Heart function region modeling method based on deep generative adversarial network

Abstract

The invention discloses a cardiac functional region modeling method based on a deep generative adversarial network. The method comprises the following steps: S1, collecting cardiac medical image data; s2, constructing a cardiac medical image data set; s3, constructing an improved deep generative adversarial network; s4, inputting the preprocessed cardiac medical image data into an improved deep generative adversarial network, and outputting a cardiac functional region segmentation prediction result by a generator of the improved deep generative adversarial network; s5, optimizing network parameters of the generator and the discriminator; and S6, performing reasoning calculation on the preprocessed cardiac medical image data, outputting a cardiac function region segmentation result, and performing post-processing on the cardiac function region segmentation result based on discrimination results of the global discriminator and the local discriminator to generate a final cardiac function region modeling result. According to the method, the segmentation precision of the heart function region, the consistency of cross-modal image segmentation and the reliability of a final segmentation result are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cardiac functional area modeling, and in particular to a cardiac functional area modeling method based on a deep generative adversarial network. Background Art

[0002] With the rapid development of deep learning technology, advanced artificial intelligence methods have gradually been introduced into the field of medical image analysis to improve the accuracy of disease diagnosis, treatment planning and postoperative evaluation. Accurate modeling of cardiac functional areas is a key task in the diagnosis of cardiovascular diseases, mainly including the automatic segmentation and functional evaluation of the left ventricle, right ventricle, left atrium, right atrium and valve areas of the heart. However, the automated analysis of cardiac images still faces many challenges in existing technologies.

[0003] Currently, deep learning-based medical image segmentation methods, such as U-Net, ResNet, and their variants, have been widely used for cardiac functional area segmentation tasks. These methods rely on large-scale labeled data for supervised learning to build high-precision cardiac image segmentation models. However, existing technologies still have the following major drawbacks:

[0004] First, the segmentation accuracy of existing deep learning models is easily affected by data quality and has insufficient generalization ability. Since medical imaging data are usually limited by imaging equipment, imaging parameters, and individual anatomical differences, which lead to large deviations in data distribution, the trained deep learning models may be difficult to apply to new data distributions. In addition, the existing models have poor segmentation effects on low-quality images and are difficult to meet clinical needs.

[0005] Secondly, the consistency of functional area segmentation in multimodal medical images is poor. There are significant differences in the resolution, tissue contrast and artifact characteristics of images of different modalities, making it difficult for traditional deep learning models to maintain consistent segmentation results between different modalities. This leads to deviations in the segmented areas during cross-modal analysis, affecting the accuracy of subsequent cardiac function measurements.

[0006] Third, existing methods are insufficient in processing the boundary details of cardiac functional areas. The organizational structure of cardiac images is complex, especially the segmentation of the valve and cardiac cavity junction areas often has fuzzy boundaries, which makes it difficult for traditional segmentation networks to accurately capture the cardiac anatomical structure. The current mainstream segmentation method based on global loss optimization fails to fully consider local details and is prone to misclassification or blurred boundaries in local areas.

[0007] In summary, existing medical image segmentation technologies have problems in cardiac functional area modeling tasks, such as limited segmentation accuracy, insufficient cross-modal consistency, poor local boundary processing capabilities, and strong dependence on data annotation, making it difficult to meet the needs of clinical diagnosis and personalized treatment planning. Summary of the Invention

[0008] One purpose of the present invention is to propose a cardiac functional area modeling method based on a deep generative adversarial network. The present invention significantly improves the segmentation accuracy of cardiac functional areas, the consistency of cross-modal image segmentation, and the reliability of the final segmentation results.

[0009] According to an embodiment of the present invention, a cardiac functional area modeling method based on a deep generative adversarial network includes the following steps:

[0010] S1. Collect cardiac medical image data and perform preprocessing of the collected cardiac medical image data including image registration, normalization, noise removal, and contrast enhancement;

[0011] S2. Use expert annotation methods to segment and annotate cardiac functional regions in the preprocessed cardiac medical image data to construct a cardiac medical image dataset;

[0012] S3. Build an improved deep generative adversarial network based on cardiac medical imaging dataset;

[0013] S4. Inputting the preprocessed cardiac medical image data into the improved deep generative adversarial network, the generator of the improved deep generative adversarial network outputs a cardiac functional region segmentation prediction result, and inputting the corresponding expert-annotated cardiac functional region segmentation result in the cardiac medical image data and the cardiac functional region segmentation prediction result output by the generator into the global discriminator and the local discriminator respectively for authenticity discrimination;

[0014] S5. Perform adversarial training on the generator, the global discriminator, and the local discriminator according to the discrimination results output by the global discriminator and the local discriminator of the improved deep generative adversarial network, and optimize the network parameters of the generator and the discriminator;

[0015] S6. Perform inference calculation on the pre-processed cardiac medical image data, output the cardiac functional area segmentation result, and post-process the cardiac functional area segmentation result based on the discrimination results of the global discriminator and the local discriminator to generate the final cardiac functional area modeling result.

