Methods, systems, equipment, and media for enhancing medical images and training models.

Abstract

This invention relates to a method, system, device, and medium for enhancing medical images and training a model. The training method includes: acquiring non-contrast-enhanced medical images and their paired contrast-enhanced images; inputting the non-contrast-enhanced medical images into a single-branch generator to extract image features and generate an initial virtual enhanced image; inputting the initial virtual enhanced image and the contrast-enhanced image into a registration network to align the initial virtual enhanced image to the coordinate space of the contrast-enhanced image, generating an aligned virtual enhanced image; obtaining the discriminant values ​​of the real image and the virtual image through a discriminator; calculating the adversarial loss between the real image discriminant value and the real label, the virtual image discriminant value and the virtual label, and the correlation loss between the aligned virtual enhanced image and the contrast-enhanced image; updating the single-branch generator based on the adversarial loss and the correlation loss to obtain a medical image enhancement model. This invention can generate high-quality virtual enhanced images without the need for contrast agents.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method, system, device, and medium for enhancing medical images and training models. Background Technology

[0002] Traditional medical image enhancement relies on contrast agents to improve image contrast and enhance lesion visibility. However, the use of contrast agents carries certain safety risks. For example, pregnant women or patients allergic to contrast agents cannot tolerate them, thus preventing contrast-enhanced imaging. Furthermore, contrast-enhanced imaging requires additional scanning time and contrast agent injection procedures, increasing examination costs, prolonging patient wait times, and placing a greater burden on hospital operations. Existing technologies typically utilize deep learning models to generate virtual enhanced images to replace contrast agents. However, current methods for generating virtual enhanced images suffer from high computational demands, cross-modal image misalignment, and unstable image quality, severely limiting the clinical application of this technology. Therefore, there is a need for a method, system, device, and medium for medical image enhancement and model training. Summary of the Invention

[0003] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a method, system, device and medium for enhancing medical images and training models, which improves the problem of poor image accuracy of virtual enhanced images generated by the prior art.

[0004] To achieve the above and other related objectives, this invention provides a training method for a medical image enhancement model. The training method includes: acquiring non-contrast-enhanced medical images and their paired contrast-enhanced images; wherein the non-contrast-enhanced medical images are acquired using a contrast-free medical imaging method, and the contrast-enhanced images are contrast-enhanced versions of the corresponding medical images, and the contrast-enhanced images have corresponding preset real labels; inputting the non-contrast-enhanced medical images into a single-branch generator, extracting the image features of the non-contrast-enhanced medical images, and generating an initial virtual enhanced image; wherein the initial virtual enhanced image has corresponding preset virtual labels; and inputting the initial virtual enhanced image and the corresponding contrast-enhanced image into... The image is fed into a registration network to align the initial virtual augmented image to the coordinate space of the corresponding contrast-enhanced image, generating an aligned virtual augmented image. A discriminator is used to obtain the real image discrimination value corresponding to the contrast-enhanced image and the virtual image discrimination value corresponding to the initial virtual augmented image. Adversarial losses between the real image discrimination value and its corresponding real label, and between the virtual image discrimination value and its corresponding virtual label, are calculated. Correlation losses between the aligned virtual augmented image and the contrast-enhanced image are also calculated. Based on these adversarial and correlation losses, a single-branch generator is updated to obtain a trained medical image enhancement model. The medical image enhancement model is a single-branch generator.

[0005] In one embodiment of the present invention, the non-contrast-treated medical image includes at least two modalities of images. The step of inputting the non-contrast-treated medical image into a single-branch generator, extracting the image features of the non-contrast-treated medical image, and generating an initial virtual enhanced image includes: stitching the images of each modality together, inputting the stitched image into the single-branch generator, extracting the image features of the stitched image, and generating an initial virtual enhanced image corresponding to the non-contrast-treated medical image.

[0006] In one embodiment of the present invention, the step of inputting an initial virtual enhanced image and a corresponding contrast enhanced image into a registration network, aligning the initial virtual enhanced image to the coordinate space of the corresponding contrast enhanced image, and generating an aligned virtual enhanced image includes: inputting the contrast enhanced image into a first feature extraction module of the registration network to extract contrast image features; inputting the initial virtual enhanced image into a second feature extraction module of the registration network to extract virtual enhanced image features; inputting the contrast image features and the virtual enhanced image features into a similarity calculation module of the registration network to calculate the similarity error between the virtual enhanced image features and the contrast image features; inputting the similarity error into a deformation field generation module of the registration network to generate a deformation field based on the similarity error; wherein the deformation field is used to characterize the offset of the initial virtual enhanced image relative to the contrast enhanced image; and inputting the deformation field into a resampling module of the registration network to resample the initial virtual enhanced image according to the deformation field to generate an aligned virtual enhanced image.

[0007] In one embodiment of the present invention, the step of resampling the initial virtual augmented image based on the deformation field to generate an aligned virtual augmented image includes: for each pixel of the initial virtual augmented image: extracting the offset corresponding to the pixel in the deformation field, and determining whether the offset is an integer; if so, correcting the coordinates of the pixel based on the offset to obtain the corrected coordinates of the pixel; otherwise, interpolating the offset based on interpolation, and using the interpolated offset to correct the coordinates of the pixel to obtain the corrected coordinates of the pixel; and forming an aligned virtual augmented image based on the corrected coordinates of all pixels and their corresponding pixel values.

[0008] In one embodiment of the present invention, the discriminator is a local perception discriminator. The discriminator obtains the real image discrimination value corresponding to the contrast-enhanced image, including: inputting the contrast-enhanced image into the cropping module of the local perception discriminator to cut the contrast-enhanced image and generate multiple local contrast-enhanced images; inputting the multiple local contrast-enhanced images into the discrimination module of the local perception discriminator to discriminate each local contrast-enhanced image and generate the real image discrimination value.

[0009] In one embodiment of the present invention, the discriminator is a local perception discriminator. The discriminator obtains the virtual image discrimination value corresponding to the initial virtual augmented image, including: inputting the initial virtual augmented image into the cropping module of the local perception discriminator to cut the initial virtual augmented image to generate multiple local virtual augmented images; inputting the multiple local virtual augmented images into the discrimination module of the local perception discriminator to discriminate each local virtual augmented image and generate a virtual image discrimination value.

