Robust placenta implantation detection method based on semi-supervised multi-scale generative adversarial network

Abstract

The invention discloses a robust placenta implantation detection method based on a semi-supervised multi-scale generative adversarial network. The method comprises the following steps: step 1, establishing a placenta implantation image diagnosis data set; and 2, establishing a placenta implantation image diagnosis model based on the semi-supervised generative adversarial network. And step 3, multi-scale feature extraction. 4, carrying out semi-supervised training; and 5, applying to automatic diagnosis of placenta implantation. Aiming at the problems of strong expert experience, insufficient sample data and weak model robustness in placenta implantation detection, the method combines semi-supervised learning and a multi-scale generative adversarial network, generates a high-quality image sample through a small amount of labeled data and a large amount of unlabeled data, and enhances the discrimination capability of the model through multi-scale feature extraction. According to the method, the accuracy of placenta implantation detection and the generalization ability of the model can be effectively improved, the dependence on labeled data is reduced, the sample efficiency is enhanced, the expert intervention requirement is reduced, and a more reliable and accurate intelligent tool is provided for clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to a robust placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network, and belongs to the technical field of artificial intelligence. Background Art

[0002] Placenta accreta detection is crucial in obstetrics, because abnormal placenta accreta may lead to serious complications such as postpartum hemorrhage, hysterectomy, and even life-threatening complications for mothers and infants. Therefore, early and accurate identification of the type and severity of placenta accreta can not only optimize the delivery plan, but also significantly reduce the risks in obstetric surgery and ensure the safety of mothers and infants. However, current detection methods face multiple technical difficulties: insufficient image resolution makes it difficult to accurately identify early or mild placenta accreta, and inconsistent diagnostic standards can easily lead to inconsistent diagnostic results between different doctors, increasing the possibility of misdiagnosis and missed diagnosis. In addition, individual differences further increase the difficulty of diagnosis. The depth, location, and maternal uterine wall characteristics of placenta accreta all affect the effectiveness of image analysis, and conventional detection methods are not suitable for comprehensive application during pregnancy due to potential risks. In response to these problems, artificial intelligence (AI) technology has been gradually introduced in recent years to assist placenta accreta detection with the help of deep learning and intelligent image analysis. However, AI is not yet fully mature in placenta accreta detection, and the main bottlenecks are the lack of labeled data and insufficient model robustness. High-quality labeled data for placenta implantation detection requires manual labeling by experienced doctors, which is not only a cumbersome process, but also requires extremely high labeling accuracy due to the variety of placenta types and high image complexity. In addition, due to insufficient data volume, existing AI models have poor generalization capabilities when dealing with physiological differences between different individuals and are easily affected by abnormal features or slight implantation variations, resulting in insufficient stability and reliability of the model in clinical applications.

[0003] With the rapid development of deep learning, large-scale labeled data has become the key to training complex models. However, in practical application scenarios, obtaining sufficient labeled data is often costly, time-consuming and laborious. Semi-supervised learning (SSL) is a compromise solution that improves the training efficiency and generalization ability of the model by combining a small amount of labeled data with a large amount of unlabeled data. Semi-supervised learning methods rely on the model's ability to effectively extract features from unlabeled data and guide the model's learning through certain rules or objective functions.

[0004] In recent years, semi-supervised learning technology has made significant progress. Representative methods include consistency regularization, semi-supervised learning assisted by generative models, and semi-supervised methods based on self-supervised learning. The consistency regularization method effectively utilizes unlabeled data by introducing perturbations so that the model has consistent output for small changes in the input samples. Semi-supervised learning assisted by generative models uses generative models such as generative adversarial networks (GANs) to generate pseudo labels for unlabeled data, thereby improving the model's discriminative ability. Semi-supervised methods based on self-supervised learning enable the model to learn effective representations from unlabeled data by designing pre-training tasks, thereby improving the performance of downstream tasks.

[0005] Generative Adversarial Networks (GANs) is a generative model proposed by Ian Goodfellow et al. in 2014. GANs consists of a generator and a discriminator, where the generator attempts to generate fake samples that are consistent with the distribution of real data, while the discriminator attempts to distinguish between real samples and generated samples. During the adversarial training process, the generator continuously improves its generation ability to deceive the discriminator so that the generated samples gradually approach the distribution of real samples. GANs have been widely used and self-supervised performance verified in many fields, such as image generation, data enhancement, style transfer, etc. GANs networks have high-precision task performance. In particular, in medical image analysis, GANs are used to synthesize medical images with high fidelity to assist the problem of insufficient data. For example, GANs can be used to generate synthetic images of lesion areas for data augmentation or training more robust classifiers. In addition, multi-scale generative adversarial networks can better capture the detailed information of images by generating and discriminating images at different scales, thereby improving the quality of generated samples. Therefore, how to develop a new placenta accreta detection method by combining semi-supervised learning and GANs. This method uses generative adversarial networks to generate realistic unlabeled image samples based on a small amount of labeled data, and only requires a small amount of expert labeled data to complete the intelligent diagnosis model training, thereby making it sample efficient. Summary of the invention

[0006] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a robust placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network. This method addresses the problems of strong expert experience, insufficient sample data, and weak model robustness in placenta accreta detection. It combines semi-supervised learning with a multi-scale generative adversarial network to generate high-quality image samples through a small amount of labeled data and a large amount of unlabeled data, and enhances the model's discrimination ability through multi-scale feature extraction. The present invention can effectively improve the accuracy of placenta accreta detection and the generalization ability of the model, reduce dependence on labeled data, enhance sample efficiency, reduce the need for expert intervention, and provide a more reliable and accurate intelligent tool for clinical diagnosis.