[0016] Optionally, the S1 includes the following steps:

[0017] S11. Collect cardiac medical image data, wherein the resolution of the cardiac medical image data is set to W*H, where W is the width of the cardiac medical image and H is the height of the cardiac medical image;

[0018] S12. Performing rigid and non-rigid transformations on the acquired cardiac medical image data, wherein the rigid transformations include rotation, scaling, and translation. The non-rigid transformations are optimized for pixel-level alignment using a deformation field, so that the cardiac medical image data processed by the deformation field are anatomically aligned.

[0019] S13. Normalizing the registered cardiac medical image data to a range of [0, 1].

[0020] S14. De-noise the normalized cardiac medical image data, perform contrast enhancement on the de-noised cardiac medical image data, optimize the resolution of cardiac functional areas, and obtain enhanced cardiac medical image data I enhanced .

[0021] Optionally, S2 includes the following steps:

[0022] S21. Using expert annotation method to label cardiac medical image data enhanced Perform segmentation and annotation of cardiac functional regions, including the left ventricle, right ventricle, left atrium, right atrium, and valve regions. The expert-annotated cardiac functional region segmentation mask is represented by a binary mask:

[0023]

[0024] Among them, M(x) represents the segmentation mask value of pixel x, R heart For the cardiac functional area, each cardiac medical image data has a corresponding expert-labeled cardiac functional area segmentation mask M(x);

[0025] S22. Perform quality assessment on the annotated cardiac functional area segmentation mask M(x) and calculate the morphological closure C of the segmented area heart :

[0026]

[0027] Among them, R ref is the expert reference segmentation region, and the |·| table indicates the number of pixels in the region;

[0028] S23. Divide the labeled cardiac medical imaging data that have passed the quality assessment and its corresponding cardiac functional area segmentation mask M(x) into a training set, a validation set, and a test set according to a preset ratio.

[0029] Optionally, S3 includes the following steps:

[0030] S31. Construct an improved deep generative adversarial network, the deep generative adversarial network includes a generator G θ, global discriminator D G With the local discriminator D L , the generator G θ Realize cardiac functional area segmentation prediction based on encoding and decoding structure;

[0031] S32. Set the generator G θ The input is the preprocessed cardiac medical image data I enhanced , the output is the predicted cardiac functional area segmentation mask

[0032] S33. Setting the global discriminator D G Cardiac functional area segmentation mask for judgment prediction Whether it matches the expert-labeled cardiac functional area segmentation mask M(x);

[0033] S34. Set local discriminator D L Taking the local structure of the cardiac functional area as input, the left ventricle, right ventricle, left atrium, right atrium and valve area are independently discriminated. The local discriminator D L Used to judge the authenticity of local details between the specific cardiac function area segmentation results output by the generator and the corresponding cardiac function area segmentation results annotated by experts.

[0034] Optionally, the S32 includes the following steps:

[0035] S321. Set the input of the generator to the pre-processed cardiac medical imaging data, and construct the generator using the encoding-decoding structure. The generator consists of the encoder E θ With decoder D θ Composition, encoder E θ Extract deep features Z of cardiac medical imaging data;

[0036] S323. Set decoder D θ With encoder E θ The extracted deep feature Z is used as input to generate the segmentation prediction mask of the cardiac functional area

[0037]

[0038] Among them, T j (·) represents the j-th layer transposed convolution operation, γ j is the learning scaling factor of the jth layer, b j is the bias term, σ is the Sigmoid activation function, N d is the number of layers of the decoder;

[0039] S324. Set up a skip connection mechanism to combine feature information of different scales during decoding to maintain the integrity of the boundaries of the cardiac functional area. The skip connection calculation is as follows:

[0040]

[0041] in, is the k-1th layer output of the decoder, is the k-th layer feature map of the encoder, which is connected by the skip connection function S k After processing, it is passed to the decoder, ⊙ and ⊕ represent element-by-element multiplication and element-by-element addition, respectively, which are used to realize the multi-mode fusion of feature information, η k is the equilibrium parameter, c k It is the bias term in the fusion process, and σ1 is used to activate the fusion result to fully retain the boundary and internal structure information of the cardiac functional area.

[0042] Optionally, the S5 includes the following steps:

[0043] S51. Set the global discriminator D G The discriminant loss function L G as follows:

[0044]

[0045] Among them, M represents the heart function area segmentation mask annotated by experts, Represents the cardiac functional area segmentation prediction mask output by the generator, and Represent the probability distribution of real and generated masks respectively;

[0046] S52. Set local discriminator D L The discriminant loss function L L as follows:

[0047]

[0048] Among them, r represents each specific area in the cardiac functional area, M r Annotate the segmentation mask for the expert in the corresponding area, The local segmentation mask generated by the generator, and Represent the probability distribution of local true and generated masks respectively;

[0049] S53. Define the cardiac function area segmentation evaluation loss L seg Segmentation results output by the quantized generator The difference between the expert-annotated segmentation result M(x) is:

[0050]

[0051] Among them, ∈ is a constant to prevent division by zero;

[0052] S54. Combine the global adversarial loss, local adversarial loss and cardiac functional area segmentation evaluation loss to define the joint generative adversarial training loss function L GAN for:

[0053] L GAN =λ G L G +λ L L L +λ seg L seg ;

[0054] Among them, λ G ,λ L and λ seg are hyperparameters that regulate the weights of global and local adversarial losses and segmentation losses during the overall training process, so that the generator output gradually approaches the segmentation results annotated by experts;