[0010] In one embodiment of the present invention, the step of calculating the adversarial loss between the real image discriminant value and its corresponding real label, the virtual image discriminant value and its corresponding virtual label, and calculating the correlation loss between the aligned virtual enhanced image and the contrast enhanced image, and updating the single-branch generator based on the adversarial loss and the correlation loss to obtain a trained medical image enhancement model includes: calculating the adversarial loss between the real image discriminant value and its corresponding real label, the virtual image discriminant value and its corresponding virtual label, and updating the single-branch generator and the discriminator based on the adversarial loss; calculating the correlation loss between the aligned virtual enhanced image and the contrast enhanced image, and updating the single-branch generator and the registration network based on the correlation loss.

[0011] In one embodiment of the present invention, a method for enhancing medical images is also provided. The enhancement method includes: acquiring a non-contrast-treated medical image; inputting the non-contrast-treated medical image into a medical image enhancement model to generate a virtual enhanced image; wherein the medical image enhancement model is obtained by training the medical image enhancement model according to any one of the above-described methods, and the medical image enhancement model generates the virtual enhanced image by aligning the non-contrast-treated medical image to a preset coordinate space and extracting image features from the aligned non-contrast-treated medical image.

[0012] In one embodiment of the present invention, a training system for a medical image enhancement model is also provided. The system includes: a data acquisition module for acquiring non-contrast-enhanced medical images and their paired contrast-enhanced images; wherein the non-contrast-enhanced medical images are acquired using a contrast-free medical imaging method, and the contrast-enhanced images are contrast-enhanced versions of the corresponding medical images, and the contrast-enhanced images have corresponding preset real labels; a virtual enhanced image generation module for inputting the non-contrast-enhanced medical images into a single-branch generator, extracting the image features of the non-contrast-enhanced medical images, and generating an initial virtual enhanced image; wherein the initial virtual enhanced image has corresponding preset virtual labels; and a registration module for registering the initial virtual enhanced image and the corresponding contrast-enhanced image. The enhanced image is input into the registration network, which aligns the initial virtual enhanced image to the coordinate space of the corresponding contrast enhanced image, generating the aligned virtual enhanced image. A discrimination module is used to obtain the real image discrimination value corresponding to the contrast enhanced image and the virtual image discrimination value corresponding to the initial virtual enhanced image through a discriminator. A model training module is used to calculate the adversarial loss between the real image discrimination value and its corresponding real label, the virtual image discrimination value and its corresponding virtual label, and the correlation loss between the aligned virtual enhanced image and the contrast enhanced image. Based on the adversarial loss and the correlation loss, the single-branch generator is updated to obtain the trained medical image enhancement model. The medical image enhancement model is a single-branch generator.

[0013] In one embodiment of the present invention, a medical imaging device is also provided, the medical imaging device comprising: an imaging module for capturing non-contrast medical images of the object to be examined; and a processing module for inputting the non-contrast medical images into a medical image enhancement model to generate virtual enhanced images; wherein the medical image enhancement model generates virtual enhanced images by aligning the non-contrast medical images to a preset coordinate space and extracting image features from the aligned non-contrast medical images.

[0014] In one embodiment of the present invention, an electronic device is also provided, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device enables the training method of the medical image enhancement model or the medical image enhancement method described above.

[0015] In one embodiment of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a computer processor, the computer performs the training method of the medical image enhancement model or the medical image enhancement method described in any of the preceding claims.

[0016] As described above, the medical image enhancement and model training method, system, device, and medium of the present invention have the following beneficial effects: Compared with traditional multi-branch generative networks, the single-branch generative architecture of this scheme significantly reduces computational complexity and improves model training and inference efficiency. Furthermore, by aligning the initial virtual enhanced image with the corresponding real contrast-enhanced image through a registration network, the spatial consistency of the images is improved, greatly reducing the inter-slice misalignment problem caused by patient movement or differences in scanning timing. During model training, the generator is optimized by combining adversarial loss and correlation loss, making the generated virtual enhanced image closer to the real contrast-enhanced image while maintaining the integrity of the anatomical structure. The trained medical image enhancement model can generate high-quality virtual enhanced images without contrast agents, thus providing a safer and more efficient solution for clinical medical image analysis. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a training method for a medical image enhancement model provided in an embodiment of the present invention;

[0018] Figure 2 A schematic diagram of the overall process of training a medical image enhancement model provided in an embodiment of the present invention;

[0019] Figure 3 A schematic flowchart of a medical image enhancement method provided in an embodiment of the present invention;

[0020] Figure 4 A schematic diagram of the overall process of the medical image enhancement method provided in the embodiments of the present invention;

[0021] Figure 5 The diagram shown is a structural block diagram of a training system for a medical image enhancement model provided in an embodiment of the present invention.

[0022] Figure 6 The diagram shown is a structural schematic of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0024] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0026] The inventors discovered that traditional medical image enhancement relies on contrast agents to improve image contrast and enhance lesion visibility. However, the use of contrast agents not only increases examination costs and scan time but may also pose safety risks to certain patient groups. For example, in late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) imaging, gadolinium-based contrast agents (GBCAs) are required during imaging to highlight fibrosis, infarction, or other lesion areas. However, the use of GBCAs in LGE-CMR imaging has several drawbacks: Firstly, GBCAs can cause serious adverse reactions in patients with severe renal failure or those allergic to them, thus their use is warned or contraindicated, limiting the applicability of LGE-CMR imaging in these populations. Secondly, the use of GBCAs increases the cost of CMR imaging examinations and significantly prolongs scan time due to the need for prior patient preparation, contrast agent injection, and monitoring for potential adverse reactions. In addition, the use of GBCA requires injection by professionals, which increases labor costs in resource-limited medical environments.