[0007] The technical solution adopted by the present invention to solve the technical problem is: a placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network, which mainly includes the following steps:

[0008] Step 1: Establish a placenta accreta imaging diagnosis dataset.

[0009] Step 1-1: Dataset construction.

[0010] The Department of Imaging collected and saved the desensitized placenta accreta image data of real patients in DICOM (Digital Imaging and Communications in Medicine) magnetic resonance imaging format. The experts (chief physicians) related to placenta accreta combined real clinical information and diagnosis and treatment experience to summarize and set the key fields for disease diagnosis, and manually annotated a small number of placenta accreta image samples to obtain the expert annotation dataset D l , used for model classification ability enhancement training. Other large amounts of unlabeled placenta implantation data are used as unlabeled datasets D u , used for unsupervised model training. Unlabeled dataset D u The size of the expert annotation dataset D l size.

[0011] Step 1-2: Data preprocessing.

[0012] In order for the neural network to read and recognize medical imaging data, the DICOM image data needs to be further processed. First, read the DICOM file to extract the image data, and retain the information in the DICOM file other than the image pixels, such as age, weight, etc., according to the key fields determined by the experts. Then, the pixel values ​​are linearly converted according to the "Rescale Slope" and "Rescale Intercept" fields in the DICOM file. Due to the different resolutions of different scanning devices, the image is further resampled so that the images obtained from different scans have consistent physical dimensions in space. The image can be adjusted to a uniform pixel size through interpolation technology to ensure consistent image quality in subsequent processing.

[0013] Magnetic resonance imaging is sliced ​​to obtain 2D placenta implantation image slices. The 2D image is cropped and resized to meet the input requirements of the GAN network. Center cropping is usually used to retain the placenta implantation area, maintain the aspect ratio of the image, and avoid deformation. Pixel values ​​are usually standardized to [0,1]. In order to further enhance the generalization ability of the model, data augmentation is often performed on the image. In order to enhance sample efficiency and model generalization ability, the image is processed through geometric transformation (such as rotation, flipping, scaling, etc.) and brightness contrast adjustment to make the model more robust to different imaging conditions. Finally, the preprocessed placenta implantation image data is stored as a tensor file to speed up subsequent data reading and network training to improve network training efficiency.

[0014] Step 2: Establish a placenta accreta imaging diagnosis model based on a semi-supervised generative adversarial network.

[0015] The generative adversarial network consists of a generator G and a discriminator D. The generator G(z) can receive input noise z~p z (z) Generated and real placenta implantation data distribution p d Similar placenta accreta image samples, the goal is to make the generated placenta accreta samples as similar as possible to the real placenta accreta, so that they cannot be identified by the discriminator. In the absence of sufficient expert-annotated data, clinical medicine requires the intervention of an expert team. The generator can generate realistic new data through limited annotation learning to achieve the effect of enhancing the generalization ability of the model, so as to achieve sample-efficient placenta accreta detection.

[0016] Step 2-1: Design the generator G. The goal of the generator is to deceive the discriminator so that the discriminator cannot distinguish whether the patient is sick based on the placenta implantation data. The loss function of the generator is to maximize the probability that the discriminator considers the generated sample to be a real sample. The present invention uses a binary classification adversarial loss to evaluate the performance of the generator:

[0017]

[0018] Step 2-2: Design the discriminator D. The discriminator D(x) aims to distinguish the diseased placenta implantation data x~p d (x) and normal placenta implantation data x~G(z).

[0019] For labeled placenta implantation data, the discriminator must not only judge whether it is real, but also predict the correct category to which it belongs, and use cross entropy loss to minimize the classification error of the discriminator for real samples:

[0020]

[0021] For unlabeled samples, the discriminator classifies them as real samples or generated samples. At this time, the adversarial loss function of the discriminator is

[0022]

[0023] Step 2-3: Jointly train the generator and discriminator. The adversarial training objective function of the generator and discriminator is:

[0024]

[0025] Step 3: Multi-scale feature extraction.

[0026] Multi-scale feature extraction can capture and analyze structural features of different sizes and levels in medical images, enabling the model to better understand and characterize complex medical image data, process medical images with unevenly distributed lesion areas, enhance the robustness to images of different resolutions, and enable the model to have higher accuracy in lesion detection and classification tasks.

[0027] Step 3-1: In order to more accurately capture the multi-scale features of medical images, perform multi-scale feature decomposition on the image data. Assume that the feature representation of the image at different resolutions is Where s represents different scale levels (from global structure to local details) and S is the total number of scales. The generator G generates image samples at each scale. Where z is the input noise vector or potential vector. Based on this, the discriminator D s The generated image samples are evaluated separately at each scale s to judge their authenticity and quality. Therefore, the discriminator is able to perceive and identify the placenta accreta lesion area at multiple scales simultaneously, ensuring a comprehensive evaluation of generated samples at different resolutions, so that it can distinguish between real medical images and generated fake samples.

