Contrast image intelligent analysis method and system based on generative counterfactual explanation

Abstract

The application discloses a contrast image intelligent analysis method and system based on generative counterfactual explanation, comprising the following steps: collecting paired medical image data of early contrast agent filling period and late contrast agent diffusion period in an oviduct; extracting and calculating the difference value of the late image features and the early image features to construct dynamic features; extracting early image spatial features, and designing a multi-scale gating fusion up-sampling mechanism to gradually fuse the spatial features and the dynamic features, dynamically guide the reconstruction of the spatial features, generate counterfactual images while ensuring the continuity of the structure; quantifying the pixel-level difference between the generated counterfactual images and the late images; and better assisting doctors in diagnosis by quantifying the pixel-level difference between the real and ideal states and visually presenting the difference part, thereby improving the efficiency and accuracy of clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis and artificial intelligence-assisted diagnosis technology, specifically to a method and system for intelligent analysis of hysterosalpingography images based on generative counterfactual interpretation, used to identify and analyze hysterosalpingography images in medical imaging, assisting clinicians in improving the accuracy and efficiency of diagnosis. Background Technology

[0002] In existing technologies, hysterosalpingography (HSG) analysis techniques mainly include X-ray hysterosalpingography, two-dimensional ultrasound HSG, three-dimensional ultrasound HSG, four-dimensional ultrasound HSG, and magnetic resonance-guided hysterosalpingography. These methods all have certain limitations in practical applications. They generally face problems such as reliance on manual operation, poor image interpretability, insufficient dynamic observation, high rates of missed and misdiagnosed cases, and lack of automated analysis. Furthermore, after the images are output by the above methods, doctors need to make manual judgments, which lacks interpretability. Therefore, there is an urgent need for intelligent analysis methods and systems for HSG images based on generative counterfactual interpretation to solve the problems existing in the above-mentioned technologies. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent analysis method and system for hysterosalpingography (HSG) images based on generative counterfactual interpretation, which solves the problems of lack of interpretability and high misdiagnosis rate in existing HSG analysis systems, and realizes more efficient and faster intelligent analysis of HSG images.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: an intelligent analysis method for contrast-enhanced images based on generative counterfactual interpretation, comprising the following steps:

[0005] Acquire contrast images, including early images of the early contrast agent filling phase and late images of the late contrast agent diffusion phase in the fallopian tube;

[0006] The difference between late-stage image features and early-stage image features is extracted and calculated to construct dynamic features. These dynamic features are 64-dimensional temporal dynamic representations obtained by extracting high-level semantic feature differences from early-stage and late-stage images through a deep neural network, followed by feature compression and normalization. Each dimension corresponds to a different contrast agent diffusion dynamics mode.

[0007] Early image spatial features are extracted, and a multi-scale gated fusion upsampling mechanism is designed to fuse spatial features and dynamic features step by step, dynamically guiding the reconstruction of spatial features and generating counterfactual images while ensuring structural continuity.

[0008] The pixel-level differences between the quantized counterfactual image and the later image.

[0009] Preferably, the difference range is [-1, 1]. The difference range is hard-constrained by the tanh activation function. Temporary out-of-bounds values ​​that occur during training are handled by a combination of gradient clipping and L2 normalization to ensure the stability of the system under extreme cases.

[0010] Preferably, the fusion at each level includes:

[0011] Perform a pixel shuffle operation;

[0012] Feature optimization is performed using 3×3 convolution combined with instance normalization.

[0013] The extracted spatial features and dynamic features are adaptively weighted and fused; and

[0014] By automatically learning the optimal fusion weights through Sigmoid-activated convolutional layers, dynamic features can automatically adjust the reconstruction intensity of different anatomical structures.

[0015] This invention also discloses an intelligent imaging image analysis system based on generative counterfactual interpretation, comprising:

[0016] The image acquisition module acquires and automatically matches pairs of early and late hysterosalpingography images to form a dataset;

[0017] An improved generative adversarial network module includes a generator and a discriminator. The generator includes an encoder, a temporal difference module, and a decoder. The encoder is configured to extract spatial features of an image, and the temporal difference module is configured to calculate the difference between late-stage image features and early-stage image features to construct dynamic features. The decoder incorporates a multi-scale gated fusion upsampling mechanism to fuse spatial features and dynamic features step by step to generate counterfactual images.

[0018] The discriminator uses a deep convolutional architecture to achieve multi-level identification of image authenticity.

[0019] The clinical analysis module calculates the pixel-level differences between the generated counterfactual images and the actual late-stage images, and quantifies these pixel-level differences.