[0027] Despite the numerous challenges of contrast-enhanced imaging, very few published documents or patents have attempted to generate Virtual Native Enhancement (VNE) images without using GBCA, primarily due to the novelty and complexity of the technology. Existing methods typically employ a multi-branch architecture, transforming multiple different contrast-free CMR modalities (such as T1-weighted, T2-weighted, or Cine MRI sequences) to the target enhancement modality. Common architectures often use multiple independent U-Net branches to extract information from different modalities and integrate this information at the feature and pixel levels using a fusion network to synthesize a virtual enhanced image that resembles the appearance of LGE images. However, this approach fails to adequately address the problem of intra-layer misalignment. In practical medical applications, patient movement can cause misalignment between modalities, significantly reducing model performance and affecting the quality and reliability of the generated virtual enhanced images. Furthermore, the discriminators in existing methods typically only output single binary data (e.g., real / fake) for judgment, which is too coarse for handling quality variations in different regions of the generated image. If the discriminator cannot accurately assess the quality of local regions, the texture fidelity of the generated image will decrease, resulting in poor texture details in certain areas. Furthermore, the multi-branch architecture has high computational complexity, affecting generation efficiency and significantly increasing computational costs. Moreover, due to the lack of a dedicated module for handling inter-layer misalignment, existing methods still suffer from significant shortcomings in the stability and reliability of generated virtual augmented images, leading to poor image quality.

[0028] To address the aforementioned issues, this invention provides a training method for a medical image enhancement model. Compared to traditional multi-branch generative networks, the single-branch generative architecture of this method significantly reduces computational complexity and improves model training and inference efficiency. Furthermore, a registration network aligns the initial virtual enhanced image with the corresponding real contrast-enhanced image, improving spatial consistency and greatly reducing inter-slice misalignment caused by patient movement or differences in scanning timing. During model training, the generator is optimized using adversarial and correlation losses to make the generated virtual enhanced image closer to the real contrast-enhanced image while maintaining the integrity of anatomical structures. The trained medical image enhancement model can generate high-quality virtual enhanced images without contrast agents, thus providing a safer and more efficient solution for clinical medical image analysis.

[0029] Please see Figure 1 The segmentation method for medical images includes the following steps:

[0030] S101. Acquire non-contrast medical images and their paired contrast-enhanced images; wherein, the non-contrast medical images are acquired using a contrast-free medical imaging method, the contrast-enhanced images are the contrast-enhanced versions of the corresponding medical images, and the contrast-enhanced images have corresponding preset real labels.

[0031] Contrast-free medical images refer to those acquired through conventional contrast-free medical imaging methods, such as contrast-free MRI (Magnetic Resonance Imaging), CT (Computed Tomography), or ultrasound. Contrast-enhanced images are obtained under the same scanning conditions with the addition of a contrast agent, providing clearer contrast between tissues and lesions. Further, the contrast-free medical images and contrast-enhanced images of this invention refer to patient image data acquired through medical imaging equipment such as CT or MRI. After acquiring the original contrast-free medical images or contrast-enhanced images, preprocessing is required to adapt them to subsequent model processing, resulting in standardized contrast-free medical images and standardized contrast-enhanced images. Preprocessing includes, but is not limited to, grayscale normalization, noise removal from the contrast-free medical images or contrast-enhanced images, and contrast enhancement. Those skilled in the art can adaptively select the corresponding preprocessing method based on the needs of the actual application scenario, and no limitation is made here.

[0032] S102. Input the non-contrast medical image into the single-branch generator to generate the initial virtual enhanced image; wherein the initial virtual enhanced image has a corresponding preset virtual label.

[0033] The acquired non-contrast-enhanced medical images are input into a single-branch generator. This generator automatically extracts and learns the spatial and textural information of the images through a series of convolution and feature fusion operations, generating an initial virtual augmented image corresponding to the non-contrast-enhanced medical image. This image visually simulates the contrast enhancement effect after contrast agent application. Virtual labels refer to the real / false labeling of the initial virtual augmented image. For example, the initial virtual augmented image can be divided into a preset number (e.g., n) of local virtual augmented images, each with a corresponding virtual label. This virtual label can be set to a first value (e.g., 0), indicating that the image is a generated virtual augmented image, not a real contrast-enhanced image. Similarly, for each initial virtual augmented image, its matching contrast-enhanced image is labeled with a real label, which refers to the real / false labeling of the contrast-enhanced image. For example, the contrast-enhanced image can be divided into the same number (e.g., n) of local contrast-enhanced images as the initial virtual augmented image, each with a corresponding real label. This real label can be set to a second value (e.g., 1), indicating that the image is a real contrast-enhanced image. These annotations are used to subsequently train the single-branch generator and discriminator to optimize the image generation capabilities of the medical image enhancement model, thereby generating more accurate virtual enhanced images.

[0034] In one embodiment of the present invention, the non-contrast-enhanced medical image includes images of at least two modalities. The non-contrast-enhanced medical image is input into a single-branch generator to generate an initial virtual enhanced image. This process includes: stitching together the images of each modality, and inputting the stitched image into the single-branch generator to generate the initial virtual enhanced image corresponding to the non-contrast-enhanced medical image. Considering that non-contrast-enhanced medical images have different modalities, in order to fully utilize multimodal medical images (e.g., simultaneously containing T1 mapping, T2 mapping, Cine, etc.), these images of different modalities are first stitched together on their corresponding channels to generate a multi-channel input image integrating information from each modality. This stitched multi-channel image is then input into the single-branch generator. The generator automatically learns the correlation between each modality through a series of convolution and feature extraction operations, thereby generating an initial virtual enhanced image that is visually similar to a contrast-enhanced image. This novel single-branch generator proposed in this invention replaces the existing multi-branch architecture. This single-branch generator can receive multiple non-contrast-enhanced multimodal images and efficiently synthesize virtual enhanced images. By integrating feature extraction and generation processes into a unified network, the complexity of data processing is significantly reduced, and processing efficiency is improved, making the trained medical image enhancement model more suitable for practical applications in clinical settings. This invention employs a lightweight single-branch generator architecture, optimizing the training process and significantly reducing computational overhead. This makes the single-branch generator faster and requires fewer computational resources, making it more efficient than traditional multi-branch methods.

[0035] S103. Input the initial virtual augmented image and the corresponding contrast augmented image into the registration network, and align the corresponding initial virtual augmented image according to the contrast augmented image to generate the aligned virtual augmented image.

[0036] A major challenge in generating virtual enhanced images is slice misalignment caused by patient movement. To address this, this invention proposes a registration network that preprocesses the initial virtual enhanced image through elastic alignment. Specifically, the initial virtual enhanced image and its corresponding real contrast-enhanced image are input into the registration network. The registration network calculates the spatial difference between the virtual and contrast-enhanced images to generate a deformation field for alignment. This field is used to transform and resample the initial virtual enhanced image to obtain the aligned virtual enhanced image. This alignment process effectively eliminates slice misalignment caused by patient movement, respiration, or differences in scanning timing. This approach enhances the alignment and consistency of the generated virtual enhanced images, making the system more robust when handling changes in patient anatomy and imaging conditions.