[0028] In the multi-scale placenta accreta image recognition task, the loss function of the GAN-based semi-supervised diagnosis model can be further expanded to a multi-scale loss. And the loss function of the discriminator It can be expressed as the weighted sum of multiple scale losses:

[0029]

[0030] Among them, λ s Represents the weight of scale s.

[0031] The loss function of the joint training of the semi-supervised diagnosis model based on the GAN network can be further expanded as follows:

[0032]

[0033] Through a multi-scale generative network, the global and local information of placenta implantation images can be captured, improving the model's generation, discrimination and diagnostic performance.

[0034] Step 4: Perform semi-supervised training, combining limited labeled data with a large amount of unlabeled data, and optimize model performance through consistency regularization and pseudo-label generation mechanism of generative adversarial network. In the semi-supervised learning framework, not only supervised learning is performed on labeled data, but also reasonable pseudo-label generation and model training are performed on unlabeled data to maximize the use of data.

[0035] Step 4-1: Use consistency regularization to enhance the robustness of the model to input perturbations. By adding a small perturbation ∈ to the input sample x, the model ensures the consistency of the output before and after the perturbation, thereby improving the generalization ability and stability of the model. Let the perturbed sample be x+∈, then the goal of consistency regularization is to minimize the difference between the model's output for the original sample x and the perturbation sample x+∈. Formally, the consistency regularization loss function is It can be defined as:

[0036]

[0037] Among them, G(x) represents the model's predicted output for the input sample x, G(x+∈) represents the model's predicted output for the perturbation sample x+∈, Dl is the labeled data set, and the perturbation ∈ obeys Gaussian random noise with small variance. By minimizing the loss The model will maintain output consistency for small perturbations, improving its robustness on unknown samples.

[0038] Step 4-2: Next, use the generative adversarial network to train the unlabeled data D u Generate pseudo labels. The generator G of GAN generates pseudo labels for unlabeled data, and the discriminator D guides the optimization of the generator by distinguishing between real samples and generated samples. Specifically, the sample D generated by the generator u Through the discriminator Du Evaluation is performed to ensure that the generated pseudo-label samples are indistinguishable from real samples. Loss function for pseudo-label generation It can be expressed as:

[0039]

[0040] Among them, G(x) represents the sample generated by the generator on the unlabeled data x, and the discriminator D judges the authenticity of the sample through its classification task. By generating pseudo labels for unlabeled data and training them, the generator can be further optimized to generate more realistic samples and improve the quality of pseudo labels.

[0041] Step 4-3: After pseudo-label generation is completed, the labeled data and unlabeled data are jointly trained to optimize the model parameters. The loss function of joint training consists of supervised learning loss, pseudo-label loss, and consistency regularization loss. The joint loss function can be expressed as:

[0042] L total =L GAN +λ 1 L consistencty +λ 2 L semi ,

[0043] where λ 1 and λ 2 is a hyperparameter that controls the weight of pseudo-label loss and consistency regularization loss. By optimizing this joint loss, the model can learn effective diagnostic features from labeled data and generate effective pseudo-labels from unlabeled data to improve overall performance, thereby improving diagnostic accuracy and robustness under semi-supervised learning. The model parameters are optimized by combining labeled data and pseudo-label data for joint training.

[0044] Step 5: Automated diagnosis for placenta accreta.

[0045] After the model training is completed, it is applied to the automated diagnosis task of clinical placenta accreta. Specifically, the medical image sample x to be diagnosed is first new The input is fed into the trained generative adversarial network, where the generator G and the discriminator D extract and evaluate the image features at multiple scales. The generator generates a feature representation that conforms to the anatomical structure of the placenta accreta based on the input image, and the discriminator classifies and evaluates the generated features to determine whether they conform to the pathological pattern of the placenta accreta. The model captures the pathological information of the placenta accreta by generating pseudo images or high-dimensional feature expressions, and outputs a preliminary diagnosis result. The output result of the GAN network G(x new ) can be expressed as a probability distribution p(y|x new), where y is the diagnostic label of placenta implantation, with a value of normal or abnormal. The generator G generates features using new image samples and passes them to the discriminator D, which performs multi-scale and multi-level authenticity assessments. In this way, the model can give a diagnosis probability based on pathological features for each input sample, and the diagnosis result can be formalized as:

[0046]

[0047] in This is the final automated diagnosis result given by the model. This diagnosis result is based on the model's comprehensive analysis of the imaging features of the placenta accreta, covering global and local information as well as multi-scale features. Since the model has been trained with a large amount of labeled and unlabeled data, it has strong generalization ability and can effectively identify the pathological differences of placenta accreta between different individuals.

[0048] In order to improve the accuracy of diagnosis, the results output by the model must also be combined with the doctor's clinical observations and medical experience. During the film reading process, the doctor can refer to the diagnostic suggestions provided by the model, and combine the patient's specific clinical symptoms, medical history, and other auxiliary examination results to give a final comprehensive diagnosis. Therefore, this automated diagnostic system can not only greatly improve the efficiency of image analysis, but also provide doctors with valuable auxiliary diagnostic information, especially in the screening and early detection of placenta accreta, which can significantly reduce the workload of doctors and improve the accuracy and consistency of the overall diagnosis.