[0055] S55. Based on joint generation adversarial training loss L GAN The parameters of the generator are optimized using the gradient descent method, while the parameters of the global discriminator and the local discriminator are optimized using the gradient ascent method. The update rules are:

[0056]

[0057] Among them, θ represents the parameters of the generator, φ G With φ L Represent the parameters of the global discriminator and the local discriminator, η θ ,η G and η L is the corresponding learning rate;

[0058] S56. Repeat steps S51 to S55 until the cardiac function area segmentation prediction mask output by the generator and the segmentation mask M(x) annotated by the expert are in the joint generation adversarial training loss L GAN The convergence is achieved under the constraint of , achieving the adversarial training goal, that is, the generator output gradually approaches the real cardiac functional area segmentation result.

[0059] Optionally, the S6 includes the following steps:

[0060] S61. Using the generator optimized by adversarial training Perform inference calculation on the pre-processed cardiac medical image data to obtain the cardiac functional area segmentation prediction mask;

[0061] S62. Use the global discriminator after training Evaluate the overall authenticity between the generated cardiac function region segmentation prediction mask and the expert-annotated cardiac function region segmentation mask, and calculate the global discriminant score S G ;

[0062] S63. Use the local discriminator after training Evaluate the local authenticity of the generated cardiac function region segmentation prediction mask on different cardiac function regions and calculate the local segmentation score S L (r);

[0063]

[0064] Among them, R r is the pixel set of the cardiac functional area r, is the segmentation mask of the generator in region r, M r (x) is the true mask of the corresponding area marked by the expert, and Represents the gradient information of the predicted mask and the true mask, which is used to measure the boundary matching degree of the local segmentation result, λ B A weighting factor for scoring local segmentation based on boundary information;

[0065] S64. Calculate the final cardiac function region segmentation confidence score S based on the global discrimination score and each local segmentation score final :

[0066]

[0067] Among them, α G and α L (r) are the weighting factors for global and local segmentation scores, respectively;

[0068] S65. Adjust the cardiac functional area segmentation results output by the adaptive generator When the final cardiac function area segmentation confidence score S final When the value is lower than the set threshold, the segmentation result is optimized:

[0069]

[0070] in, represents the initial cardiac functional area segmentation prediction mask obtained by generator reasoning, M ref (x) is the reference segmentation mask verified with high confidence in history, representing the ideal model of the expert-annotated structure, β is the adaptive adjustment factor, T seg is the preset segmentation confidence threshold, which is used to distinguish high confidence from low confidence segmentation cases, and δ is the sharpness control parameter;

[0071] By constructing a piecewise nonlinear weighted function, when the final cardiac functional region segmentation confidence score is lower than the set threshold, the correction factor approaches 1, thereby correcting the initial cardiac functional region segmentation prediction mask to be closer to the reference segmentation mask. When the final cardiac functional region segmentation confidence score is close to or exceeds the set threshold, the correction factor approaches 0, maintaining the stability of the initial segmentation result, so that the final optimized cardiac functional region segmentation result meets the clinical anatomical structure requirements;

[0072] S66. Optimize the cardiac functional area segmentation results Perform morphological processing, eliminate boundary noise, and convert it into medical visualization format to obtain the final cardiac function area modeling result M final .

[0073] The beneficial effects of the present invention are:

[0074] (1) Based on the traditional generative adversarial network, the present invention introduces a dual adversarial mechanism of global discriminator and local discriminator, so that the generator not only optimizes the segmentation accuracy of the overall cardiac functional area during the learning process, but also maintains a higher boundary retention ability in key cardiac anatomical areas (valve, cardiac chamber junction). The global discriminator is responsible for judging the authenticity of the segmentation mask of the overall cardiac functional area, while the local discriminator adopts a regional refinement strategy to independently evaluate the left ventricle, right ventricle, left atrium, right atrium and valve areas, thereby ensuring that the segmentation results are more accurate in the structural details of each cardiac functional area. By optimizing the loss function to balance the global and local segmentation accuracies, the model can improve the accuracy of local segmentation while ensuring the consistency of the overall anatomical structure.

[0075] (2) The present invention introduces a multi-scale jump connection mechanism into the generator structure, so that cardiac anatomical features of different scales can be fully integrated during the segmentation process, ensuring that highly consistent segmentation effects can be achieved under different imaging modalities. The role of the jump connection is to combine the encoding features of different levels during the decoding process so that the model can capture the global information of the cardiac functional area while retaining the fine structure of the local boundary. Through this mechanism, the generator can effectively align the feature distributions of images of different modalities when cross-modal images are input, thereby improving the generalization ability of the model and reducing the segmentation deviation caused by different imaging modalities.