[0037] In one embodiment of the present invention, the step of inputting an initial virtual augmented image and a corresponding contrast-enhanced image into a registration network, aligning the corresponding initial virtual augmented image according to the contrast-enhanced image, and generating an aligned virtual augmented image includes:

[0038] The contrast-enhanced image is input into the first feature extraction module of the registration network to extract the contrast image features of the contrast-enhanced image;

[0039] The initial virtual augmented image is input into the second feature extraction module of the registration network to extract the virtual augmented image features of the initial virtual augmented image;

[0040] The contrast image features and the virtual enhanced image features are input into the similarity calculation module of the registration network to calculate the similarity error between the virtual enhanced image features and the contrast image features;

[0041] The similarity error is input into the deformation field generation module of the registration network to generate a deformation field based on the similarity error; wherein, the deformation field is used to characterize the offset of the initial virtual enhanced image relative to the contrast enhanced image;

[0042] The deformation field is input into the resampling module of the registration network, and the initial virtual augmented image is resampled according to the deformation field to generate an aligned virtual augmented image.

[0043] The initial virtual augmented image and its corresponding real contrast-enhanced image are input into the registration network. A first feature extraction module and a second feature extraction module extract the contrast-enhanced image features of the contrast-enhanced image and the virtual augmented image features of the initial virtual augmented image, respectively. The parameters of the first and second feature extraction modules can be the same or different, depending on the structural adaptability of the feature extraction modules. The first / second feature extraction modules can be convolutional neural networks or deep neural networks, etc., without limitation. A similarity calculation module calculates a similarity error based on the similarity between the two images in the feature space to quantify the degree of consistency in spatial structure between the initial virtual augmented image and the contrast-enhanced image. The similarity calculation module can be a mean squared error calculation layer or a similarity calculation layer after a convolutional layer, etc., without limitation. Using a deformation field generation module, the registration network learns and generates a deformation field based on the similarity error. This deformation field indicates the direction in which the initial virtual augmented image needs to move relative to the contrast-enhanced image at each pixel coordinate. The calculated deformation field is then applied to the initial virtual augmented image, and resampling is performed to generate the aligned virtual augmented image. The deformation field generation module can be a U-net structure or a Transformer structure. Specifically, the registration network maps the corresponding pixels in the initial virtual enhanced image to new coordinate positions based on the positional offset indicated by the deformation field, and uses the resampling module to calculate the pixel values ​​at these new coordinates using interpolation methods. Through this process, the virtual enhanced image can be made to spatially overlap with the real contrast-enhanced image as much as possible, thus achieving alignment. This ensures that the aligned virtual enhanced image is spatially consistent with the patient's anatomical structures.

[0044] In one embodiment of the present invention, the step of resampling the initial virtual augmented image based on the deformation field to generate an aligned virtual augmented image includes:

[0045] For each pixel of the initial virtual augmented image:

[0046] Extract the offset of the pixel in the deformation field and determine whether the offset is an integer:

[0047] If so, the coordinates of the pixel are corrected based on the offset to obtain the corrected coordinates of the pixel;

[0048] Otherwise, the offset is interpolated using the interpolation method, and the coordinates of the pixel are corrected using the interpolated offset to obtain the corrected coordinates of the pixel.

[0049] Based on the coordinates of all corrected pixels and their corresponding pixel values, an aligned virtual augmented image is formed.

[0050] For each pixel in the initial virtual augmented image, its offset in the deformation field is read. If both the horizontal and vertical components of this offset are integers, the pixel's coordinates are directly corrected according to these integer values ​​to obtain the corrected coordinate position. Conversely, if the offset is not an integer, the interpolated offset is calculated using bilinear or bicubic interpolation algorithms before updating the coordinates. This method ensures that the adjustment of each pixel's position more accurately matches the spatial structure of the real contrast-enhanced image, avoiding misalignment or information loss caused by simple integer translation. Based on the corrected coordinates and original pixel values ​​of all pixels, the aligned virtual augmented image is reconstructed using this method. This ensures that the virtual augmented image remains as consistent as possible with the real contrast-enhanced image while maintaining its detail.

[0051] S104. The discriminator obtains the real image discrimination value corresponding to the contrast-enhanced image and the virtual image discrimination value corresponding to the initial virtual enhanced image.

[0052] A real contrast-enhanced image and an initial virtual enhanced image are input into a discriminator. For each input, the discriminator learns the texture, edge, and brightness distribution features of the contrast-enhanced image locally or globally, outputting a numerical evaluation result for the contrast-enhanced image. Similarly, the discriminator learns the texture, edge, and brightness distribution features of the initial virtual enhanced image locally or globally, outputting a numerical evaluation result for the initial virtual enhanced image. Specifically, the discriminator's evaluation result for the real contrast-enhanced image is the real image discrimination value, and its evaluation result for the initial virtual enhanced image is the virtual image discrimination value. These two values ​​allow for a quantitative comparison of the differences between the real and initial virtual enhanced images, and are used to adjust the generator's weights during the adversarial learning phase, making the generated virtual enhanced image more closely resemble the real contrast-enhanced image.

[0053] In one embodiment of the present invention, the discriminator is a local perception discriminator, which obtains the real image discrimination value corresponding to the contrast-enhanced image, including:

[0054] The contrast-enhanced image is input into the cropping module of the local perception discriminator, and the contrast-enhanced image is cut to generate multiple local contrast-enhanced images;

[0055] Multiple local contrast-enhanced images are input into the discrimination module of the local perception discriminator to discriminate each local contrast-enhanced image and generate a real image discrimination value.