[0049] Beneficial effects:

[0050] 1. The present invention improves the robustness and effectiveness of the intelligent diagnosis model through multi-scale feature learning, thereby solving the problems of traditional diagnostic methods that are highly professional, difficult to quantify the condition, and difficult to accurately identify. It can also improve the accuracy of diagnosis, reduce dependence on labeled data, enhance data efficiency, and provide a more effective intelligent placenta accreta detection tool for clinical practice.

[0051] 2. The present invention is aimed at the field of placenta accreta detection and adopts semi-supervised learning and multi-scale generative adversarial networks for the first time. It can utilize a small amount of clinical data annotated by experts and a large amount of unlabeled data, effectively reducing the dependence of the placenta accreta intelligent diagnosis model on expert experience and knowledge during the training stage, enhancing the efficiency of clinical samples, and effectively improving the accuracy of placenta accreta detection and the robustness of the intelligent model.

[0052] 3. In the placenta implant feature extraction stage, the present invention adopts multi-scale feature extraction, which can effectively capture the different detail levels of the placenta implant image at different resolutions, and significantly improves the recognition ability and accuracy of complex lesion areas through multi-scale placenta implant detail fusion.

[0053] 4. In the placenta implantation detection stage, the present invention adopts a generative adversarial network, which can effectively solve the problem of insufficient expert data, reduce the dependence on a large amount of labeled data and expert teams, and significantly reduce the model training overhead.

[0054] 5. The simulation results show that the placenta accreta detection method based on semi-supervised multi-scale generative adversarial network proposed in the present invention has good generalization ability and efficient diagnostic performance in the placenta accreta detection task, and is ahead of the benchmark solution in diagnostic accuracy. Therefore, the automated diagnosis system can not only greatly improve the efficiency of image analysis, but also provide doctors with valuable auxiliary diagnosis information, especially in the early confirmation process of placenta accreta, which can significantly reduce the workload of doctors and improve the accuracy and consistency of the overall diagnosis.

[0055] 6. In order to improve the accuracy of diagnosis, the results output by the model must also be combined with the doctor's clinical observation and medical experience. During the film reading process, the doctor can refer to the diagnostic suggestions provided by the model, and combine the patient's specific clinical symptoms, medical history, and other auxiliary examination results to give a final comprehensive diagnosis. Therefore, this automated diagnostic system can not only greatly improve the efficiency of image analysis, but also provide doctors with valuable auxiliary diagnostic information, especially in the screening and early detection of placenta implantation, which can significantly reduce the workload of doctors and improve the accuracy and consistency of the overall diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a block diagram of the placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network proposed in the present invention.

[0057] Figure 2 This is the DICOM placenta accreta image data display and feature preprocessing display used in the present invention.

[0058] Figure 3 This is a graph showing the loss convergence of the placenta implantation detection method based on the semi-supervised multi-scale generative adversarial network proposed in the present invention over the training cycle.

[0059] Figure 4 This is a comparison chart of the diagnostic accuracy of the placenta accreta detection method based on semi-supervised multi-scale generative adversarial network proposed in the present invention and the existing benchmark solution. DETAILED DESCRIPTION

[0060] The invention will be further described in detail below in conjunction with the accompanying drawings.

[0061] Example 1

[0062] like Figure 2As shown, a method for constructing a data set in a robust placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network in this embodiment comprises the following steps:

[0063] Step 1: Dataset construction.

[0064] The Department of Imaging collected and saved the desensitized placenta accreta image data of real patients in DICOM (Digital Imaging and Communications in Medicine) magnetic resonance imaging format. The experts (chief physicians) related to placenta accreta combined real clinical information and diagnosis and treatment experience to summarize and set the key fields for disease diagnosis, and manually annotated a small number of placenta accreta image samples to obtain the expert annotation dataset D l , used for model classification ability enhancement training. Other large amounts of unlabeled placenta implantation data are used as unlabeled datasets D u , used for unsupervised model training. Unlabeled dataset D u The size of the expert annotation dataset D l size.

[0065] Step 2: Data preprocessing.

[0066] In order for the neural network to read and recognize medical imaging data, the DICOM image data needs to be further processed. First, read the DICOM file to extract the image data, and retain the information in the DICOM file other than the image pixels, such as age, weight, etc., according to the key fields determined by the experts. Then, the pixel values ​​are linearly converted according to the "Rescale Slope" and "Rescale Intercept" fields in the DICOM file. Due to the different resolutions of different scanning devices, the image is further resampled so that the images obtained from different scans have consistent physical dimensions in space. The image can be adjusted to a uniform pixel size through interpolation technology to ensure consistent image quality in subsequent processing.