[0076] (3) In the post-segmentation processing stage, the present invention proposes an adaptive optimization method based on dynamic confidence. The segmentation confidence scores calculated by global and local discriminators are used to adaptively correct the low-confidence segmentation results. When the confidence of the segmentation result is lower than the set threshold, the current segmentation mask is dynamically adjusted by referring to the historical high-confidence segmentation results to make it closer to the real anatomical structure. In addition to model training, the reliability of the segmentation results is further improved, the sensitivity of the model to noise data is reduced, and the structural rationality of the cardiac functional area is ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0078] Figure 1 This is a flowchart of a cardiac functional area modeling method based on a deep generative adversarial network proposed in the present invention. DETAILED DESCRIPTION

[0079] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0080] refer to Figure 1 A cardiac functional area modeling method based on a deep generative adversarial network comprises the following steps:

[0081] S1. Collect cardiac medical image data and perform preprocessing of the collected cardiac medical image data including image registration, normalization, noise removal, and contrast enhancement;

[0082] S2. Use expert annotation methods to segment and annotate cardiac functional regions in the preprocessed cardiac medical image data to construct a cardiac medical image dataset;

[0083] S3. Build an improved deep generative adversarial network based on cardiac medical imaging dataset;

[0084] S4. Input the preprocessed cardiac medical image data into the improved deep generative adversarial network. The generator of the improved deep generative adversarial network outputs a cardiac functional region segmentation prediction result. The corresponding expert-annotated cardiac functional region segmentation result in the cardiac medical image data and the cardiac functional region segmentation prediction result output by the generator are respectively input into the global discriminator and the local discriminator for authenticity judgment.

[0085] S5. Perform adversarial training on the generator, global discriminator, and local discriminator based on the discrimination results output by the global discriminator and local discriminator of the improved deep generative adversarial network, and optimize the network parameters of the generator and discriminator;

[0086] S6. Perform inference calculation on the preprocessed cardiac medical imaging data, output the cardiac functional area segmentation results, and post-process the cardiac functional area segmentation results based on the discrimination results of the global discriminator and the local discriminator to generate the final cardiac functional area modeling results.

[0087] In this embodiment, S1 includes the following steps:

[0088] S11. Collect cardiac medical image data, the resolution of which is set to W*H, where W is the width of the cardiac medical image and H is the height of the cardiac medical image;

[0089] S12. Performing rigid and non-rigid transformations on the acquired cardiac medical image data. Rigid transformations include rotation, scaling, and translation. Non-rigid transformations perform pixel-level alignment optimization using a deformation field, so that the cardiac medical image data processed by the deformation field are anatomically aligned.

[0090] S13. Normalizing the registered cardiac medical image data to a range of [0, 1].

[0091] S14. De-noise the normalized cardiac medical image data, perform contrast enhancement on the de-noised cardiac medical image data, optimize the resolution of cardiac functional areas, and obtain enhanced cardiac medical image data I enhanced .

[0092] In this embodiment, S2 includes the following steps:

[0093] S21. Using expert annotation method to label cardiac medical image data enhanced Perform segmentation and annotation of cardiac functional areas. The cardiac functional areas include the left ventricle, right ventricle, left atrium, right atrium and valve area. The expert-annotated cardiac functional area segmentation mask is represented by a binary mask:

[0094]

[0095] Among them, M(x) represents the segmentation mask value of pixel x, R heart For the cardiac functional area, each cardiac medical image data has a corresponding expert-labeled cardiac functional area segmentation mask M(x);

[0096] S22. Perform quality assessment on the annotated cardiac functional area segmentation mask M(x) and calculate the morphological closure C of the segmented area heart :

[0097]

[0098] Among them, R ref is the expert reference segmentation region, and the |·| table indicates the number of pixels in the region;

[0099] S23. Divide the labeled cardiac medical imaging data that have passed the quality assessment and its corresponding cardiac functional area segmentation mask M(x) into a training set, a validation set, and a test set according to a preset ratio.

[0100] In this embodiment, S3 includes the following steps:

[0101] S31. Construct an improved deep generative adversarial network, which includes a generator G θ , global discriminator D G With the local discriminator D L , generator G θ Realize cardiac functional area segmentation prediction based on encoding and decoding structure;

[0102] S32. Set the generator G θ The input is the preprocessed cardiac medical image data I enhanced , the output is the predicted cardiac functional area segmentation mask

[0103] S33. Setting the global discriminator D G Cardiac functional area segmentation mask for judgment prediction Whether it matches the expert-labeled cardiac functional area segmentation mask M(x);

[0104] S34. Set local discriminator D L Taking the local structure of the cardiac functional area as input, the left ventricle, right ventricle, left atrium, right atrium and valve area are independently discriminated. The local discriminator D L Used to judge the authenticity of local details between the specific cardiac function area segmentation results output by the generator and the corresponding cardiac function area segmentation results annotated by experts.

[0105] In this embodiment, S32 includes the following steps:

[0106] S321. Set the input of the generator to the pre-processed cardiac medical imaging data, and use the encoding-decoding structure to construct the generator. The generator consists of the encoder E θ With decoder D θ Composition, encoder Eθ Extract deep features Z of cardiac medical imaging data;

[0107] S323. Set decoder D θ With encoder E θ The extracted deep feature Z is used as input to generate the segmentation prediction mask of the cardiac functional area

[0108]

[0109] Among them, T j (·) represents the j-th layer transposed convolution operation, γ j is the learning scaling factor of the jth layer, b j is the bias term, σ is the Sigmoid activation function, N d is the number of layers of the decoder;

[0110] S324. Set up a skip connection mechanism to combine feature information of different scales during decoding to maintain the integrity of the boundaries of the cardiac functional area. The skip connection calculation is as follows:

[0111]

[0112] in, is the k-1th layer output of the decoder, is the k-th layer feature map of the encoder, which is connected by the skip connection function S k After processing, it is passed to the decoder, ⊙ and ⊕ represent element-by-element multiplication and element-by-element addition, respectively, which are used to realize the multi-mode fusion of feature information, η k is the equilibrium parameter, c k It is the bias term in the fusion process, and σ1 is used to activate the fusion result to fully retain the boundary and internal structure information of the cardiac functional area.