[0056] Considering that the quality of virtual enhanced images is crucial to diagnostic accuracy, this invention proposes a locally perceptive fully convolutional discriminator to further improve the quality of virtual enhanced images. This discriminator evaluates local regions rather than making a single judgment on the entire image. Specifically, a cropping module divides the entire contrast-enhanced image into multiple local contrast-enhanced images according to a certain size and stride. For example, a 256×256 image is segmented into several 64×64 or 32×32 feature maps, each of which is a local contrast-enhanced image. The cropping module can be configured to use adaptive region segmentation or fixed-size cropping, etc. The discriminator uses a discrimination module to judge each local contrast-enhanced image separately, utilizing its internal convolutional neural network to analyze local brightness, texture, edge details, etc., and outputs a local real image discrimination value to characterize the realism of the region. The discrimination module can be a convolutional neural network, a Transformer-based local discriminator, a GAN-based adversarial discriminator, etc., and is not limited here. After discriminating all local enhanced images, the discriminant values ​​of these local ground truth images are arranged sequentially according to row and column order based on the spatial position of each local enhanced image in the original contrast-enhanced image, forming a matrix corresponding to the size of the original contrast-enhanced image, and this matrix is ​​used as the ground truth image discriminant value. Furthermore, the generated local ground truth image discriminant values ​​can be weighted or averaged to determine the ground truth image discriminant value. Preferably, to achieve fine-grained observation of the real contrast-enhanced image, the matrix formed by the discriminant values ​​of each local ground truth image can be used as the ground truth image discriminant value. This method enables more refined judgment than single global discrimination, providing richer local information for subsequent model training and image quality assessment, improving the realism of the synthesized image and reducing artifacts. By using a locally aware discriminator, even the smallest details are accurately captured, improving the robustness of the entire image generation process.

[0057] In one embodiment of the present invention, the discriminator is a local perception discriminator. The discriminator obtains a virtual image discrimination value corresponding to the initial virtual augmented image, including:

[0058] The initial virtual augmented image is input into the cropping module of the local perception discriminator, which cuts the initial virtual augmented image to generate multiple local virtual augmented images;

[0059] Multiple local virtual augmented images are input into the discrimination module of the local perception discriminator to discriminate each local virtual augmented image and generate virtual image discrimination values.

[0060] The initial virtual augmented image is divided into multiple local virtual augmented images according to a certain size and stride by a cropping module. For example, a 256×256 image is divided into several 64×64 or 32×32 feature maps, each of which is a local virtual augmented image. The cropping module can be configured for adaptive region segmentation or fixed-size cropping. The discriminator uses a discrimination module to classify each local virtual augmented image, analyzing its brightness, texture, edges, and other details using an internal convolutional neural network, and outputting a local virtual image discrimination value to characterize the realism of the region. The discrimination module can be a convolutional neural network, a Transformer-based local discriminator, or a GAN-based adversarial discriminator, etc., without limitation. After classifying all local virtual augmented images, the discrimination values ​​are arranged sequentially according to the position of each local virtual augmented image within the initial virtual augmented image, forming a matrix corresponding to the size of the initial virtual augmented image, and this matrix is ​​used as the virtual image discrimination value. Furthermore, the generated local virtual image discriminant values ​​can be weighted or averaged to determine the virtual image discriminant value. Preferably, to achieve fine-grained observation of the virtual enhanced image, the matrix formed by the local virtual image discriminant values ​​can be used as the virtual image discriminant value. This method enables more refined judgment than single global discrimination, providing richer local information for subsequent model training and image quality assessment, improving the realism of the synthesized image and reducing artifacts. By using a locally aware discriminator, even the smallest details are accurately captured, improving the robustness of the entire image generation process.

[0061] S105. Calculate the adversarial loss between the real image discriminant value and the corresponding real label, the virtual image discriminant value and the corresponding virtual label, and calculate the correlation loss between the aligned virtual enhanced image and the contrast enhanced image. Update the single-branch generator based on the adversarial loss and the correlation loss to obtain the trained medical image enhancement model; wherein, the medical image enhancement model is a single-branch generator.

[0062] The model's performance in contrast and texture detail is measured by calculating the adversarial loss between the discriminator's output of real image discrimination values ​​and their corresponding real labels, and between virtual image discrimination values ​​and their corresponding virtual labels. The similarity between the aligned virtual enhanced image and the real contrast-enhanced image is measured by calculating the correlation loss. The single-branch generator and discriminator are iteratively trained several times based on the adversarial loss, and then iteratively trained several times based on the correlation loss. This iterative alternating training process ultimately generates a well-trained medical image enhancement model.

[0063] In one embodiment of the present invention, the calculation of adversarial loss between the real image discriminant value and its corresponding real label, and between the virtual image discriminant value and its corresponding virtual label, and the calculation of correlation loss between the aligned virtual enhanced image and the contrast-enhanced image, and the updating of the single-branch generator based on the adversarial loss and correlation loss to obtain the trained medical image enhancement model, includes:

[0064] Calculate the adversarial loss between the real image discrimination value and its corresponding real label, and between the virtual image discrimination value and its corresponding virtual label, and update the single-branch generator and discriminator based on the adversarial loss;

[0065] The correlation loss between the aligned virtual augmented image and the contrast augmented image is calculated, and the single-branch generator and registration network are updated based on the calculated difference.

[0066] The discriminator outputs the real image discrimination value and the virtual image discrimination value respectively. The real label and the virtual label are the target signals used to guide the discriminator to learn the features of real contrast-enhanced images and non-contrast images. By calculating the adversarial loss between them, the model can determine the similarity between the generated virtual enhanced image and the real contrast-enhanced image and the degree of preservation of non-contrast information. Specifically, based on the adversarial loss of the real image discrimination value and the corresponding real label, and the virtual image discrimination value and the corresponding virtual label, the adversarial loss can be calculated according to formula (1):

[0067]

[0068] Where G represents a single-branch generator, D represents a discriminator, and x represents the input medical image. Indicates contrast-enhanced image, The probability of the discriminator assigning a real contrast-enhanced image is obtained from the real labels. D(G(x)) represents the probability of the discriminator assigning a generated virtual enhanced image, obtained from the virtual labels.

[0069] The single-branch generator and registration network can be updated based on the correlation loss between the aligned virtual augmented image and the contrast-enhanced image. The correlation loss is shown in Equation (2):

[0070]

[0071] Where G represents a single-branch generator, and x represents the input medical image. This represents a contrast-enhanced image, where R represents the registration network and its output deformation field. This represents the resampling process. The generator and registration network are trained by continuously minimizing the correlation loss. It is understood that the input medical images can be multimodal, and to adapt to the model's input requirements, the input medical images can be cropped, for example, the region to be analyzed, such as the heart, can be cropped from the medical image. The cropped medical image is then input into a single-branch generator for analysis, where the correlation loss includes, but is not limited to, L1 loss, L2 loss, or structural similarity loss.