[0067] Slice processing obtains 2D placenta implantation image slices. The 2D image is cropped and resized to meet the input requirements of the GAN network. Center cropping is usually used to retain the placenta implantation area, maintain the aspect ratio of the image, and avoid deformation. The pixel values ​​are usually standardized to [0,1]. In order to further enhance the generalization ability of the model, data augmentation is often performed on the image. In order to enhance sample efficiency and model generalization ability, the image is processed through geometric transformation (such as rotation, flipping, scaling, etc.) and brightness contrast adjustment, so that the model has stronger robustness to different imaging conditions. Finally, the preprocessed placenta implantation image data is stored as a tensor file to speed up subsequent data reading and network training to improve network training efficiency.

[0068] Example 2

[0069] A robust placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network in this embodiment establishes a placenta accreta imaging diagnosis model based on a semi-supervised generative adversarial network. The method includes the following steps:

[0070] The generative adversarial network consists of a generator G and a discriminator D. The generator G(z) can receive input noise z~p z (z) Generated and real placenta implantation data distribution p d Similar placenta accreta image samples, the goal is to make the generated placenta accreta samples as similar as possible to the real placenta accreta, so that they cannot be identified by the discriminator. In the absence of sufficient expert-annotated data, clinical medicine requires the intervention of an expert team. The generator can generate realistic new data through limited annotation learning to achieve the effect of enhancing the generalization ability of the model, so as to achieve sample-efficient placenta accreta detection.

[0071] Step 1: Design the generator G. The goal of the generator is to deceive the discriminator so that the discriminator cannot distinguish whether the patient is sick based on the placenta implantation data. The loss function of the generator is to maximize the probability that the discriminator considers the generated sample to be a real sample. The present invention uses a binary classification adversarial loss to evaluate the performance of the generator:

[0072]

[0073] Step 2: Design the discriminator D. The discriminator D(x) aims to distinguish the diseased placenta implantation data x~p d (x) and normal placenta implantation data x~G(z).

[0074] For labeled placenta implantation data, the discriminator must not only judge whether it is real, but also predict the correct category to which it belongs, and use cross entropy loss to minimize the classification error of the discriminator for real samples:

[0075]

[0076] For unlabeled samples, the discriminator classifies them as real samples or generated samples. At this time, the adversarial loss function of the discriminator is

[0077]

[0078] Step 3: Jointly train the generator and discriminator. The adversarial training objective function of the generator and discriminator is:

[0079]

[0080] Example 3

[0081] A robust placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network in this embodiment is based on Example 2 to establish a multi-scale placenta accreta image feature extraction method, which includes the following steps:

[0082] The present invention utilizes multi-scale feature extraction to capture and analyze structural features of different sizes and levels in medical images, enabling the model to better understand and characterize complex medical image data, process medical images with unevenly distributed lesion areas, enhance the robustness to images of different resolutions, and enable the model to have higher accuracy in lesion detection and classification tasks.

[0083] Step 1: In order to more accurately capture the multi-scale features of medical images, perform multi-scale feature decomposition on the image data. Assume that the feature representation of the image at different resolutions is Where x represents different scale levels (from global structure to local details), and S is the total number of scales. The generator G generates image samples at each scale. Where z is the input noise vector or potential vector. Based on this, the discriminator D s The generated image samples are evaluated separately at each scale s to judge their authenticity and quality. Therefore, the discriminator is able to perceive and identify the placenta accreta lesion area at multiple scales simultaneously, ensuring a comprehensive evaluation of generated samples at different resolutions, so that it can distinguish between real medical images and generated fake samples.

[0084] In the multi-scale placenta accreta image recognition task, the loss function of the GAN-based semi-supervised diagnosis model can be further expanded to a multi-scale loss. And the loss function of the discriminator It can be expressed as the weighted sum of multiple scale losses:

[0085]

[0086]

[0087] Among them, λ s Represents the weight of scale s.

[0088] The loss function of the joint training of the semi-supervised diagnosis model based on the GAN network can be further expanded as follows:

[0089]

[0090] Through a multi-scale generative network, the global and local information of placenta implantation images can be captured, improving the model's generation, discrimination and diagnostic performance.

[0091] Example 4

[0092] A robust placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network in this embodiment is based on Example 3 and performs semi-supervised learning. The method includes the following steps:

[0093] Semi-supervised training is performed, combining limited labeled data with a large amount of unlabeled data, and optimizing model performance through consistency regularization and pseudo-label generation mechanism of generative adversarial networks. In the semi-supervised learning framework, not only supervised learning is performed on labeled data, but also reasonable pseudo-label generation and model training are performed on unlabeled data to maximize the use of data.

[0094] Step 1: Use consistency regularization to enhance the robustness of the model to input perturbations. By adding a small perturbation ∈ to the input sample x, the model ensures the consistency of the output before and after the perturbation, thereby improving the generalization ability and stability of the model. Let the perturbed sample be x+∈, then the goal of consistency regularization is to minimize the difference between the model's output for the original sample x and the perturbation sample x+∈. Formally, the consistency regularization loss function is It can be defined as:

[0095]

[0096] Among them, G(x) represents the model's predicted output for the input sample x, G(x+∈) represents the model's predicted output for the perturbation sample x+∈, and D l is a labeled dataset, and the perturbation ∈ is a Gaussian random noise with small variance. By minimizing the loss The model will maintain output consistency for small perturbations, improving its robustness on unknown samples.