[0020] The visualization module generates a four-panel comparison chart consisting of the original early stage image, the actual late stage image, the generated counterfactual image, and a difference heatmap.

[0021] The report generation module generates structured PDF reports based on a four-panel comparison chart.

[0022] Preferably, the image data in the dataset is uniformly adjusted to a resolution of 256×256 using OpenCV's LANCZOS4 interpolation algorithm, and grayscale normalization is performed to linearly transform the pixel values ​​to the range [0,1].

[0023] The dataset was expanded by randomly horizontally flipping, rotating, and adjusting the brightness, and the image data in the dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

[0024] Preferably, the encoder employs a three-level residual downsampling module with 64, 128, and 256 channels respectively; the temporal difference module compresses spatial features through global average pooling, then reduces the dimensionality to a 64-dimensional feature vector through a fully connected layer, and stabilizes the difference within the range of [-1, 1] through layer normalization and the LeakyReLU activation function; the decoder employs a three-level progressive upsampling, and performs pixel shuffling during each upsampling stage; it uses 3×3 convolution combined with instance normalization for feature optimization; it adaptively weights and fuses the extracted spatial features and dynamic features; and it automatically learns the optimal fusion weights through a Sigmoid-activated convolutional layer, enabling the dynamic features to automatically adjust the reconstruction intensity of different anatomical structures.

[0025] Preferably, the Sobel edge constraint operator is integrated into the output layer of the generator to enhance the anatomical accuracy of the fallopian tube edges through horizontal and vertical gradient operations.

[0026] Preferably, a progressive feature abstraction process is designed in the discriminator to construct a four-level convolutional downsampling structure; including:

[0027] The first stage uses 64-dimensional convolution;

[0028] The second stage uses 128-dimensional convolution;

[0029] The third level uses 256-dimensional convolution; and

[0030] The first three levels of feature transformation are combined with instance normalization to eliminate the interference caused by the difference in contrast agent concentration, and the LeakyReLU activation function is used to maintain gradient stability.

[0031] The fourth output layer uses 1D convolution to output a 31×31 discriminative feature map to achieve authenticity identification.

[0032] Preferably, the system further includes a training module for training the generative adversarial network module, comprising:

[0033] Parameters are updated using the AdmaW optimizer; the initial learning rate is 0.0001.

[0034] Set the batch size to 8 and the maximum training epochs to 100; implement a dynamic learning rate decay strategy with linear decay; adopt an early stopping mechanism with a patience value of 15 epochs.

[0035] The generator and discriminator parameters are updated alternately, the training loss function value and validation set index are recorded in real time, and the optimal model parameters are saved.

[0036] Preferably, a triple collaborative constraint mechanism is set in the generator, including reconstruction loss function constraint, adversarial loss function constraint, and edge-aware function constraint. The reconstruction loss maintains the consistency of image structure; the adversarial loss ensures that the generated distribution is close to the real distribution; the edge-aware loss maintains the sharpness of the anatomical structure through the Sobel operator, and maintains the optimal balance of reconstruction loss, adversarial loss, and edge loss through an adaptive weight adjustment algorithm during training.

[0037] Preferably, the optimization balance specifically involves: automatically differentiating and calculating the gradient norms of reconstruction loss, adversarial loss, and edge loss, normalizing them using a function, and generating dynamic weight coefficients; continuously monitoring the ratio of the three gradients, and immediately performing an optimization balance operation when the gradient norm ratio of reconstruction loss, adversarial loss, and edge loss deviates from the preset range of 1.0:0.8~1.2:0.25~0.35.

[0038] Beneficial effects: The method of this invention extracts the difference between late-stage and early-stage image features, and uses this dynamic feature to capture and accurately understand the diffusion process of contrast agents. Then, by designing a multi-scale gated fusion upsampling mechanism, spatial features and dynamic features are fused step by step, dynamically guiding the reconstruction of spatial features and generating images while ensuring structural continuity. Finally, difference map calculation is achieved. By quantifying the pixel-level differences between the real and ideal states, the display effect is continuously trained and optimized on the generated basic four-panel contrast view, gradually increasing the sensitivity of the difference heatmap. By visually presenting the differences, it better assists doctors in diagnosis and improves the efficiency and accuracy of clinical diagnosis.