[0113] In this embodiment, S5 includes the following steps:

[0114] S51. Set the global discriminator D G The discriminant loss function L G as follows:

[0115]

[0116] Among them, M represents the heart function area segmentation mask annotated by experts, Represents the cardiac functional area segmentation prediction mask output by the generator, and Represent the probability distribution of real and generated masks respectively;

[0117] S52. Set local discriminator D L The discriminant loss function LL as follows:

[0118]

[0119] Among them, r represents each specific area in the cardiac functional area, M r Annotate the segmentation mask for the expert in the corresponding area, The local segmentation mask generated by the generator, and Represent the probability distribution of local true and generated masks respectively;

[0120] S53. Define the cardiac function area segmentation evaluation loss L seg Segmentation results output by the quantized generator The difference between the expert-annotated segmentation result M(x) is:

[0121]

[0122] Among them, ∈ is a constant to prevent division by zero;

[0123] S54. Combine the global adversarial loss, local adversarial loss and cardiac functional area segmentation evaluation loss to define the joint generative adversarial training loss function L GAN for:

[0124] L GAN =λ G L G +λ L L L +λ seg L seg ;

[0125] Among them, λ G ,λ L and λ seg are hyperparameters that regulate the weights of global and local adversarial losses and segmentation losses during the overall training process, so that the generator output gradually approaches the segmentation results annotated by experts;

[0126] S55. Based on joint generation adversarial training loss L GAN The parameters of the generator are optimized using the gradient descent method, while the parameters of the global discriminator and the local discriminator are optimized using the gradient ascent method. The update rules are:

[0127]

[0128] Among them, θ represents the parameters of the generator, φ G With φ L Represent the parameters of the global discriminator and the local discriminator, η θ ,η G and η Lis the corresponding learning rate;

[0129] S56. Repeat steps S51 to S55 until the cardiac function area segmentation prediction mask output by the generator and the segmentation mask M(x) annotated by the expert are in the joint generation adversarial training loss L GAN The convergence is achieved under the constraint of , achieving the adversarial training goal, that is, the generator output gradually approaches the real cardiac functional area segmentation result.

[0130] In this embodiment, S6 includes the following steps:

[0131] S61. Using the generator optimized by adversarial training Perform inference calculation on the pre-processed cardiac medical image data to obtain the cardiac functional area segmentation prediction mask;

[0132] S62. Use the global discriminator after training Evaluate the overall authenticity between the generated cardiac function region segmentation prediction mask and the expert-annotated cardiac function region segmentation mask, and calculate the global discriminant score S G ;

[0133] S63. Use the local discriminator after training Evaluate the local authenticity of the generated cardiac function region segmentation prediction mask on different cardiac function regions and calculate the local segmentation score S L (r);

[0134]

[0135] Among them, R r is the pixel set of the cardiac functional area r, is the segmentation mask of the generator in region r, M r (x) is the true mask of the corresponding area marked by the expert, and Represents the gradient information of the predicted mask and the true mask, which is used to measure the boundary matching degree of the local segmentation result, λ B A weighting factor for scoring local segmentation based on boundary information;

[0136] S64. Calculate the final cardiac function region segmentation confidence score S based on the global discrimination score and each local segmentation score final :

[0137]

[0138] Among them, α G and α L (r) are the weighting factors for global and local segmentation scores, respectively;

[0139] S65. Adjust the cardiac functional area segmentation results output by the adaptive generator When the final cardiac function area segmentation confidence score S final When the value is lower than the set threshold, the segmentation result is optimized:

[0140]

[0141] in, represents the initial cardiac functional area segmentation prediction mask obtained by generator reasoning, M ref (x) is the reference segmentation mask verified with high confidence in history, representing the ideal model of the expert-annotated structure, β is the adaptive adjustment factor, T seg is the preset segmentation confidence threshold, which is used to distinguish high confidence from low confidence segmentation cases, and δ is the sharpness control parameter;

[0142] By constructing a piecewise nonlinear weighted function, when the final cardiac functional region segmentation confidence score is lower than the set threshold, the correction factor approaches 1, thereby correcting the initial cardiac functional region segmentation prediction mask to be closer to the reference segmentation mask. When the final cardiac functional region segmentation confidence score is close to or exceeds the set threshold, the correction factor approaches 0, maintaining the stability of the initial segmentation result, so that the final optimized cardiac functional region segmentation result meets the clinical anatomical structure requirements;

[0143] S66. Optimize the cardiac functional area segmentation results Perform morphological processing, eliminate boundary noise, and convert it into medical visualization format to obtain the final cardiac function area modeling result M final .

