network encoder to generate a steganalytic image;

[0014] Step S5: Construct a reversible neural network decoder. Input the steganalyte image into the reversible neural network decoder, extract the fused features and encrypted secret information through reverse mapping, and reconstruct the original medical image and secret information respectively.

[0015] Step S6: Construct a structure-aware joint loss function, train the reversible neural network to obtain a trained reversible neural network, deploy a reversible neural network encoder at the transmitting end, and deploy a reversible neural network decoder at the receiving end to complete the encryption and decryption of medical images.

[0016] As a preferred technical solution of the present invention, the specific steps of step S1 are as follows:

[0017] Step S1.1: Receive the raw medical images The dimension is H×W, where H and W are the height and width of the original medical image, respectively, and the pixel grayscale value range is [0,255]; and the secret information to be embedded. ;

[0018] Step S1.2: Process the original medical images The normalization process is performed as follows:

[0019] ;

[0020] in, This represents the intensity value of the original medical image at (x, y). This represents the intensity value of the normalized medical image at (x, y). Represents the original medical image The lowest strength value in, Represents the original medical image The highest strength value in;

[0021] Step S1.3: Pre-extract the preset region of interest based on the clinical task. The mask of the region of interest is represented by the following formula:

[0022] ;

[0023] In the formula, Indicates the region of interest; This represents the mask at (x,y) in the region of interest;

[0024] Step S1.4: Perform spatial alignment using the following formula:

[0025] ;

[0026] In the formula, Represents the PET image variation function. This represents the intensity value at (x, y) of the PET image after affine transformation and resampling to the reference coordinates. , , , , , This represents the parameters of a two-dimensional affine transformation.

[0027] As a preferred technical solution of the present invention, the specific steps of step S2 are as follows:

[0028] Step S2.1: Extract low-level texture features As shown in the following formula:

[0029] ;

[0030] In the formula, This represents the low-level encoder function consisting of the first 2-3 layers of convolutions and ReLU activation in a multi-scale feature extraction network; This represents a normalized medical image;

[0031] Step S2.2: Extract mid-layer edge features As shown in the following formula:

[0032] ;

[0033] In the formula, Represents downsampling and mesoscale convolution functions;

[0034] Step S2.3: Extract high-level semantic features As shown in the following formula:

[0035] ;

[0036] In the formula, This represents a function from a residual module or a Transformer module.

[0037] Step S2.4: Transfer low-level texture features Mid-layer edge features High-level semantic features By concatenating and fusing along the channel dimension, the fused feature is obtained as follows:

[0038] ;

[0039] In the formula, Indicates fusion characteristics, This indicates a splicing operation.

[0040] As a preferred technical solution of the present invention, the specific steps of step S3 are as follows:

[0041] Step S3.1: The data format of secret information includes text, tags, and images; if the data format of secret information is text or tags, then binary encoding is used; if the data format of secret information is an image, then a pixel matrix is ​​used directly.

[0042] Step S3.2: For the secret information in text or tag data format, perform AES encryption using the Advanced Encryption Standard (AES), expressed as follows:

[0043] ;

[0044] In the formula, This represents the secret information encrypted with AES. This indicates a symmetric encryption operation based on a pre-shared key. Decryption requires the same pre-shared key to perform the inverse operation. Indicates secret information;

[0045] Step S3.3: For the secret information in the image data format, a Gaussian perturbation is used for diffusion perturbation encryption, expressed as the following formula:

[0046] ;

[0047] In the formula, This represents the secret information encrypted using diffusion perturbation; Indicates Gaussian noise. The variance of Gaussian noise is represented;

[0048] Step S3.4: In practical applications, AES encryption is combined with diffusion perturbation encryption, as shown in the following formula:

[0049] ;

[0050] In the formula, This represents secret information obtained by combining AES encryption with diffusion perturbation encryption.

[0051] As a preferred embodiment of the present invention, the specific steps of step S4 are as follows:

[0052] Step S4.1: encrypt the secret information obtained in step S3. Fusion features obtained in step S2 Concatenate into joint input:

[0053] ;

[0054] In the formula, Indicates joint input. This indicates a splicing operation along the channel dimension;

[0055] Step S4.2: Construct a reversible neural network encoder to input the joint input. The input is fed into a reversible neural network encoder to generate a steganalytic image. :

[0056] ;

[0057] In the formula, It is a combination function of a two-way coupled structure;

[0058] Step S4.3: The reversible neural network is reversible, meaning it has an inverse mapping function such that the following equation holds:

[0059] ;

[0060] In the formula, The inverse mapping function representing the combination function of a two-way coupled structure;

[0061] Extract the encrypted secret information according to the above formula. With fusion features .

[0062] As a preferred embodiment of the present invention: the reversible neural network encoder defines the output steganalyte image using the following residual method. :

[0063] ;

[0064] In the formula, The perturbation term obtained from learning, This represents a normalized medical image.