[0039] In this invention system, a seamless end-to-end workflow is achieved. Starting with data preprocessing, the system first completes image registration and standardization, laying the foundation for subsequent analysis. Then, the model training process automatically optimizes parameters and evaluates performance on the validation set, forming a continuous closed loop. Finally, the clinical analysis module is implemented to quantify the pixel-level differences between counterfactual images and real late-stage images. The entire process requires no manual intervention. Attached Figure Description

[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0041] In the attached diagram:

[0042] Figure 1 This is a flowchart of the system training process of this invention;

[0043] Figure 2 This is a flowchart of the data preprocessing process in the system of this invention;

[0044] Figure 3 This is a flowchart of the construction process of the improved generative adversarial network module in the system of this invention;

[0045] Figure 4 This is a flowchart of the adversarial training process in the system of this invention;

[0046] Figure 5 This is a flowchart of the verification and evaluation process in the system of this invention;

[0047] Figure 6 This is a system module diagram of the present invention;

[0048] Figure 7 This is a diagram illustrating the implementation of importing early and late contrast images into the system.

[0049] Figure 8 This invention uses the original image to generate a counterfactual graph through a generative adversarial network module;

[0050] Figure 9 This is a comparison result diagram of the four-panel diagram of the present invention. Detailed Implementation

[0051] The embodiments of the present invention will now be described with reference to the accompanying drawings. The terminology used in the embodiments section is for illustrative purposes only and is not intended to limit the scope of the invention. The embodiments of this application will now be described with reference to the accompanying drawings.

[0052] Example 1: Intelligent Analysis Method for Contrast Imaging Based on Generative Counterfactual Interpretation, referencing Figure 1 The flowchart shown is a flowchart of the present invention, which includes the following steps:

[0053] S1. Data Collection:

[0054] Early images of the early contrast agent filling phase and late images of the late contrast agent diffusion phase in the fallopian tubes were acquired and paired.

[0055] S2. Data preprocessing, refer to... Figure 2 The diagram shown is a flowchart of the data preprocessing process, including:

[0056] Image size standardization is performed by using OpenCV's LANCZOS4 interpolation algorithm to uniformly adjust the image to a resolution of 256×256 and then performing grayscale normalization to linearly transform the pixel values ​​to the range of [0,1].

[0057] Data augmentation was performed by expanding the training dataset through random horizontal flipping by 10 degrees, rotation, and brightness adjustment.

[0058] To divide the dataset, this invention divides the dataset into a training set, a validation set, and a test set in a ratio of 7:2:1.

[0059] S3. Design an improved generative adversarial network module, referencing... Figure 3 The diagram shown illustrates the construction flowchart of the improved generative adversarial network module, which involves the following operations:

[0060] The differences between late-stage and early-stage image features are extracted and calculated, with the difference ranging from [-1, 1]. Dynamic features are constructed by extracting high-level semantic feature differences from early and late-stage images using a deep neural network, followed by feature compression and normalization to obtain a 64-dimensional temporal dynamic representation. Each dimension corresponds to a different contrast agent diffusion dynamics pattern.

[0061] Early image spatial features are extracted, and a multi-scale gated fusion upsampling mechanism is designed to fuse spatial features and dynamic features step by step, dynamically guiding the reconstruction of spatial features and generating counterfactual images while ensuring structural continuity.

[0062] Each level of fusion includes:

[0063] Perform pixel shuffling to improve feature map resolution;

[0064] Feature optimization is performed using 3×3 convolution combined with instance normalization.

[0065] The extracted spatial features and dynamic features are adaptively weighted and fused; and

[0066] By automatically learning the optimal fusion weights through Sigmoid-activated convolutional layers, dynamic features can automatically adjust the reconstruction intensity of different anatomical structures.

[0067] S4. Model training, refer to Figure 4 The diagram shown is a flowchart of adversarial training. The specific training process includes:

[0068] The AdmaW optimizer was used for parameter updates with β1 = 0.5 and β2 = 0.999. The first-order moment decay rate of β1 = 0.5 ensures the optimization process remains highly sensitive to recent gradients, retaining only 50% of historical momentum information with each parameter update. This setting significantly enhances the model's responsiveness to sudden gradient changes in medical images, particularly in scenarios involving sudden contrast agent enhancement or sharp changes in vessel edges, enabling rapid adjustment of the update direction. Simultaneously, the second-order moment decay rate of β2 = 0.999 maintains long-term memory of the squared gradients. By retaining 99.9% of historical squared gradient information, a stable adaptive learning rate adjustment mechanism is constructed, effectively balancing the update magnitude of parameters across different layers. The initial learning rate was 0.0001.