[0072] Furthermore, considering that each patient's cardiovascular structure is unique, this invention employs a patient-specific image generation mechanism. During model training and inference phases, slices from different modalities of the same patient are selected, ensuring that the interslice spacing is within a preset threshold to guarantee the matching of paired data. This approach allows for better capture and reproduction of the patient's unique anatomical structures, improving the clinical relevance and accuracy of the synthesized virtual enhanced images. This ensures that the generated virtual enhanced images are structurally consistent with the actual contrast-enhanced images. This method is applicable to various clinical scenarios, especially for patient groups who cannot use contrast agents, thereby ensuring the availability of personalized imaging data for diagnosis and treatment planning.

[0073] like Figure 2As shown, taking contrast-free CMR modalities as an example, the training process of the medical image enhancement model described in this invention is as follows: Multimodal medical images, exemplarily such as T1 maps (T1map), T2 maps (T2map), and dynamic images (Cine), are cropped to obtain corresponding cropped images. For example, the cropped images of the multimodal medical images generate cropped T1-maps, cropped T2-maps, and cropped Cine t1. To perform supervised learning and train a single-branch generator, the cropped images can be concatted and input into the single-branch generator to generate an initial virtual enhanced image (e.g., VNE). The contrast-enhanced image (e.g., LGE) is cropped to obtain a cropped contrast-enhanced image (e.g., Cropped LGE), which, along with the initial virtual enhanced image, is input into a panch-based discriminator to generate an adversarial loss. This adversarial loss is then used to train the generator and discriminator. Furthermore, the initial virtual enhanced image is input into a registration network for enhanced alignment to generate an aligned virtual enhanced image (e.g., Aligned VNE). The correlation loss between this aligned and cropped contrast-enhanced image is calculated to update the registration network and generator. This iterative training process yields a well-trained medical image enhancement model.

[0074] Please see Figure 3 The present invention also provides a method for enhancing medical images, the enhancement method comprising:

[0075] S301. Acquire non-contrast-based medical images;

[0076] S302. Input the non-contrast medical image into the medical image enhancement model to generate a virtual enhanced image; wherein, the medical image enhancement model is obtained by the training method of the medical image enhancement model described above, and the medical image enhancement model generates a virtual enhanced image by aligning the non-contrast medical image to a preset coordinate space and extracting image features from the aligned non-contrast medical image.

[0077] like Figure 4As shown, taking contrast-free CMR modalities as an example, the processing procedure in the inference stage of this invention is as follows: Non-contrast-free medical images are cropped, such as T1 maps, T2 maps, and dynamic images (Cine), generating cropped T1-maps, cropped T2-maps, and cropped Cine t1 images. In the inference stage (Generation process), the cropped images are concatenated and input into a trained medical image enhancement model (Single-branch generator) to generate virtual enhanced images (such as VNE). These can then be used for subsequent clinical analysis and diagnosis.

[0078] This invention also provides a medical imaging device, comprising: an imaging module for capturing non-contrast medical images of the subject; and a processing module for inputting the non-contrast medical images into a medical image enhancement model to generate virtual enhanced images. The medical image enhancement model generates virtual enhanced images by aligning the non-contrast medical images to a preset coordinate space and extracting image features from the aligned non-contrast medical images. The imaging module acquires non-contrast medical images of the patient, such as contrast-free MRI or CT images. The processing module uses the medical image enhancement model to process the acquired images and generate virtual enhanced images. The medical image enhancement model aligns the input non-contrast medical images to a preset coordinate space to ensure spatial consistency between different images. By extracting the aligned image features and generating virtual enhanced images with similar contrast enhancement effects based on these features, patients can obtain high-quality enhanced images without the use of contrast agents. It is understood that the medical imaging equipment can be any existing medical image acquisition equipment, including but not limited to X-ray machines, CT scanners, MRI machines, ultrasound imaging equipment, etc., as long as the imaging equipment is suitable for medical image acquisition, and is not limited here. The processing module can be implemented based on the computing unit in the existing medical imaging equipment, or integrated into the equipment as an independent image processing unit, and the specific implementation method is not limited. As long as the processing module has the functions described in this invention, that is, on the basis of existing image processing functions, further integrating a medical image enhancement model, and being able to align, extract features, and generate virtual enhanced images for non-contrast-treated medical images, the medical imaging equipment having the above-mentioned processing module can be considered to fall within the protection scope of this invention.

[0079] In summary, the novel algorithm proposed in this invention is more robust to virtual enhanced image generation than existing methods, especially in maintaining high-quality image synthesis even with patient movement. The quality of the generated virtual enhanced images surpasses that of contrast-enhanced images obtained from accelerated or even normal scans. This enhanced robustness ensures that even with poor-quality input medical images, the generated virtual enhanced images maintain high quality, thereby improving the accuracy of clinical assessment and reducing the need for repeated scans. This not only saves time and costs but also improves the reliability of image diagnosis. Furthermore, the single-branch generator used in this invention significantly reduces computational requirements compared to existing methods, substantially reducing GPU memory usage and computation time. Since rapid imaging and real-time analysis are crucial for medical efficiency, this streamlined architecture accelerates image processing, which is particularly advantageous in clinical settings. Moreover, reduced computational requirements lead to lower energy consumption and less heat dissipation, making the entire medical image enhancement model more sustainable and reducing overall operating costs. Additionally, the method has low hardware requirements, enabling its widespread application in medical institutions with limited computing resources. The AI ​​algorithm of this invention covers a wide range of disease types and has broad applicability in clinical practice. For example, this method can generate patient-specific virtual enhanced images for various diseases such as cardiomyopathy, myocardial infarction, aortic valve disease, and myocardial amyloidosis. This makes the algorithm highly practical in multiple medical fields such as pathology and radiology, thereby enhancing its application value in patient diagnosis and disease management.