[0065] As a preferred embodiment of the present invention, the specific steps of step S5 are as follows:

[0066] Step S5.1: Construct a reversible neural network decoder. Based on the reversibility of the reversible neural network, process the steganalyte image through an inverse mapping function, as shown in the following formula:

[0067] ;

[0068] In the formula, The reconstructed and fused feature map obtained by processing the inverse mapping function, The reconstructed secret information is processed by the inverse mapping function; The inverse mapping function representing the combination function of a two-way coupled structure;

[0069] Step S5.2: Utilize the reconstruction function The reconstructed fusion feature map is then converted back to the original medical image format:

[0070] ;

[0071] In the formula, For reconstruction function, Represents the original medical image used for reconstruction;

[0072] Step S5.3: If the secret information is encrypted with AES or diffusion perturbation, then perform the following reverse process:

[0073] ;

[0074] In the formula, This represents the secret information decoded. This represents the inverse process of AES encryption;

[0075] At the same time, the encrypted part of the diffusion perturbation is restored by averaging or filtering to obtain the final secret information.

[0076] As a preferred embodiment of the present invention, the specific steps of step S6 are as follows:

[0077] Step S6.1: Constructing the hidden loss As shown in the following formula:

[0078] ;

[0079] in, Represents the similarity function of balanced structures. Indicates control weight, This represents the steganalysis output of a reversible neural network encoder. This represents a normalized medical image;

[0080] Step S6.2: Construct the reconstruction loss As shown in the following formula:

[0081] ;

[0082] In the formula, This represents the reconstructed secret information output by the reversible neural network decoder. This represents the secret information output by the reversible neural network encoder. Indicates the L1 distance;

[0083] Step S6.3: Construct the ROI structural loss As shown in the following formula:

[0084] ;

[0085] In the formula, This represents the mask at (x,y) in the region of interest; This represents the intensity value of the steganalysis at (x, y). This represents the intensity value of the normalized medical image at (x, y);

[0086] Step S6.4: Construct the structure-aware joint loss function as follows:

[0087] ;

[0088] In the formula, This represents the structure-aware joint loss function. , Reconstruction losses ROI structural loss The weighting coefficients.

[0089] This invention also designs a medical image steganography system based on reversible deep networks, comprising:

[0090] One or more processors and a memory, the memory storing operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, a reversible neural network is obtained through the following steps, including a reversible neural network encoder and a reversible neural network decoder, to implement the aforementioned medical image steganography method based on a reversible deep network:

[0091] Step S1: Receive the original medical image and the secret information to be embedded, and normalize the original medical image while performing spatial alignment.

[0092] Step S2: Construct a multi-scale feature extraction network. Input the normalized medical image into the multi-scale feature extraction network to extract the low-level texture features, mid-level edge features and high-level semantic features of the medical image, and then stitch them together in the channel dimension to form a fused feature.

[0093] Step S3: Encrypt the secret information to be embedded based on its data format;

[0094] Step S4: Construct a reversible neural network encoder, concatenate the fused features and the encrypted secret information in the channel dimension to form a joint representation, and input the joint representation into the reversible neural network encoder to generate a steganalytic image;

[0095] Step S5: Construct a reversible neural network decoder. Input the steganalyte image into the reversible neural network decoder, extract the fused features and encrypted secret information through reverse mapping, and reconstruct the original medical image and secret information respectively.

[0096] Step S6: Construct a structure-aware joint loss function, train the reversible neural network to obtain a trained reversible neural network, deploy a reversible neural network encoder at the transmitting end, and deploy a reversible neural network decoder at the receiving end to complete the encryption and decryption of medical images.

[0097] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0098] By introducing a multi-scale feature extraction network, which simultaneously extracts low-level texture, mid-level edge, and high-level semantic information, a comprehensive model of the multi-level structural features of medical images is achieved, ensuring that key anatomical structures are not destroyed during steganography. Compared to traditional single-scale convolution or direct embedding steganography methods, this approach significantly improves the structural fidelity and diagnostic usability of steganographic images, maintaining high embedding capacity while also considering the clinical value of medical images.

[0099] By utilizing reversible neural networks (INNs) for steganographic embedding, which possesses bidirectional information mapping capabilities, a completely reversible reconstruction of the original image and the embedded information can be achieved. Unlike existing GAN or CNN-based steganography methods, this invention avoids information loss during the encoding and decoding process of image content, meeting the stringent regulatory requirements for image traceability in medical scenarios. It represents a key innovation in steganography systems that combines concealment and reproducibility.

[0100] By introducing a combined protection mechanism of AES symmetric encryption and Gaussian diffusion perturbation before steganography, the embedded information cannot be decrypted without a key, even if it is detected or extracted, effectively enhancing the system's resistance to attacks. Compared to information security mechanisms that rely solely on steganography, this invention achieves dual security protection of "content leakage prevention + presence detection prevention" through the collaborative design of encryption and steganography, improving the system's robustness during medical data transmission.