[0069] The batch size is set to 8, and the maximum number of training epochs is set to 100. A dynamic learning rate decay strategy is implemented, with linear decay. An early stopping mechanism is used, with a patience value of 15 epochs to prevent overfitting.

[0070] The generator and discriminator parameters are updated alternately, the training loss function value and validation set index are recorded in real time, and the optimal model parameters are saved.

[0071] S5. Quantize the pixel-level differences between the generated counterfactual image and the later image, and evaluate the performance over a preset number of epochs, referencing... Figure 5 As shown in the flowchart, the performance is evaluated on the validation set every 5 epochs in this embodiment, forming a continuous closed loop. Finally, the clinical analysis module is implemented to quantify the pixel-level differences between the counterfactual images and the real late-stage images. The whole process does not require manual intervention. The structural similarity index (SSIM) of the generated images is also measured.

[0072] Example 2: Reference Figure 6 As shown, the intelligent image analysis system based on generative counterfactual interpretation includes:

[0073] Image acquisition module: Automatically matches paired early and late hysterosalpingography images, verifies image resolution, grayscale range and pairing validity, ensures the integrity and consistency of input data, standardizes the data source for subsequent processing, and constitutes a dataset; and preprocesses the dataset using the method described in Example 1.

[0074] An improved generative adversarial network (GAN) module includes a generator and a discriminator. The generator comprises an encoder, a temporal difference module, and a decoder. The encoder uses a three-level residual downsampling module with 64, 128, and 256 channels, respectively. The temporal difference module calculates the feature differences between early and late hysterosalpingography (HSG) images, compresses spatial features through global average pooling, and then reduces the dimensionality to a 64-dimensional feature vector through a fully connected layer. Layer normalization and the LeakyReLU activation function ensure that the feature difference is stable within the range of [-1, 1]. The decoder employs a multi-scale gated fusion upsampling mechanism, dynamically guiding spatial feature reconstruction through temporal difference features to achieve image resolution. While restoring the original size, the continuity of the structure is ensured through three levels of progressive upsampling. In each level of upsampling, the resolution of the feature map is first improved by pixel shuffling, and then 3×3 convolution combined with instance normalization is used for feature optimization. The mechanism adaptively weights and fuses the spatial features extracted by the encoder with the dynamic features generated by the temporal difference module. The optimal fusion weights are automatically learned by the Sigmoid-activated convolutional layer, so that the dynamic features can intelligently adjust the reconstruction intensity of different anatomical structures. In the final output layer, the Sobel edge constraint operator is integrated to enhance the anatomical accuracy of the fallopian tube edge through horizontal and vertical gradient operations.

[0075] The improved generative adversarial network module integrates prior knowledge from the medical field and employs a dynamic gradient balancing mechanism to intelligently adjust the balance between different optimization objectives, making the model training process more stable and efficient. In terms of network module architecture design, a multi-scale discrimination system is introduced to achieve comprehensive control over medical image features. In response to the special requirements of medical imaging, professional anatomical knowledge is transformed into structured constraints, focusing on enhancing the generation quality of key anatomical regions. Through edge-preserving algorithms and vascular region enhancement mechanisms, it is ensured that the generated images maintain the authenticity of the overall structure while clearly presenting important pathological features.

[0076] The discriminator employs a deep convolutional architecture to achieve multi-level identification of image authenticity. Through a progressive feature abstraction process, a four-level convolutional downsampling structure is constructed: the first level uses 64-dimensional convolution, the second level uses 128-dimensional convolution, and the third level uses 256-dimensional convolution. Each level of feature transformation is combined with instance normalization to eliminate interference caused by differences in contrast agent concentration, and the LeakyReLU activation function is used to maintain gradient stability. Finally, the output layer uses 1-dimensional convolution to output a 31×31 discriminative feature map to achieve authenticity identification.

[0077] Clinical Analysis Module: Calculates the pixel-level differences between the generated counterfactual image and the real late-stage image, and outputs the probability P as the quantification.