[0144] Example 1:

[0145] In 2023, in the cardiovascular department of a well-known domestic tertiary hospital, a doctor encountered a difficult problem during the preoperative evaluation of a patient suspected of valvular heart disease. The patient was a 65-year-old male with a history of hypertension and coronary heart disease. After admission, he underwent MRI and CT cardiac imaging examinations. Due to the low signal-to-noise ratio, poor tissue contrast, and blurred valve area boundaries in the patient's imaging data, traditional image segmentation methods had difficulty accurately identifying the boundaries of the left atrium, left ventricle, mitral valve, and aortic valve. The doctor tried manual correction many times, but still could not achieve the accuracy required for surgical planning.

[0146] Faced with this challenge, the hospital's medical image analysis team decided to adopt the method of the present invention to automatically segment the patient's cardiac images through cardiac functional area modeling based on a deep generative adversarial network, so as to improve the accuracy of image analysis and the reliability of surgical planning.

[0147] Radiology engineers extracted the patient's MRI and CT image data and input the data into the medical imaging AI analysis system. The system first preprocessed the images, including image registration, normalization, noise removal, and contrast enhancement. Among them, the system detected that the patient's CT image resolution was low and performed bicubic interpolation resampling by default to match the MRI resolution.

[0148] After the image data is preprocessed, it is input into the trained GAN segmentation model. The model automatically identifies the functional areas of the heart and begins segmentation calculations.

[0149] The system has preliminarily completed the automatic segmentation of the left atrium, left ventricle, right atrium, and right ventricle. Preliminary results show that the Dice similarity coefficient is approximately 87.6%, but the boundaries of the mitral valve and aortic valve are still blurred.

[0150] The system calls a global-local discriminator to evaluate the segmentation results. The global discriminator performs a consistency check on the overall anatomical structure, while the local discriminator performs a refined analysis of the valve area and finds that the uncertainty of the valve boundary is high, with a confidence score of 71.2% (lower than the threshold of 80%).

[0151] Due to insufficient confidence scores, the system automatically triggers the dynamic confidence adaptive correction mechanism to correct the segmentation of the mitral valve and aortic valve, referencing historical high-confidence segmentation data and adjusting the adversarial loss optimization of the boundary area.

[0152] The system automatically performs morphological optimization on the segmentation results, eliminates small isolated areas, and uses a connectivity filtering algorithm to remove mis-segmented parts of non-cardiac tissue areas.

[0153] The final heart modeling results are stored in the hospital PACS system in DICOM format, and doctors can directly call them in the surgical planning software and perform three-dimensional visualization analysis.

[0154] The imaging analysis team showed the attending physician the AI-automated segmentation modeling results of the cardiac functional areas. The doctor compared them with traditional manual annotation and found that:

[0155] 1. The anatomical structure of the heart's functional area is clearer, and the sharpness of the valve area boundary is improved, making surgical planning more intuitive.

[0156] 2. The system automatically identifies and optimizes the junction area between the left atrium and the mitral valve, avoiding the structural fracture problem that may occur in traditional segmentation methods.

[0157] 3. The analysis time is shortened from 45 minutes to 7 minutes, which greatly improves the doctor's work efficiency and reduces manual correction time.

[0158] To further verify the effectiveness of this invention, we compared it with traditional segmentation methods (U-Net, ResNet) and statistically analyzed the performance of the model on the entire test dataset (60 patients):

[0159]

[0160] In actual cases, the Dice coefficient of the patient's cardiac imaging data in the valve area increased from 85.4% of the traditional method to 92.3%. Doctors confirmed that this result significantly improved the accuracy of preoperative evaluation and formulated a more accurate heart valve replacement plan for the patient.

[0161] This example truly demonstrates the advantages of the present invention in actual hospital clinical applications. Compared with traditional methods, the present invention has higher automatic segmentation accuracy in complex cardiac functional areas (especially valve areas), and at the same time has stronger cross-modal adaptability and computational efficiency, helping doctors obtain high-quality cardiac function modeling results in a short period of time, and providing strong technical support for the accurate diagnosis and treatment of cardiovascular diseases.

[0162] Based on the traditional generative adversarial network, the present invention introduces a dual adversarial mechanism of global discriminator and local discriminator, so that the generator not only optimizes the segmentation accuracy of the overall cardiac functional area during the learning process, but also maintains a higher boundary retention ability in key cardiac anatomical areas (valve, cardiac chamber junction). The global discriminator is responsible for judging the authenticity of the segmentation mask of the overall cardiac functional area, while the local discriminator adopts a regional refinement strategy to independently evaluate the left ventricle, right ventricle, left atrium, right atrium and valve areas, thereby ensuring that the segmentation results are more accurate in the structural details of each cardiac functional area. By optimizing the loss function to balance the global and local segmentation accuracies, the model can improve the accuracy of local segmentation while ensuring the consistency of the overall anatomical structure.

[0163] The present invention introduces a multi-scale skip connection mechanism into the generator structure, so that cardiac anatomical features of different scales can be fully integrated during the segmentation process, ensuring highly consistent segmentation effects under different imaging modalities. The role of the skip connection is to combine the coding features of different levels during the decoding process so that the model can capture the global information of the cardiac functional area while retaining the fine structure of the local boundary. Through this mechanism, the generator can effectively align the feature distributions of images of different modalities when cross-modal images are input, thereby improving the generalization ability of the model and reducing the segmentation deviation caused by different imaging modalities.