[0097] Step 2: Generate adversarial networks to train unlabeled data D u Generate pseudo labels. The generator G of GAN generates pseudo labels for unlabeled data, and the discriminator D guides the optimization of the generator by distinguishing between real samples and generated samples. Specifically, the sample D generated by the generatoru Through the discriminator D u Evaluation is performed to ensure that the generated pseudo-label samples are indistinguishable from real samples. Loss function for pseudo-label generation It can be expressed as:

[0098]

[0099] Among them, G(x) represents the sample generated by the generator on the unlabeled data x, and the discriminator D judges the authenticity of the sample through its classification task. By generating pseudo labels for unlabeled data and training them, the generator can be further optimized to generate more realistic samples and improve the quality of pseudo labels.

[0100] Step 3: After pseudo-label generation is completed, the labeled data and unlabeled data are jointly trained to optimize the model parameters. The loss function of joint training consists of supervised learning loss, pseudo-label loss, and consistency regularization loss. The joint loss function can be expressed as:

[0101] L total =L GAN +λ 1 L consistency +λ 2 L semi ,

[0102] where λ 1 and λ 2 is a hyperparameter that controls the weight of pseudo-label loss and consistency regularization loss. By optimizing this joint loss, the model can learn effective diagnostic features from labeled data and generate effective pseudo-labels from unlabeled data to improve overall performance, thereby improving diagnostic accuracy and robustness under semi-supervised learning. The model parameters are optimized by combining labeled data and pseudo-label data for joint training.

[0103] Example 5

[0104] A robust placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network in this embodiment is based on Example 4 and proposes an automated diagnosis method for placenta accreta. The method comprises the following steps:

[0105] After the model training is completed, it is applied to the automated diagnosis task of clinical placenta accreta. Specifically, the medical image sample x to be diagnosed is first newThe input is fed into the trained generative adversarial network, where the generator G and the discriminator D extract and evaluate the image features at multiple scales. The generator generates a feature representation that conforms to the anatomical structure of the placenta accreta based on the input image, and the discriminator classifies and evaluates the generated features to determine whether they conform to the pathological pattern of the placenta accreta. The model captures the pathological information of the placenta accreta by generating pseudo images or high-dimensional feature expressions, and outputs a preliminary diagnosis result. The output result of the GAN network G(x new ) can be expressed as a probability distribution p(y|x new ), where y is the diagnostic label of placenta implantation, with a value of normal or abnormal. The generator G generates features using new image samples and passes them to the discriminator D, which performs multi-scale and multi-level authenticity assessments. In this way, the model can give a diagnosis probability based on pathological features for each input sample, and the diagnosis result can be formalized as:

[0106]

[0107] in This is the final automated diagnosis result given by the model. This diagnosis result is based on the model's comprehensive analysis of the imaging features of the placenta accreta, covering global and local information as well as multi-scale features. Since the model has been trained with a large amount of labeled and unlabeled data, it has strong generalization ability and can effectively identify the pathological differences of placenta accreta between different individuals.

[0108] In order to improve the accuracy of diagnosis, the results output by the model must also be combined with the doctor's clinical observations and medical experience. During the film reading process, the doctor can refer to the diagnostic suggestions provided by the model, and combine the patient's specific clinical symptoms, medical history, and other auxiliary examination results to give a final comprehensive diagnosis. Therefore, this automated diagnostic system can not only greatly improve the efficiency of image analysis, but also provide doctors with valuable auxiliary diagnostic information, especially in the screening and early detection of placenta accreta, which can significantly reduce the workload of doctors and improve the accuracy and consistency of the overall diagnosis.

[0109] The technical effects of the present invention are further described in detail below in combination with simulation experiments and data, including:

[0110] 1. Simulation hardware conditions

[0111] The simulation experiment of the present invention is carried out on the simulation platform of Python 3.7 and TensorFlow 2.7. The CPU model of the computer is Intel Core i7-12700 processor, and the GPU model is NVIDIA Geforce RTX 4070.

[0112] 2. Simulation system parameters

[0113] The training data set consists of 300 2D placenta accreta images, each of which is obtained by annotation by an expert as described in Example 1, and is obtained from healthy subjects using a standard imaging protocol. Among them, the 2D placenta accreta images are obtained by nuclear magnetic resonance imaging slices, and the voxel resolution of the 2D stacked low-resolution image is fixed at 1.25×1.25×10.00mm. The test data set consists of 1200 2D stack short-axis images that have not been processed by experts. Both the training data and the test data have been desensitized and only have the necessary diagnostic information.

[0114] 3. Simulation content

[0115] Figure 3 This is a graph showing the loss convergence of the placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network proposed in the present invention over the training cycle. As shown in the figure, the placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network proposed in the present invention can make full use of the labeled samples and achieve good convergence effect in a shorter training cycle.

[0116] Figure 4 This is a comparison chart of the diagnostic accuracy of the placenta accreta detection method based on the semi-supervised multi-scale generative adversarial network proposed in the present invention and the existing benchmark scheme with the size of expert labeled data. Compared with the existing supervised learning method that relies on labeled samples, the placenta accreta detection method based on the semi-supervised multi-scale generative adversarial network proposed in the present invention has the characteristics of high sample efficiency, can efficiently learn image classification capabilities on a small sample training set, and make full use of large-scale unlabeled data sets to further improve diagnostic efficiency, with a diagnostic accuracy rate of 93.7%. Traditional neural networks based on supervised learning cannot be well trained under limited expert samples.