[0078]

[0079] x represents the real late-stage image, y represents the generated counterfactual image, H represents the image height, and W represents the image width;

[0080] Visualization module: Generates a four-panel comparison chart of the original early stage, the actual late stage, the generated counterfactual image, and the difference heatmap;

[0081] The report generation module generates a structured PDF report based on a four-panel comparison chart, containing key images and probabilities P. Observing the key images and probabilities P in the PDF report can assist doctors in making judgments and analyses. For example:

[0082]

[0083] Training module: Trains the generative adversarial network module, including:

[0084] The AdmaW optimizer was used for parameter updates with β1 = 0.5 and β2 = 0.999. The first-order moment decay rate of β1 = 0.5 ensures the optimization process remains highly sensitive to recent gradients, retaining only 50% of historical momentum information with each parameter update. This setting significantly enhances the model's responsiveness to sudden gradient changes in medical images, particularly in scenarios involving sudden contrast agent enhancement or sharp changes in vessel edges, enabling rapid adjustment of the update direction. Simultaneously, the second-order moment decay rate of β2 = 0.999 maintains long-term memory of the squared gradients. By retaining 99.9% of historical squared gradient information, a stable adaptive learning rate adjustment mechanism is constructed, effectively balancing the update magnitude of parameters across different layers. The initial learning rate was 0.0001.

[0085] The batch size is set to 8, and the maximum number of training epochs is set to 100. A dynamic learning rate decay strategy is implemented, with linear decay. An early stopping mechanism is used, with a patience value of 15 epochs to prevent overfitting.

[0086] The generator and discriminator parameters are updated alternately, the training loss function value and validation set index are recorded in real time, and the optimal model parameters are saved.

[0087] Configure multi-objective loss functions in the YAML file: reconstruction loss (L1 norm) with a weight of 1.0, adversarial loss with a weight of 0.1, and edge-aware loss with a weight of 0.3. Counterfactual medical image generation is achieved through a triple collaborative constraint mechanism, using a dynamic balancing strategy. During training, an adaptive weight adjustment algorithm maintains an optimal balance between reconstruction loss, adversarial loss, and edge loss. The reconstruction loss (L1 norm) maintains image structural consistency; the adversarial loss ensures the generated distribution approximates the true distribution; and the edge-aware loss maintains anatomical sharpness through the Sobel operator.

[0088] The quality control module can evaluate performance in a preset number of epochs, such as evaluating performance on the validation set every 5 epochs to form a continuous closed loop; the final clinical analysis module quantifies the pixel-level differences between counterfactual images and real late-stage images, and the whole process does not require manual intervention.

[0089] Measure the structural similarity index (SSIM) of the generated images.

[0090]

[0091] x represents the generated image, y represents the real image, and N represents the total number of pixels in the image; μ x To generate the pixel mean of image x, μ y Let y be the pixel mean of the real image. To generate the pixel variance of image x, Let σ be the pixel variance of the real image y. xy Let C1 and C2 be the covariance of the two images; C1 and C2 are constants, and the optimal SSIM is 0.64.

[0092] In a specific case, the hysterosalpingography of patient A was analyzed based on the system of Example 2;

[0093] refer to Figure 7 As shown, firstly, in the image acquisition module, click the import button to obtain the initial imaging image and contrast agent diffusion image of patient A, and automatically match the paired early and late fallopian tube imaging images;

[0094] Then, clicking the "Analyze" button preprocesses the image and, in conjunction with the quality control module, verifies the input data and the reasonableness of the output, ensuring the reliability of the system and achieving standardized processing of medical images. Simultaneously, it calls the best-trained improved generative adversarial network module to generate high-quality counterfactual images to simulate the ideal contrast agent diffusion state, for reference. Figure 8 As shown;

[0095] The clinical analysis module calculates the pixel-level differences between the generated image and the actual late-stage image, outputting a probability P as the quantification. The calculated probability is 4.6%. Combined with the visualization module, a pop-up window generates a four-panel comparison chart showing the original early-stage image, the actual late-stage image, the generated image, and a difference heatmap, providing intuitive model output. The final output is structured analysis results, such as... Figure 9 As shown;

[0096] In the report generation module, you can generate a structured PDF report containing four comparison charts by clicking the report generation button.

[0097] 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. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. An intelligent analysis method for contrast-enhanced images based on generative counterfactual interpretation, characterized in that: Includes the following steps: Acquire contrast images, including early images of the early contrast agent filling phase and late images of the late contrast agent diffusion phase in the fallopian tube; The difference between late-stage image features and early-stage image features is extracted and calculated to construct dynamic features. The dynamic features are 64-dimensional temporal dynamic representations obtained by extracting high-level semantic feature differences from early-stage and late-stage images through deep neural networks, and then performing feature compression and normalization. Each dimension corresponds to a different contrast agent diffusion dynamic mode. as well as Early image spatial features are extracted, and a multi-scale gated fusion upsampling mechanism is designed to fuse spatial features and dynamic features step by step, dynamically guiding the reconstruction of spatial features and generating counterfactual images while ensuring structural continuity. The pixel-level differences between the quantized counterfactual image and the later image; Generate a four-panel comparison chart consisting of the original early stage, the actual late stage, the generated counterfactual image, and the difference heatmap; Generate a structured PDF report based on a four-part comparison chart.