[0101] An explicit reverse decoding process was designed, utilizing a reversible network for reverse mapping to obtain image features and embedded information, supplemented by decryption and reconstruction modules, effectively ensuring high-precision recovery of the original image and embedded data. Compared to the limitations of some existing deep steganography schemes that can only extract hidden data but cannot restore the image, the decoding mechanism of this invention is more in line with the application requirements of "zero information loss" in medical imaging.

[0102] By introducing a structure-aware loss function, which jointly considers pixel similarity, hidden information restoration accuracy, and structural consistency of the ROI region, the steganalyte image achieves good diagnostic quality at both the global and local levels. This loss design overcomes the shortcomings of traditional steganalysis methods that only focus on overall image quality, particularly strengthening the fidelity guarantee for key diagnostic regions (such as lesions), and effectively avoiding interference from steganalysis on clinical judgment.

[0103] In multimodal application scenarios, spatial alignment between different modalities such as PET and CT is achieved through image registration mechanisms, making cross-modal steganography possible. Compared to existing methods that are usually limited to single-modal image processing, this approach expands the applicability of steganography in multi-center collaboration and joint diagnosis, and is an important enhancement to the scenario adaptability of existing technologies.

[0104] This invention can be applied to secure storage of medical images, remote consultation, multi-center collaboration, and medical blockchain evidence storage. Attached Figure Description

[0105] Figure 1 This is a flowchart of a medical image steganography method based on reversible deep networks provided according to an embodiment of the present invention. Detailed Implementation

[0106] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0107] The medical image steganography method based on reversible deep networks provided in this embodiment of the invention refers to... Figure 1 Perform the following steps S1-S6 to construct a reversible neural network (INN), including a reversible neural network encoder and a reversible neural network decoder, to complete the encryption and decryption of medical images:

[0108] Step S1: Receive the original medical image and the secret information to be embedded, and normalize the original medical image while performing spatial alignment.

[0109] The specific steps are as follows:

[0110] Step S1.1: Receive the raw medical images The dimension is H×W, where H and W are the height and width of the original medical image, respectively, and the pixel grayscale value range is [0,255]; and the secret information to be embedded. ;

[0111] The medical image formats include common imaging types such as CT, MRI, PET, and ultrasound images. Supported image formats include DICOM, PNG, and TIFF, with image resolutions ranging from 256×256 to 1024×1024 pixels. Secret information may include patient diagnostic labels, physician annotations, structured data, or another image. The secret information can be in the form of an image or binary code, with image size varying within the range of 256×256 to 1024×1024 pixels.

[0112] Step S1.2: For standardized processing, the original medical images are processed... Normalization is performed, mapping pixel values ​​to a floating-point range between 0 and 1 to improve the stability of subsequent network training and inference, as shown in the following formula:

[0113] ;

[0114] in, This represents the intensity value of the original medical image at (x, y). This represents the intensity value of the normalized medical image at (x, y). Represents the original medical image The lowest strength value in, Represents the original medical image The highest strength value in;

[0115] Step S1.3: Pre-extract a pre-defined region of interest (ROI) mask based on the clinical task, such as lesions, organ boundaries, or specific anatomical structures. This mask can be automatically generated using existing segmentation models or obtained through manual annotation. When processing multimodal images such as PET and CT, spatial registration between images is also required. Affine or non-rigid registration methods are preferred to achieve a one-to-one correspondence between anatomical structures across modalities, providing a foundation for subsequent feature alignment and joint embedding. The ROI mask is represented by the following formula:

[0116] ;

[0117] In the formula, Indicates the region of interest; This represents the mask at (x,y) in the region of interest;

[0118] Step S1.4: For multimodal scenes (such as PET and CT), spatial alignment needs to be achieved through registration, using an affine transformation model, as shown in the following formula:

[0119] ;

[0120] In the formula, Represents the PET image variation function. This represents the intensity value at (x, y) of the PET image after affine transformation and resampling to the reference coordinates. , , , , , Represents the parameters of a two-dimensional affine transformation;

[0121] Spatial alignment ensures that images of different modalities correspond precisely in terms of anatomical structure, which facilitates subsequent joint embedding.

[0122] Step S2: Construct a multi-scale feature extraction network. Input the normalized medical image into the multi-scale feature extraction network to extract the low-level texture features, mid-level edge features and high-level semantic features of the medical image, and then stitch them together in the channel dimension to form a fusion feature. Through the hierarchical modeling method, it can be ensured that the steganalysis network retains key anatomical structures during the embedding process, while also having semantic understanding capabilities.