[0117] Based on the above simulation results and analysis, the placenta accreta detection method based on the semi-supervised multi-scale generative adversarial network proposed in the present invention can utilize a small amount of expert-annotated clinical data and a large amount of unlabeled data compared to other benchmark methods, effectively reducing the dependence of the placenta accreta intelligent diagnosis model on expert experience and knowledge during the training stage, enhancing the clinical sample efficiency, and effectively improving the accuracy of placenta accreta detection and the robustness of the intelligent model. The simulation results show that the placenta accreta detection method based on the semi-supervised multi-scale generative adversarial network proposed in the present invention has good generalization ability and efficient diagnostic performance in the placenta accreta detection task, and is ahead of the benchmark solution in diagnostic accuracy. Therefore, the automated diagnostic system can not only greatly improve the efficiency of image analysis, but also provide doctors with valuable auxiliary diagnostic information, especially in the early detection process of placenta accreta, which can significantly reduce the workload of doctors and improve the accuracy and consistency of the overall diagnosis.

Claims

1. A placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network, characterized in that: The method comprises the following steps: Step 1: Establish a placenta accreta imaging diagnostic dataset; Step 1-1: Dataset construction; Step 1-2: Data preprocessing; Step 2: Establish a placenta accreta imaging diagnosis model based on a semi-supervised generative adversarial network; The generative adversarial network consists of a generator G and a discriminator D. The generator G(z) can receive input noise z~p z (z) Generate placenta implantation image samples similar to the real placenta implantation data distribution pd, so that the generated placenta implantation samples are as similar as possible to the real placenta implantation situation, so that they cannot be recognized by the discriminator, and the generator generates realistic new data through limited annotation learning; Step 2-1: Design generator G; The goal of the generator is to deceive the discriminator so that the discriminator cannot distinguish whether the patient is sick based on the placenta implantation data. The loss function of the generator is to maximize the probability that the discriminator believes that the generated sample is a real sample. The binary classification adversarial loss is used to evaluate the performance of the generator: Step 2-2: Design the discriminator D; The discriminator D(x) aims to distinguish diseased placenta implantation data x~p d (x) and normal placenta implantation data x~G(z); For labeled placenta implantation data, the discriminator must not only judge whether it is real, but also predict the correct category to which it belongs, and use cross entropy loss to minimize the classification error of the discriminator for real samples: For unlabeled samples, the discriminator classifies them as real samples or generated samples. At this time, the adversarial loss function of the discriminator is: Step 2-3: Jointly train the generator and discriminator. The adversarial training objective function of the generator and discriminator is: Step 3: Multi-scale feature extraction; Multi-scale feature extraction can capture and analyze structural features of different sizes and levels in medical images, enabling the model to better understand and characterize complex medical image data, process medical images with unevenly distributed lesion areas, enhance the robustness of images with different resolutions, and enable the model to have higher accuracy in lesion detection and classification tasks. Step 3-1: In order to more accurately capture the multi-scale features of medical images and perform multi-scale feature decomposition on image data, it is assumed that the feature representation of the image at different resolutions is Where s represents different scale levels (from global structure to local details), S is the total number of scales, and the generator G generates image samples at each scale Where Z is the input noise vector or potential vector, based on which the discriminator D s The generated image samples are evaluated at each scale s to determine their authenticity and quality. The discriminator can perceive and identify the placenta implantation lesion area at multiple scales at the same time, ensuring a comprehensive evaluation of the generated samples at different resolutions, so as to distinguish between real medical images and generated fake samples. In the multi-scale placenta accreta image recognition task, the loss function of the semi-supervised diagnosis model based on GAN can be further expanded to multi-scale loss. The loss function of the generator is And the loss function of the discriminator It can be expressed as the weighted sum of multiple scale losses: Among them, λ s represents the weight of scale s; The loss function of the joint training of the semi-supervised diagnosis model based on the GAN network can be further expanded as follows: Through the multi-scale generative network, the global and local information of the placenta implantation images are captured to improve the model's generation, discrimination and diagnostic performance. Step 4: Perform semi-supervised training, combining limited labeled data with a large amount of unlabeled data, and optimize model performance through consistency regularization and pseudo-label generation mechanism of generative adversarial network. In the semi-supervised learning framework, not only supervised learning is performed on labeled data, but also reasonable pseudo-label generation and model training are performed on unlabeled data to maximize the use of data; Step 4-1: Use consistency regularization to enhance the robustness of the model to input perturbations. By adding a small perturbation ∈ to the input sample x, the model ensures the consistency of the output before and after the perturbation, thereby improving the generalization ability and stability of the model. Suppose the perturbed sample is x+∈, then the goal of consistency regularization is to minimize the difference between the model's output for the original sample x and the perturbation sample x+∈. Formally, the consistency regularization loss function is It can be defined as: Among them, G(x) represents the model's predicted output for the input sample x, G(x+∈) represents the model's predicted output for the perturbation sample x+∈, and D l is a labeled dataset, the perturbation ∈ obeys Gaussian random noise with small variance, and the loss is minimized The model will maintain output consistency for small perturbations, improving its robustness on unknown samples; Step 4-2: Next, use the generative adversarial network to train the unlabeled data D u To generate pseudo labels, the generator G of GAN generates pseudo labels for unlabeled data, and the discriminator D guides the optimization of the generator by distinguishing between real samples and generated samples. Specifically, the sample D generated by the generator u Through the discriminator D u Evaluation is performed to ensure that the generated pseudo-label samples are indistinguishable from the real samples. The loss function for pseudo-label generation is It can be expressed as: Among them, G(x) represents the sample generated by the generator on the unlabeled data x, and the discriminator D discriminates the authenticity of the sample through its classification task. By generating pseudo labels for unlabeled data and training them, the generator can be further optimized to generate more realistic samples and improve the quality of pseudo labels; Step 4-3: After pseudo-label generation is completed, the labeled data and unlabeled data are jointly trained to optimize the model parameters. The loss function of the joint training consists of supervised learning loss, pseudo-label loss, and consistency regularization loss. The joint loss function can be expressed as: L total =L GAN +λ1L consistency +λ2L semi ; Among them, λ1 and λ2 are hyperparameters that control the weights of pseudo-label loss and consistency regularization loss. By optimizing this joint loss, the model can learn effective diagnostic features from labeled data and generate effective pseudo-labels from unlabeled data to improve the overall performance, thereby improving the diagnostic accuracy and robustness under semi-supervised learning. By combining labeled data and pseudo-label data for joint training, the model parameters are optimized. Step 5: Automated diagnosis of placenta accreta; After the model training is completed, it is applied to the automated diagnosis task of clinical placenta accreta. Specifically, the medical image sample x to be diagnosed is first new The input is sent to the trained generative adversarial network, where the generator G and the discriminator D extract and evaluate the image features at multiple scales. The generator generates a feature representation that conforms to the anatomical structure of the placenta implantation based on the input image, and the discriminator classifies and evaluates the generated features to determine whether they conform to the pathological pattern of the placenta implantation. The model captures the pathological information of the placenta implantation by generating pseudo images or high-dimensional feature expressions, and outputs a preliminary diagnosis result. The output result of the GAN network G(x new ) can be expressed as a probability distribution p(y|x new ), where y is the diagnostic label of placenta implantation, with a value of normal or abnormal. The generator G generates features using new image samples and passes them to the discriminator D, which performs multi-scale and multi-level authenticity evaluation. In this way, the model can give a diagnosis probability based on pathological features for each input sample. The diagnosis result can be formalized as: in It is the automated diagnosis result finally given by the model. This diagnosis result is based on the model's comprehensive analysis of the imaging features of the placenta accreta, covering global and local information as well as multi-scale features. Since the model has been trained with a large amount of labeled data and unlabeled data, it has strong generalization ability and can effectively identify the pathological differences of placenta accreta between different individuals.