[0080] Please see Figure 5The training system 500 for the medical image enhancement model includes: a data acquisition module 510, a virtual enhanced image generation module 520, a registration module 530, a discrimination module 540, and a model training module 550. The data acquisition module 510 acquires non-contrast-enhanced medical images and their paired contrast-enhanced images; wherein the non-contrast-enhanced medical images are acquired using a contrast-free medical imaging method, and the contrast-enhanced images are contrast-enhanced versions of the corresponding medical images, each with a corresponding preset real-world label. The virtual enhanced image generation module 520 inputs the non-contrast-enhanced medical images into a single-branch generator, extracts the image features of the non-contrast-enhanced medical images, and generates an initial virtual enhanced image; wherein the initial virtual enhanced image has a corresponding preset virtual label. The registration module 530 inputs the initial virtual enhanced image and the corresponding contrast-enhanced image into a registration network, aligns the initial virtual enhanced image to the coordinate space of the corresponding contrast-enhanced image, and generates an aligned virtual enhanced image. The discrimination module 540 is used to obtain the real image discrimination value corresponding to the contrast-enhanced image and the virtual image discrimination value corresponding to the initial virtual enhanced image through a discriminator. The model training module 550 is used to calculate the adversarial loss between the real image discrimination value and the corresponding real label, the virtual image discrimination value and the corresponding virtual label, and the correlation loss between the aligned virtual enhanced image and the contrast-enhanced image, and update the single-branch generator based on the adversarial loss and the correlation loss to obtain the trained medical image enhancement model; wherein, the medical image enhancement model is a single-branch generator.

[0081] Specific limitations regarding the training system for medical image enhancement models can be found in the limitations on training methods for medical image enhancement models described above, and will not be repeated here. Each module in the aforementioned training system for medical image enhancement models can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware format, or stored in the memory of a computer device in software format, so that the processor can call the corresponding operations of each module.

[0082] It should be noted that, in order to highlight the innovative aspects of this invention, this embodiment does not include modules that are not closely related to solving the technical problems proposed by this invention, but this does not mean that there are no other modules in this embodiment.

[0083] Please see Figure 6 The electronic device 600 may include a memory 610, a processor 620, and a bus, and may also include a computer program stored in the memory 610 and executable on the processor 620, such as a training program for a medical image enhancement model or a medical image enhancement program.

[0084] The memory 610 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 610 can be an internal storage unit of the electronic device 600, such as the portable hard drive of the electronic device 600. In other embodiments, the memory 610 can be an external storage device of the electronic device 600, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 600. Furthermore, the memory 610 can include both internal and external storage units of the electronic device 600. The memory 610 can be used not only to store application software and various types of data installed on the electronic device 600, such as training code for medical image enhancement models or code for enhancing medical images, but also to temporarily store data that has been output or will be output.

[0085] In some embodiments, the processor 620 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 620 is the control unit of the electronic device 600, connecting various components of the electronic device 600 via various interfaces and lines. It executes programs or modules stored in the memory 610 (e.g., training of medical image enhancement models or medical image enhancement programs), and calls data stored in the memory 610 to perform various functions of the electronic device 600 and process data.

[0086] The processor 620 executes the operating system of the electronic device 600 and various installed applications. The processor 620 executes the applications to implement the steps in the above-described training of the medical image enhancement model or the medical image enhancement method.

[0087] For example, the computer program may be divided into one or more modules, which are stored in the memory 610 and executed by the processor 620 to complete this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the electronic device 600. For example, the computer program may be divided into a data acquisition module 510, a virtual augmented image generation module 520, a registration module 530, a discrimination module 540, and a model training module 550.

[0088] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium, which can be non-volatile or volatile. The software functional module, stored in the storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute part of the functions of the training of the medical image enhancement model or the medical image enhancement method described in the various embodiments of this application.

[0089] In summary, this invention discloses a method, system, device, and medium for enhancing medical images and training models. The virtual enhanced images generated by this invention have higher contrast and clarity compared to traditional virtual enhanced images, improving the diagnostic value of the images and helping medical imaging products gain a competitive advantage in the market. In clinical applications, traditional contrast-enhanced MRI examinations typically require lengthy scans and contrast agent injections. This invention completely eliminates the need for contrast agents, significantly shortening scan time, improving examination efficiency, reducing hospital operating costs, optimizing the overall medical service process, reducing the burden on medical staff, and shortening patient waiting times, thus improving the patient experience. Furthermore, this invention is particularly suitable for patients who cannot use contrast agents, such as those with renal insufficiency, contrast agent allergies, or other contraindications, expanding its clinical applicability and facilitating the promotion and application of image enhancement technology in more medical institutions. Therefore, this invention, by optimizing image alignment, improving image quality, reducing computational costs, and expanding applicability, technically surpasses existing methods, demonstrating significant advantages in performance, efficiency, cost, and applicability. This innovative technology not only improves the quality and reliability of image diagnosis but also optimizes medical workflows, profoundly impacting the promotion and application of medical image enhancement technology. Therefore, this invention effectively overcomes the various shortcomings of existing technologies and possesses high industrial applicability.

[0090] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A training method for a medical image enhancement model, characterized in that, The training method includes: Acquire non-contrast medical images and their paired contrast-enhanced images; wherein, the non-contrast medical images are acquired using a contrast-free medical imaging method, and the contrast-enhanced images are the contrast-enhanced versions of the corresponding medical images, and the contrast-enhanced images have corresponding preset real labels; Non-contrast-enhanced medical images are input into a single-branch generator, and image features of the non-contrast-enhanced medical images are extracted to generate an initial virtual enhanced image; wherein, the initial virtual enhanced image has a corresponding preset virtual label; The initial virtual augmented image and the corresponding contrast-enhanced image are input into the registration network, and the initial virtual augmented image is aligned to the coordinate space of the corresponding contrast-enhanced image to generate the aligned virtual augmented image. The discriminator obtains the real image discrimination value corresponding to the contrast-enhanced image and the virtual image discrimination value corresponding to the initial virtual-enhanced image. The adversarial loss between the real image discriminant value and its corresponding real label, and the adversarial loss between the virtual image discriminant value and its corresponding virtual label are calculated. The correlation loss between the aligned virtual enhanced image and the contrast enhanced image is also calculated. The single-branch generator is updated based on the adversarial loss and the correlation loss to obtain the trained medical image enhancement model. The medical image enhancement model is a single-branch generator.