[0123] A multi-scale feature extraction network is used to model the normalized medical images. Depending on the scenario, U-Net or Swin Transformer can be selected as the backbone network. U-Net employs a layer-by-layer convolutional and skip connection structure, suitable for single-modal image processing; Swin Transformer uses a window attention mechanism, possessing cross-region feature perception capabilities, suitable for multi-modal image fusion representation. The entire extraction process is divided into three sub-modules: the first module extracts low-level texture features, capturing image details and texture changes; the second module extracts mid-level edge features, identifying lesion boundaries and organ contours; the third module extracts high-level semantic features, modeling the global spatial structure of the image. These three types of features are concatenated along the channel dimension to form a fused feature tensor for subsequent steganography embedding. Step S2 is detailed below:

[0124] Step S2.1: Extract low-level texture features These features typically have small receptive fields, allowing the extraction of local image details, such as tissue texture and edge noise patterns, using shallow convolutional neural networks suitable for capturing high-frequency signals. Specifically, the formula is as follows:

[0125] ;

[0126] In the formula, This represents the low-level encoder function consisting of the first 2-3 layers of convolutions and ReLU activation in a multi-scale feature extraction network; This represents a normalized medical image;

[0127] Step S2.2: Extract mid-layer edge features It captures organ contours and lesion edge structures through downsampling and mesoscale convolution extractors suitable for expressing mesoscale structural relationships. Specifically, it is expressed as follows:

[0128] ;

[0129] In the formula, Represents downsampling and mesoscale convolution functions;

[0130] Step S2.3: Extract high-level semantic features This involves using deeper residual or Transformer modules to model the global anatomical context of the image, such as the spatial structure of the heart and lungs, and left-right symmetry. Specifically, the formula is as follows:

[0131] ;

[0132] In the formula, This represents a function from a residual module or a Transformer module.

[0133] Step S2.4: The above features have the same spatial resolution (which can be aligned through upsampling), and the low-level texture features are... Mid-layer edge features High-level semantic features By concatenating and fusing along the channel dimension, the fused feature is obtained as follows:

[0134] ;

[0135] In the formula, Indicates fusion characteristics, This indicates a splicing operation. The merged result... It features rich texture boundaries and semantic support, serving as input for subsequent reversible steganography embedding modules.

[0136] Medical images contain multi-level features, and different lesions show significant differences at different scales. Multi-scale feature extraction networks can simultaneously capture texture details, edge contours, and global anatomical semantic information. The following are two application examples of multi-scale feature extraction networks:

[0137] In one embodiment, the multi-scale feature extraction network uses U-Net as the backbone network:

[0138] Network structure: U-Net encoder-decoder framework, which uses 3×3 convolutional kernels and 2×2 pooling to extract features layer by layer, and fuses shallow details and deep semantics through skip connections;

[0139] Channel configuration: 64 channels for low-level, 128 channels for mid-level, and 256 channels for high-level;

[0140] Applicable scenarios: Single-modal image processing (pure CT images or pure MRI images), with a stable network structure and high computational efficiency.

[0141] In another embodiment, the multi-scale feature extraction network uses the Swing Transformer as the backbone network:

[0142] Network structure: A Transformer architecture based on window attention, which utilizes local window and cross-window shift mechanisms to achieve global feature modeling;

[0143] Channel configuration: 96 channels for low-rise buildings, 192 channels for mid-rise buildings, and 384 channels for high-rise buildings;

[0144] Applicable scenarios: cross-modal input (such as steganography from PET images to CT images), better at capturing remote dependencies and cross-modal feature alignment relationships.

[0145] Step S3: Encrypt the secret information to be embedded based on its data format;

[0146] The specific steps are as follows:

[0147] Step S3.1: The data format of secret information includes text, tags, and images; if the data format of secret information is text or tags, then binary encoding is used; if the data format of secret information is an image, then a pixel matrix is ​​used directly.

[0148] Step S3.2: For the secret information in text or tag data format, perform AES encryption using the Advanced Encryption Standard (AES), expressed as follows:

[0149] ;

[0150] In the formula, This represents the secret information encrypted with AES. This indicates a symmetric encryption operation based on a pre-shared key, which requires the same pre-shared key to be used for inverse operation during decryption;

[0151] Step S3.3: For secret information in image data formats or scenarios requiring additional randomness, use Gaussian perturbation for diffusion perturbation encryption, expressed as the following formula:

[0152] ;

[0153] In the formula, This represents the secret information encrypted using diffusion perturbation; Indicates Gaussian noise. The variance of Gaussian noise is represented. Control the noise amplitude to ensure that disturbances do not affect the reversibility of decoding;

[0154] Step S3.4: In practical applications, AES encryption is combined with diffusion perturbation encryption to achieve dual protection, as shown in the following formula:

[0155] ;

[0156] In the formula, This represents secret information obtained by combining AES encryption with diffusion perturbation encryption.

[0157] Encrypting secret information before steganography can improve system security. Even if the steganographic image is intercepted, the original information cannot be obtained without decryption. The following are two application examples of encrypting secret information:

[0158] In one embodiment, the AES symmetric encryption method is used:

[0159] Algorithm type: AES-128 or AES-256 encryption, using a shared key for encryption and decryption;

[0160] Applicable scenarios: Medical information transmission with high security requirements.

[0161] In another embodiment, AES + diffusion perturbation combination encryption is used:

[0162] Process: First, the secret information is encrypted with AES, and then Gaussian perturbation is superimposed to simulate the diffusion process.

[0163] Parameter range: The perturbation standard deviation σ is 0.01~0.05, which can be adjusted according to the balance between concealment and decoding accuracy.