[0164] In the post-segmentation processing stage, the present invention proposes an adaptive optimization method based on dynamic confidence. The segmentation confidence scores calculated by global and local discriminators are used to adaptively correct low-confidence segmentation results. When the confidence of the segmentation result is lower than the set threshold, the current segmentation mask is dynamically adjusted by referring to the historical high-confidence segmentation results to make it closer to the actual anatomical structure. In addition to model training, the reliability of the segmentation results is further improved, the sensitivity of the model to noise data is reduced, and the structural rationality of the cardiac functional area is ensured.

[0165] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A cardiac functional area modeling method based on deep generative adversarial network, characterized in that: The steps include: S1. Collect cardiac medical image data and perform preprocessing of the collected cardiac medical image data including image registration, normalization, noise removal, and contrast enhancement; S2. Use expert annotation methods to segment and annotate cardiac functional regions in the preprocessed cardiac medical image data to construct a cardiac medical image dataset; S3. Build an improved deep generative adversarial network based on cardiac medical imaging dataset; S4. Inputting the preprocessed cardiac medical image data into the improved deep generative adversarial network, the generator of the improved deep generative adversarial network outputs a cardiac functional region segmentation prediction result, and inputting the corresponding expert-annotated cardiac functional region segmentation result in the cardiac medical image data and the cardiac functional region segmentation prediction result output by the generator into the global discriminator and the local discriminator respectively for authenticity discrimination; S5. Perform adversarial training on the generator, the global discriminator, and the local discriminator according to the discrimination results output by the global discriminator and the local discriminator of the improved deep generative adversarial network, and optimize the network parameters of the generator and the discriminator; S6. Perform inference calculation on the pre-processed cardiac medical image data, output the cardiac functional area segmentation result, and post-process the cardiac functional area segmentation result based on the discrimination results of the global discriminator and the local discriminator to generate the final cardiac functional area modeling result.

2. A cardiac functional area modeling method based on deep generative adversarial network according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect cardiac medical image data, wherein the resolution of the cardiac medical image data is set to W*H, where W is the width of the cardiac medical image and H is the height of the cardiac medical image; S12. Performing rigid and non-rigid transformations on the acquired cardiac medical image data, wherein the rigid transformations include rotation, scaling, and translation. The non-rigid transformations are optimized for pixel-level alignment using a deformation field, so that the cardiac medical image data processed by the deformation field are anatomically aligned. S13. Normalizing the registered cardiac medical image data to a range of [0, 1]. S14. De-noise the normalized cardiac medical image data, perform contrast enhancement on the de-noised cardiac medical image data, optimize the resolution of cardiac functional areas, and obtain enhanced cardiac medical image data I enhanced .

3. The cardiac functional area modeling method based on deep generative adversarial network according to claim 1, characterized in that: The S2 comprises the following steps: S21. Using expert annotation method to label cardiac medical image data enhanced Perform segmentation and annotation of cardiac functional regions, including the left ventricle, right ventricle, left atrium, right atrium, and valve regions. The expert-annotated cardiac functional region segmentation mask is represented by a binary mask: Among them, M(x) represents the segmentation mask value of pixel x, R heart For the cardiac functional area, each cardiac medical image data has a corresponding expert-labeled cardiac functional area segmentation mask M(x); S22. Perform quality assessment on the annotated cardiac functional area segmentation mask M(x) and calculate the morphological closure C of the segmented area heart : Among them, R ref is the expert reference segmentation region, and the |·| table indicates the number of pixels in the region; S23. Divide the labeled cardiac medical imaging data that have passed the quality assessment and its corresponding cardiac functional area segmentation mask M(x) into a training set, a validation set, and a test set according to a preset ratio.

4. The cardiac functional area modeling method based on deep generative adversarial network according to claim 1, characterized in that: The S3 includes the following steps: S31. Construct an improved deep generative adversarial network, the deep generative adversarial network includes a generator G θ , global discriminator D G With the local discriminator D L , the generator G θ Realize cardiac functional area segmentation prediction based on encoding and decoding structure; S32. Set the generator G θ The input is the preprocessed cardiac medical image data I enhanced , the output is the predicted cardiac functional area segmentation mask S33. Setting the global discriminator D G Cardiac functional area segmentation mask for judgment prediction Whether it matches the expert-labeled cardiac functional area segmentation mask M(x); S34. Set local discriminator D L Taking the local structure of the cardiac functional area as input, the left ventricle, right ventricle, left atrium, right atrium and valve area are independently discriminated. The local discriminator D L Used to judge the authenticity of local details between the specific cardiac function area segmentation results output by the generator and the corresponding cardiac function area segmentation results annotated by experts.