2. The training method for the medical image enhancement model according to claim 1, characterized in that, The non-contrast-enhanced medical image includes images of at least two modalities. The step of inputting the non-contrast-enhanced medical image into a single-branch generator, extracting the image features of the non-contrast-enhanced medical image, and generating an initial virtual enhanced image includes: stitching the images of each modality together, inputting the stitched image into the single-branch generator, extracting the image features of the stitched image, and generating an initial virtual enhanced image corresponding to the non-contrast-enhanced medical image.

3. The training method for the medical image enhancement model according to claim 1, characterized in that, The step of inputting the initial virtual augmented image and the corresponding contrast-enhanced image into the registration network, aligning the initial virtual augmented image to the coordinate space of the corresponding contrast-enhanced image, and generating the aligned virtual augmented image includes: The contrast-enhanced image is input into the first feature extraction module of the registration network to extract the contrast image features of the contrast-enhanced image; The initial virtual augmented image is input into the second feature extraction module of the registration network to extract the virtual augmented image features of the initial virtual augmented image; The contrast image features and the virtual enhanced image features are input into the similarity calculation module of the registration network to calculate the similarity error between the contrast image features and the virtual enhanced image features; The similarity error is input into the deformation field generation module of the registration network to generate a deformation field based on the similarity error; wherein, the deformation field is used to characterize the offset of the initial virtual enhanced image relative to the contrast enhanced image; The deformation field is input into the resampling module of the registration network, and the initial virtual augmented image is resampled according to the deformation field to generate an aligned virtual augmented image.

4. The training method for the medical image enhancement model according to claim 3, characterized in that, The step of resampling the initial virtual augmented image based on the deformation field to generate an aligned virtual augmented image includes: For each pixel of the initial virtual augmented image: Extract the offset of the pixel in the deformation field and determine whether the offset is an integer: If so, the coordinates of the pixel are corrected based on the offset to obtain the corrected coordinates of the pixel; Otherwise, the offset is interpolated using the interpolation method, and the coordinates of the pixel are corrected using the interpolated offset to obtain the corrected coordinates of the pixel. Based on the coordinates of all corrected pixels and their corresponding pixel values, an aligned virtual augmented image is formed.

5. The training method for the medical image enhancement model according to claim 1, characterized in that, The discriminator is a local perception discriminator. Through the discriminator, the true image discrimination value corresponding to the contrast-enhanced image is obtained, including: The contrast-enhanced image is input into the cropping module of the local perception discriminator, and the contrast-enhanced image is cut to generate multiple local contrast-enhanced images; Multiple local contrast-enhanced images are input into the discrimination module of the local perception discriminator to discriminate each local contrast-enhanced image and generate a real image discrimination value.

6. The training method for the medical image enhancement model according to claim 1, characterized in that, The discriminator is a local perception discriminator. Through the discriminator, the virtual image discrimination value corresponding to the initial virtual augmented image is obtained, including: The initial virtual augmented image is input into the cropping module of the local perception discriminator, which cuts the initial virtual augmented image to generate multiple local virtual augmented images; Multiple local virtual augmented images are input into the discrimination module of the local perception discriminator to discriminate each local virtual augmented image and generate virtual image discrimination values.

7. The training method for the medical image enhancement model according to claim 1, characterized in that, The calculation of adversarial loss between real image discriminant values ​​and their corresponding real labels, and between virtual image discriminant values ​​and their corresponding virtual labels, as well as the calculation of correlation loss between aligned virtual enhanced images and contrast enhanced images, and updating the single-branch generator based on the adversarial loss and correlation loss, includes: Calculate the adversarial loss between the real image discrimination value and its corresponding real label, and between the virtual image discrimination value and its corresponding virtual label, and update the single-branch generator and the discriminator based on the adversarial loss; The correlation loss between the aligned virtual augmented image and the contrast-enhanced image is calculated, and the single-branch generator and the registration network are updated based on the correlation loss.

8. A method for enhancing medical images, characterized in that, The enhancement method includes: Acquiring non-contrast medical images; The non-contrast-treated medical images are input into a medical image enhancement model to generate virtual enhanced images. The medical image enhancement model generates virtual enhanced images by aligning the non-contrast-treated medical images to a preset coordinate space and extracting image features from the aligned non-contrast-treated medical images.

9. A training system for a medical image enhancement model, characterized in that, The system includes: The data acquisition module is used to acquire non-contrast medical images and their paired contrast-enhanced images. The non-contrast medical images are acquired using a contrast-free medical imaging method, while the contrast-enhanced images are the contrast-enhanced versions of the corresponding medical images and have corresponding preset real labels. The virtual augmented image generation module is used to input non-contrast-treated medical images into a single-branch generator, extract the image features of the non-contrast-treated medical images, and generate an initial virtual augmented image; wherein, the initial virtual augmented image has a corresponding preset virtual label; The registration module is used to input the initial virtual augmented image and the corresponding contrast-enhanced image into the registration network, align the initial virtual augmented image to the coordinate space of the corresponding contrast-enhanced image, and generate the aligned virtual augmented image. The discrimination module is used to obtain the real image discrimination value corresponding to the contrast-enhanced image and the virtual image discrimination value corresponding to the initial virtual-enhanced image through the discriminator. The model training module is used to calculate the adversarial loss between the real image discriminant value and the corresponding real label, the virtual image discriminant value and the corresponding virtual label, and the correlation loss between the aligned virtual enhanced image and the contrast enhanced image. Based on the adversarial loss and the correlation loss, the single-branch generator is updated to obtain the trained medical image enhancement model; wherein, the medical image enhancement model is a single-branch generator.

10. A medical imaging device, characterized in that, The medical imaging device includes: The imaging module is used to capture non-contrast medical images of the subject being examined; The processing module is used to input the non-contrast-treated medical images into the medical image enhancement model to generate virtual enhanced images; wherein, the medical image enhancement model generates virtual enhanced images by aligning the non-contrast-treated medical images to a preset coordinate space and extracting image features from the aligned non-contrast-treated medical images.

11. An electronic device, characterized in that, The electronic device includes: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the training method of the medical image enhancement model as claimed in any one of claims 1 to 7 or the medical image enhancement method as claimed in claim 8.

12. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by the computer's processor, causes the computer to perform the training method of the medical image enhancement model as described in any one of claims 1 to 7 or the medical image enhancement method as described in claim 8.

Images