[0164] Advantages: Double encryption significantly improves resistance to selective attacks.

[0165] Step S4: Construct a reversible neural network encoder, concatenate the fused features and the encrypted secret information in the channel dimension to form a joint representation, and input the joint representation into the reversible neural network encoder to generate a steganalytic image;

[0166] The fused features are concatenated with encrypted secret information along the channel dimension. Encryption methods can include symmetric encryption algorithms such as AES-128 or AES-256, or Gaussian perturbations can be added to enhance randomness and resistance to attacks. The embedding module employs a reversible neural network structure, such as RealNVP or Glow networks. These networks consist of multiple coupled layers, each applying a reversible transformation to a portion of the input features while retaining another portion for direct transmission. Inter-layer parameter sharing or progressive design ensures that both the features and the embedded information can be fully recovered through reverse computation. The resulting output image is the steganographic image, which can be used for remote transmission, cloud storage, or blockchain notarization.

[0167] The specific steps of step S4 are as follows:

[0168] Step S4.1: The encoded or encrypted secret information obtained in step S3... Fusion features obtained in step S2 Concatenate into joint input:

[0169] ;

[0170] In the formula, Indicates joint input. This indicates a splicing operation along the channel dimension;

[0171] Step S4.2: Construct a reversible neural network encoder (forward path) to input the joint input. The input is fed into a reversible neural network encoder to generate a steganalytic image. :

[0172] ;

[0173] In the formula, It is a combination function of a bidirectional coupled structure, composed of reversible modules, such as NICE, RealNVP or Glow reversible network structures;

[0174] In one embodiment, to control the embedding strength and maintain image quality, the reversible neural network encoder can define the output steganalyte image using the following residual method. :

[0175] ;

[0176] In the formula, The norm of the perturbation term obtained through learning can be adjusted through the loss function; This represents a normalized medical image.

[0177] Step S4.3: The reversible neural network is reversible, meaning it has an inverse mapping function such that the following equation holds:

[0178] ;

[0179] In the formula, The inverse mapping function representing the combination function of a two-way coupled structure;

[0180] Extract the encrypted secret information according to the above formula. With fusion features This allows for the extraction of embedded information and image representations without losing any information.

[0181] Reversible neural networks (INNs) can achieve bidirectional mapping while maintaining information integrity, making them ideal for medical image steganography. Two application examples of reversible neural networks are provided below:

[0182] In one embodiment, the reversible neural network employs the RealNVP reversible network:

[0183] Structural features: Based on a coupling layer and an affine transformation, the invertible structure is divided into two parts for each layer's input. One part is directly passed, and the other part is combined after undergoing an affine transformation.

[0184] Embedding process: fusing features With secret information A steganalyte is generated by stitching together the data along the channel dimension and then performing a multi-layer coupling transformation. .

[0185] In another embodiment, the reversible neural network employs the Glow reversible network:

[0186] Structural features: Employs 1×1 reversible convolution and affine coupling transformation, supporting deeper networks and higher embedding capacity;

[0187] Applicable scenarios: Suitable for high-resolution images (≥512×512) or scenarios with large embedded information capacity.

[0188] Step S5: Construct a reversible neural network decoder. Input the steganalyte image into the reversible neural network decoder, extract the fused features and encrypted secret information through reverse mapping, and reconstruct the original medical image and secret information respectively.

[0189] At the receiving end, the system receives the steganalyzed image and performs inverse mapping through a homogeneous reversible neural network decoder path to recover the fused features and encrypted information. The fused features can be further used to reconstruct the original image, and the encrypted information is recovered as the original secret information through key decryption and perturbation restoration. The decoding process does not introduce any additional information loss, ensuring the integrity of the image and data.

[0190] The specific steps of step S5 are as follows:

[0191] Step S5.1: Construct a reversible neural network decoder. Based on the reversibility of the reversible neural network, process the steganalyte image through an inverse mapping function, as shown in the following formula:

[0192] ;

[0193] In the formula, The reconstructed and fused feature map obtained by processing the inverse mapping function, The reconstructed secret information (encrypted state) obtained by the inverse mapping function is processed; The inverse mapping function representing the combination function of a two-way coupled structure;

[0194] Step S5.2: Utilize the reconstruction function The reconstructed fusion feature map is then converted back to the original medical image format:

[0195] ;

[0196] In the formula, To reconstruct the function, a decoding convolutional network or an inverse transform module is used. Represents the original medical image used for reconstruction;

[0197] Step S5.3: If the secret information is encrypted with AES or diffusion perturbation, then perform the following reverse process:

[0198] ;

[0199] In the formula, This represents the secret information decoded. This represents the inverse process of AES encryption;

[0200] At the same time, the encrypted part of the diffusion perturbation is restored by averaging or filtering to obtain the final secret information.

[0201] Step S6: Construct a structure-aware joint loss function, train the reversible neural network to obtain a trained reversible neural network, deploy a reversible neural network encoder at the transmitting end, and deploy a reversible neural network decoder at the receiving end to complete the encryption and decryption of medical images.