2. The intelligent analysis method for contrast-enhanced images based on generative counterfactual interpretation according to claim 1, characterized in that: The difference range is [-1, 1]. This difference range is hard-constrained by the tanh activation function. Temporary out-of-bounds values ​​that occur during training are handled by a combination of gradient clipping and L2 normalization.

3. The intelligent analysis method for contrast-enhanced images based on generative counterfactual interpretation according to claim 1, characterized in that: Each level of fusion includes: Perform a pixel shuffle operation; Feature optimization is performed using 3×3 convolution combined with instance normalization. The extracted spatial features and dynamic features are adaptively weighted and fused; and The optimal fusion weights are automatically learned through convolutional layers activated by Sigmoid.

4. A system for implementing the intelligent analysis method for contrast-images based on generative counterfactual interpretation as described in claim 1, characterized in that: include: The image acquisition module acquires and automatically matches pairs of early and late hysterosalpingography images to form a dataset; An improved generative adversarial network module includes a generator and a discriminator. The generator includes an encoder, a temporal difference module, and a decoder. The encoder is configured to extract spatial features of an image, and the temporal difference module is configured to calculate the difference between late-stage image features and early-stage image features to construct dynamic features. The decoder has a built-in multi-scale gated fusion upsampling mechanism that fuses spatial features and dynamic features step by step to generate counterfactual images. The discriminator uses a deep convolutional architecture to achieve multi-level identification of image authenticity. The clinical analysis module calculates the pixel-level differences between the generated counterfactual images and the actual late-stage images, and quantifies these pixel-level differences. The visualization module generates a four-panel comparison chart consisting of the original early stage image, the actual late stage image, the generated counterfactual image, and a difference heatmap. The report generation module generates structured PDF reports based on a four-panel comparison chart.

5. The system according to claim 4, characterized in that: The image data in the dataset is uniformly adjusted to 256×256 resolution using OpenCV's LANCZOS4 interpolation algorithm, and grayscale normalization is performed to linearly transform the pixel values ​​to the range of [0,1]. as well as The dataset was expanded by randomly horizontally flipping, rotating, and adjusting the brightness, and the image data in the dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

6. The system according to claim 4 or 5, characterized in that: The encoder employs a three-level residual downsampling module with 64, 128, and 256 channels respectively. The temporal difference module compresses spatial features through global average pooling, then reduces the dimensionality to a 64-dimensional feature vector through a fully connected layer, and stabilizes the difference within the range of [-1, 1] through layer normalization and the LeakyReLU activation function. The decoder employs a three-level progressive upsampling, and performs pixel shuffling during each upsampling level. It uses 3×3 convolution combined with instance normalization for feature optimization. The extracted spatial features and dynamic features are adaptively weighted and fused. The optimal fusion weights are automatically learned through a Sigmoid-activated convolutional layer.

7. The system according to claim 6, characterized in that: The Sobel edge constraint operator is integrated into the output layer of the generator.

8. The system according to claim 4 or 5, characterized in that: A progressive feature abstraction process is designed in the discriminator, constructing a four-level convolutional downsampling structure; including: The first stage uses 64-dimensional convolution; The second stage uses 128-dimensional convolution; The third level uses 256-dimensional convolution; and The first three levels of feature transformation are combined with instance normalization to eliminate the interference caused by the difference in contrast agent concentration, and the LeakyReLU activation function is used to maintain gradient stability. The fourth output layer uses 1D convolution to output a 31×31 discriminative feature map.

9. The system according to claim 5, characterized in that: The system also includes a training module for training the generative adversarial network module, comprising: Parameters are updated using the AdmaW optimizer; the initial learning rate is 0.0001. Set the batch size to 8 and the maximum training epochs to 100; implement a dynamic learning rate decay strategy with linear decay; adopt an early stopping mechanism with a patience value of 15 epochs. The generator and discriminator parameters are updated alternately, the training loss function value and validation set index are recorded in real time, and the optimal model parameters are saved.

10. The system according to claim 9, characterized in that: A triple collaborative constraint mechanism is set in the generator, including reconstruction loss function constraint, adversarial loss function constraint, and edge awareness function constraint. During training, an adaptive weight adjustment algorithm is used to maintain an optimal balance between reconstruction loss, adversarial loss, and edge loss.

Images