2. A placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network according to claim 1, characterized in that: The step 1-1 comprises: the imaging department collects and saves the desensitized placenta accreta image data of real patients in DICOM (Digital Imaging and Communications in Medicine) magnetic resonance imaging format, and the experts (chief physicians) related to placenta accreta combine the real clinical information and diagnosis and treatment experience to summarize and set the key fields for disease diagnosis, and manually annotate a small number of placenta accreta image samples to obtain the expert annotated data set D l , used for model classification ability enhancement training, and other large amounts of unlabeled placenta implantation data as unlabeled dataset D u , used for unsupervised model training, unlabeled dataset D u The size of the expert annotation dataset D l size.

3. The placenta accreta detection method based on a semi-supervised multi-scale generative adversarial network according to claim 1, characterized in that: The step 1-2 includes: in order to enable the neural network to read and recognize medical imaging data, the DICOM image data needs to be further processed. First, the DICOM file is read to extract the image data. According to the key fields determined by the experts, the information other than the image pixels in the DICOM file is retained, that is, age, weight and other information. Then, the pixel values ​​are linearly converted according to the "Rescale Slope" and "Rescale Intercept" fields in the DICOM file. Due to the different resolutions of different scanning devices, the image is further resampled so that the images obtained by different scans have the same physical size in space. The image can be adjusted to a uniform pixel size through interpolation technology to ensure the consistency of the image quality in subsequent processing; Magnetic resonance imaging is sliced ​​to obtain 2D placenta implantation image slices. The 2D images are cropped and resized to meet the input requirements of the GAN network. Center cropping is usually used to retain the placenta implantation area, maintain the aspect ratio of the image, and avoid deformation. The pixel values ​​are usually standardized to [0, 1]. In order to further enhance the generalization ability of the model, the image is data enhanced. In order to enhance the sample efficiency and model generalization ability, the image is processed through geometric transformations, namely rotation, flipping, scaling, and brightness contrast adjustment, so that the model has stronger robustness to different imaging conditions. Finally, the preprocessed placenta implantation image data is stored as a tensor file to speed up subsequent data reading and network training to improve network training efficiency.

Images