[0202] To ensure the diagnostic usability of steganalytes and the complete recovery of hidden information, this invention employs multiple structure-aware loss functions to optimize the network during the training phase. The stealth loss term measures the pixel differences and structural similarity between the steganalyte and the original image, typically using a weighted average of L1 loss and the SSIM metric. The reconstruction loss term measures the accuracy between the embedded and recovered information, using mean absolute error to measure recovery quality. The ROI structure loss term targets the region of interest, constraining the grayscale and boundary consistency of that region through masking to prevent the embedding process from affecting key diagnostic areas. The final total loss is a weighted combination of these three terms, with the weights adjusted experimentally to balance embedding capacity and reconstruction accuracy.

[0203] It supports multiple deployment modes, including centralized server deployment, edge device deployment, and hybrid cloud deployment. In centralized deployment, high-performance GPU servers can be used to perform model inference and link with the PACS system; in edge device deployment, model pruning and quantization techniques can be used to deploy on low-power hardware such as Jetson to achieve real-time steganography in bedside or mobile diagnostic scenarios; in hybrid cloud architecture, the front end performs image preprocessing and encryption operations, while the back end completes deep network embedding and decoding, adapting to cross-institutional medical collaboration.

[0204] The specific steps of step S6 are as follows:

[0205] Step S6.1: Constructing the hidden loss This is used to constrain the differences between the stegana and the original image in terms of pixel-wise and structural similarity, as shown in the following formula:

[0206] ;

[0207] in, Represents the similarity function of balanced structures. Indicates control weight, This represents the steganalysis output of a reversible neural network encoder. This represents a normalized medical image;

[0208] Step S6.2: Construct the reconstruction loss This is used to constrain the accuracy of extracting secret information, specifically as follows:

[0209] ;

[0210] In the formula, This represents the reconstructed secret information output by the reversible neural network decoder. This represents the secret information output by the reversible neural network encoder. Represents the L1 distance, measuring the error in information reconstruction;

[0211] Step S6.3: Construct the ROI structural loss This is used to measure the fidelity of the ROI region, and is specifically expressed as follows:

[0212] ;

[0213] In the formula, This represents the mask at (x,y) in the region of interest; This represents the intensity value of the steganalysis at (x, y). This represents the intensity value of the normalized medical image at (x, y);

[0214] Step S6.4: Medical image diagnosis is highly sensitive to ROI (lesion, anatomical structure), therefore a comprehensive loss function needs to be designed to ensure the consistency between the steganalyte and the original image in the diagnostic region. The structure-aware joint loss function is constructed as follows:

[0215] ;

[0216] In the formula, This represents the total loss function and the structure-aware joint loss function. , Reconstruction losses ROI structural loss The weighting coefficients can be adjusted experimentally. In one embodiment, α=0.5 and β=1.0.

[0217] The following are three application examples of hardware architecture deployment:

[0218] In Example 1, a centralized server deployment is used:

[0219] Hardware components: medical image acquisition module (CT / PET / MRI interface), GPU accelerated server (such as NVIDIA A100, ≥40GB video memory), AES encryption module, PACS storage interface;

[0220] Workflow: Image acquisition → Local preprocessing → GPU execution of multi-scale feature extraction and reversible steganography embedding → Encryption → Storage / cloud transmission;

[0221] Applicable scenarios: Large hospital radiology departments, batch steganography and multi-center collaboration in research institutions.

[0222] In Example 2, a portable embedded device is used for deployment:

[0223] Hardware components: edge computing chip (such as NVIDIA Jetson Xavier NX), low-power AES module, local storage and wireless communication module;

[0224] Features: Suitable for bedside imaging, mobile healthcare and emergency scenarios, power consumption can be reduced through model quantization (FP16 / INT8).

[0225] In Example 3, a hybrid cloud deployment is adopted:

[0226] Structural design: The local module performs preprocessing and encryption, while the cloud module completes deep steganography embedding and decoding. Data transmission adopts the TLS / SSL security protocol.

[0227] Advantages: Suitable for cross-institute collaboration and remote image viewing, reducing local computing pressure.

[0228] The method designed in this invention was experimentally verified, with the following settings:

[0229] In Experiment 1, single-modal CT stegography was performed:

[0230] Input: 512×512 CT image; hidden information is 512 bytes of text;

[0231] Method configuration: U-Net+RealNVP architecture; AES-128 encryption; fixed weight loss (α=0.5, β=1.0);

[0232] Performance metrics: PSNR ≥ 42 dB after steganography, SSIM ≥ 0.985, and hidden information extraction accuracy ≥ 99.8%.

[0233] In Experiment 2, cross-modal PET→CT stegography was performed:

[0234] Input: Registered PET and CT images; Embedded information is the PET image;

[0235] Method configuration: Swing Transformer feature extraction + Glow reversible network; AES + diffusion perturbation encryption;

[0236] Performance metrics: Embedding capacity reaches 20% of image pixels; cross-modal reconstruction error MSE ≤ 1e-3.