5. The cardiac functional area modeling method based on deep generative adversarial network according to claim 4, characterized in that: The S32 includes the following steps: S321. Set the input of the generator to the pre-processed cardiac medical imaging data, and construct the generator using the encoding-decoding structure. The generator consists of the encoder E θ With decoder D θ Composition, encoder E θ Extract deep features Z of cardiac medical imaging data; S323. Set decoder D θ With encoder E θ The extracted deep feature Z is used as input to generate the segmentation prediction mask of the cardiac functional area Among them, T j (·) represents the j-th layer transposed convolution operation, γ j is the learning scaling factor of the jth layer, b j is the bias term, σ is the Sigmoid activation function, N d is the number of layers of the decoder; S324. Set up a skip connection mechanism to combine feature information of different scales during decoding to maintain the integrity of the boundaries of the cardiac functional area. The skip connection calculation is as follows: in, is the k-1th layer output of the decoder, is the k-th layer feature map of the encoder, which is connected by the skip connection function S k After processing, it is passed to the decoder and Represent element-by-element multiplication and element-by-element addition, respectively, to achieve multi-modal fusion of feature information, η k is the equilibrium parameter, c k It is the bias term in the fusion process, and σ1 is used to activate the fusion result to fully retain the boundary and internal structure information of the cardiac functional area.

6. The cardiac functional area modeling method based on deep generative adversarial network according to claim 1, characterized in that: The S5 comprises the following steps: S51. Set the global discriminator D G The discriminant loss function L G as follows: Among them, M represents the heart function area segmentation mask annotated by experts, Represents the cardiac functional area segmentation prediction mask output by the generator, and Represent the probability distribution of real and generated masks respectively; S52. Set local discriminator D L The discriminant loss function L L as follows: Among them, r represents each specific area in the cardiac functional area, M r Annotate the segmentation mask for the expert in the corresponding area, The local segmentation mask generated by the generator, and Represent the probability distribution of local true and generated masks respectively; S53. Define the cardiac function area segmentation evaluation loss L seg Segmentation results output by the quantized generator The difference between the expert-annotated segmentation result M(x) is: Among them, ∈ is a constant to prevent division by zero; S54. Combine the global adversarial loss, local adversarial loss and cardiac functional area segmentation evaluation loss to define the joint generative adversarial training loss function L GAN for: L GAN =λ G L G +λ L L L +λ seg L seg ; Among them, λ G ,λ L and λ seg are hyperparameters that regulate the weights of global and local adversarial losses and segmentation losses during the overall training process, so that the generator output gradually approaches the segmentation results annotated by experts. S55. Based on joint generation adversarial training loss L GAN The parameters of the generator are optimized using the gradient descent method, while the parameters of the global discriminator and the local discriminator are optimized using the gradient ascent method. The update rules are: Among them, θ represents the parameters of the generator, φ G With φ L Represent the parameters of the global discriminator and the local discriminator, η θ ,η G and η L is the corresponding learning rate; S56. Repeat steps S51 to S55 until the cardiac function area segmentation prediction mask output by the generator and the segmentation mask M(x) annotated by the expert are in the joint generation adversarial training loss L GAN The convergence is achieved under the constraint of , achieving the adversarial training goal, that is, the generator output gradually approaches the real cardiac functional area segmentation result.

7. The cardiac functional area modeling method based on deep generative adversarial network according to claim 1, characterized in that: The S6 comprises the following steps: S61. Using the generator optimized by adversarial training Perform inference calculation on the pre-processed cardiac medical image data to obtain the cardiac functional area segmentation prediction mask; S62. Use the global discriminator after training Evaluate the overall authenticity between the generated cardiac function region segmentation prediction mask and the expert-annotated cardiac function region segmentation mask, and calculate the global discriminant score S G ; S63. Use the local discriminator after training Evaluate the local authenticity of the generated cardiac function region segmentation prediction mask on different cardiac function regions and calculate the local segmentation score S L (r); Among them, R r is the pixel set of the cardiac functional area r, is the segmentation mask of the generator in region r, M r (x) is the true mask of the corresponding area marked by the expert, and Represents the gradient information of the predicted mask and the true mask, which is used to measure the boundary matching degree of the local segmentation result, λ B A weighting factor for scoring local segmentation based on boundary information; S64. Calculate the final cardiac function region segmentation confidence score S based on the global discrimination score and each local segmentation score final : Among them, α G and α L (r) are the weighting factors for global and local segmentation scores, respectively; S65. Adjust the cardiac functional area segmentation results output by the adaptive generator When the final cardiac function area segmentation confidence score S final When the value is lower than the set threshold, the segmentation result is optimized: in, represents the initial cardiac functional area segmentation prediction mask obtained by generator reasoning, M ref (x) is the reference segmentation mask verified with high confidence in history, representing the ideal model of the expert-annotated structure, β is the adaptive adjustment factor, T seg is the preset segmentation confidence threshold, which is used to distinguish high confidence from low confidence segmentation cases, and δ is the sharpness control parameter; By constructing a piecewise nonlinear weighted function, when the final cardiac functional region segmentation confidence score is lower than the set threshold, the correction factor approaches 1, thereby correcting the initial cardiac functional region segmentation prediction mask to be closer to the reference segmentation mask. When the final cardiac functional region segmentation confidence score is close to or exceeds the set threshold, the correction factor approaches 0, maintaining the stability of the initial segmentation result, so that the final optimized cardiac functional region segmentation result meets the clinical anatomical structure requirements; S66. Optimize the cardiac functional area segmentation results Perform morphological processing, eliminate boundary noise, and convert it into medical visualization format to obtain the final cardiac function area modeling result M final .

Images