[0237] In Experiment 3, the security enhancement mode was implemented:

[0238] Input: 512×512 MRI image; hidden information is structured diagnostic data;

[0239] Method configuration: AES + Gaussian perturbation dual encryption; dynamic weighted optimization;

[0240] Performance metrics: Success rate against selective attacks reduced by ≥60%, and the difference between the reconstructed image and the original image is invisible to the naked eye.

[0241] This invention also provides a medical image steganography system based on reversible deep networks, comprising:

[0242] One or more processors and a memory; the memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations. A reversible neural network is obtained through the following steps, including a reversible neural network encoder and a reversible neural network decoder, to implement the aforementioned medical image steganography method based on a reversible deep network:

[0243] Step S1: Receive the original medical image and the secret information to be embedded, and normalize the original medical image while performing spatial alignment.

[0244] Step S2: Construct a multi-scale feature extraction network. Input the normalized medical image into the multi-scale feature extraction network to extract the low-level texture features, mid-level edge features and high-level semantic features of the medical image, and then stitch them together in the channel dimension to form a fused feature.

[0245] Step S3: Encrypt the secret information to be embedded based on its data format;

[0246] Step S4: Construct a reversible neural network encoder, concatenate the fused features and the encrypted secret information in the channel dimension to form a joint representation, and input the joint representation into the reversible neural network encoder to generate a steganalytic image;

[0247] Step S5: Construct a reversible neural network decoder. Input the steganalyte image into the reversible neural network decoder, extract the fused features and encrypted secret information through reverse mapping, and reconstruct the original medical image and secret information respectively.

[0248] Step S6: Construct a structure-aware joint loss function, train the reversible neural network to obtain a trained reversible neural network, deploy a reversible neural network encoder at the transmitting end, and deploy a reversible neural network decoder at the receiving end to complete the encryption and decryption of medical images.

[0249] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A medical image steganography method based on reversible deep network, characterized in that, The following steps S1-S6 are performed to construct a reversible neural network, including a reversible neural network encoder and a reversible neural network decoder, to complete encryption and decryption of the medical image: Step S1: Receive the original medical image and the secret information to be embedded, and perform normalization processing on the original medical image, while performing spatial alignment; Step S2: Construct a multi-scale feature extraction network, input the normalized medical image into the multi-scale feature extraction network, respectively extract the low-level texture feature, the middle-level edge feature and the high-level semantic feature of the medical image, and splice them in the channel dimension to form a fusion feature; Step S3: For the secret information to be embedded, based on the data format of the secret information, encryption is performed; Step S4: Construct a reversible neural network encoder, splice the fusion feature and the encrypted secret information in the channel dimension to form a joint representation, and input the joint representation into the reversible neural network encoder to generate a stego image; The specific steps of step S4 are as follows: Step S4.1: encrypting the secret information obtained in step S3 fused features obtained in step S2 concatenated as joint input: ； In the formula, denotes a joint input, denotes a concatenation operation in the channel dimension; Step S4.2: constructing a reversible neural network encoder, inputting the joint input to the reversible neural network encoder, generating a stego image : ； wherein is a combination function of the bidirectional coupling structure; The reversible neural network encoder defines the stego image output in a residual manner as follows : ； In the formula, is a learned perturbation term, denotes the normalized medical image; Step S4.3: The reversible neural network has reversibility, and the reversible neural network has an inverse mapping function, so that the following formula is established: ； wherein denotes the inverse mapping function of the combined function of the bidirectional coupling structure; Extracting the encrypted secret information according to the above formula With fusion features ; Step S5: Construct a reversible neural network decoder, input the stego image into the reversible neural network decoder, extract the fusion feature and the encrypted secret information through inverse mapping, and respectively reconstruct the original medical image and the secret information; The specific steps of step S5 are as follows: Step S5.1: Construct a reversible neural network decoder, according to the reversibility of the reversible neural network, process the stego image through the inverse mapping function, as follows: ； In the formula, is an inverse mapping function processing the obtained reconstructed fusion feature mapping, is an inverse mapping function processing the obtained reconstructed secret information; is an inverse mapping function of a combination function representing a bidirectional coupling structure. Step S5.2: Utilizing the reconstruction function Converting the reconstructed fused feature map back into the original medical image form: ； wherein is a reconstruction function, denotes the reconstructed original medical image; Step S5.3: If the secret information is encrypted by AES or diffusion perturbation encryption, the following inverse process is performed: ； wherein denotes the decoded secret information, denotes the inverse process of the AES encryption; At the same time, the mean value or filtering restoration is performed on the diffusion perturbation encryption part to obtain the final secret information; Step S6: Construct a structure-aware joint loss function to train the reversible neural network, obtain the trained reversible neural network, deploy the reversible neural network encoder at the sending end and deploy the reversible neural network decoder at the receiving end to complete the encryption and decryption of the medical image.

2. The medical image steganography method based on reversible deep network according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: receiving an original medical image , with a dimension of HxW, where H and W are the height and width of the original medical image, respectively, and pixel gray value range [0, 255]; and secret information to be embedded ; Step S1.2: Normalization of the original medical image is performed, as follows: ； in, This represents the intensity value of the original medical image at (x, y). This represents the intensity value of the normalized medical image at (x, y). Represents the original medical image The lowest strength value in, Represents the original medical image The highest strength value in; Step S1.3: According to the clinical task, a preset region of interest is pre-extracted, and the mask representation of the region of interest is as follows: ； wherein represents a region of interest; represents a mask at (x, y) in the region of interest; Step S1.4: Perform spatial alignment by the following formula: ； wherein represents a PET image change function, represents the intensity value of the PET image after affine transformation and resampling to the reference coordinate at (x, y), , , , , , represents a two-dimensional affine transformation parameter.

3. The medical image steganography method based on reversible deep network according to claim 2, characterized in that, The specific steps of step S2 are as follows: Step S2.1: Extracting low-level texture features As follows: ； In the formula, denotes a low-level encoder function of the first 2-3 layers of the multi-scale feature extraction network and ReLU activation; denotes a normalized medical image; Step S2.2: Extracting middle-layer edge features as follows: ； wherein denotes a down-sampling and meso-convolution function; Step S2.3: Extracting high-level semantic features As follows: ； In the formula, denotes a residual module or a Transformer module function; Step S2.4: low-level texture features , mid-level edge features , high-level semantic features Concatenate and fuse in the channel dimension to obtain the fusion features as follows: ； In the formulae, denotes a fusion feature, denotes a stitching operation.

4. The medical image steganography method based on reversible deep network according to claim 3, characterized in that, The specific steps of step S3 are as follows: Step S3.1: The data format of the secret information includes text, label, and image; if the data format of the secret information is text or label, binary encoding form is used; if the data format of the secret information is image, pixel matrix is used directly; Step S3.2: For the secret information with text or label data format, AES encryption is performed using the Advanced Encryption Standard AES, represented as follows: ； In the formula, represents the secret information after AES encryption, represents a symmetric encryption operation based on a pre-shared key, and the same pre-shared key is used for inverse operation during decryption, represents the secret information; Step S3.3: For the secret information with image data format, diffusion perturbation encryption is performed using Gaussian perturbation, represented as follows: ； In the formula, represents the secret information after diffusion perturbation encryption; represents Gaussian noise, represents the variance of Gaussian noise; Step S3.4: In actual application, AES encryption and diffusion perturbation encryption are combined, as follows: ； In the formula, represents the secret information after the combination of AES encryption and diffusion perturbation encryption.

5. The medical image steganography method based on reversible deep network according to claim 4, characterized in that, The specific steps of step S6 are as follows: Step S6.1: Constructing the privacy loss As follows: ； wherein, denotes a balanced structural similarity function, denotes a control weight, denotes a stego image output by the reversible neural network encoder, denotes a normalized medical image; Step S6.2: Constructing reconstruction loss As follows: ； wherein denotes the reconstructed secret information output by the invertible neural network decoder, denotes the secret information output by the invertible neural network encoder, denotes the L1 distance; Step S6.3: Constructing ROI structure loss As follows: ； wherein denotes the mask at (x, y) in the region of interest; denotes the intensity value of the stego image at (x, y), denotes the intensity value of the normalized medical image at (x, y); Step S6.4: Construct a structure-aware joint loss function as follows: ； In the formula, represents a structure-aware joint loss function, , are weight coefficients of reconstruction loss , ROI structure loss , respectively.

6. A medical image steganography system based on reversible deep network, characterized by, Including: One or more processors and a memory, the memory storing instructions operable to cause the one or more processors to perform operations, by obtaining a reversible neural network, including a reversible neural network encoder, a reversible neural network decoder, implementing the reversible deep network-based medical image steganography method as claimed in any one of claims 1-5, when executed by the one or more processors: Step S1: receiving an original medical image and secret information to be embedded, and performing normalization processing on the original medical image, while performing spatial alignment; Step S2: constructing a multi-scale feature extraction network, inputting the normalized medical image into the multi-scale feature extraction network, extracting low-level texture features, middle-level edge features and high-level semantic features of the medical image respectively, and splicing the features in the channel dimension to form a fusion feature; Step S3: encrypting the secret information to be embedded based on the data format of the secret information; Step S4: constructing a reversible neural network encoder, splicing the fusion feature and the encrypted secret information in the channel dimension to form a joint representation, and inputting the joint representation into the reversible neural network encoder to generate a stego image; Step S5: constructing a reversible neural network decoder, inputting the stego image into the reversible neural network decoder, extracting the fusion feature and the encrypted secret information through inverse mapping, and respectively reconstructing the original medical image and the secret information; Step S6: constructing a structure-aware joint loss function, training the reversible neural network to obtain a trained reversible neural network, deploying the reversible neural network encoder at the sending end and deploying the reversible neural network decoder at the receiving end, and completing the encryption and decryption of the medical image.

Images