Image quality processing method, device and equipment for ultrasonic image

Abstract

The invention relates to an image quality processing method, device and equipment for an ultrasonic image. The method comprises the following steps: acquiring a to-be-processed original ultrasonic image; performing standardization processing on the original ultrasonic image through a pre-trained standardization network model to generate a target ultrasonic image; wherein the standardized network model adopts a cyclic generative adversarial network, and adversarial loss functions used in the training process at least comprise cyclic consistency loss, identity mapping loss, least square GAN loss and target task loss; the target task loss is used for representing the accuracy of the standardized image in the target task. The standardization degree of the target ultrasonic image can be improved, and the consistency and comparability of the ultrasonic image are improved through standardization processing.

Description

Technical Field

[0001] The present disclosure relates to the field of image processing technologies, and in particular, to a method, apparatus, and device for processing the image quality of ultrasonic images. Background Art

[0002] Due to differences in imaging parameters, probe characteristics, and post-processing algorithms among ultrasonic devices of different brands and models, ultrasonic images obtained for the same patient on different devices may exhibit significant stylistic differences, affecting doctors' diagnoses and long-term follow-up of patients. At the same time, with the widespread application of artificial intelligence in medical image analysis, standardized and high-quality ultrasonic images are crucial for improving the reliability of automatic diagnosis algorithms. Therefore, how to standardize the quality of ultrasonic images has become an urgent problem to be solved. Summary of the Invention

[0003] To solve the above technical problems, the present disclosure provides a method, apparatus, and device for processing the image quality of ultrasonic images.

[0004] According to one aspect of the present disclosure, there is provided a method for processing the image quality of ultrasonic images, the method comprising: obtaining an original ultrasonic image to be processed; performing standardization processing on the original ultrasonic image through a pre-trained standardization network model to generate a target ultrasonic image; wherein the standardization network model adopts a cyclic generative adversarial network, and the adversarial loss function used in the training process at least includes: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the image after standardization processing in the target task.

[0005] According to another aspect of the present disclosure, there is also provided an apparatus for processing the image quality of ultrasonic images, the apparatus comprising: an image acquisition module, configured to obtain an original ultrasonic image to be processed; a standardization processing module, configured to perform standardization processing on the original ultrasonic image through a pre-trained standardization network model to generate a target ultrasonic image.

[0006] According to another aspect of the present disclosure, there is also provided an electronic device, the electronic device comprising: a processor; a memory for storing executable instructions of the processor; the processor is configured to read the executable instructions from the memory and execute the instructions to implement the above method.

[0007] According to another aspect of the present disclosure, there is also provided a computer-readable storage medium storing a computer program for executing the above method.

[0008] The technical solutions provided by the embodiments of the present disclosure have the following advantages compared with the prior art: The technical solutions provided by the embodiments of the present disclosure include: obtaining an original ultrasound image to be processed; performing normalization processing on the original ultrasound image through a pre-trained normalization network model to generate a target ultrasound image; wherein, the normalization network model adopts a cyclic generative adversarial network, and the adversarial loss function used in the training process at least includes: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the image after normalization processing in the target task.

[0009] Based on the normalization network model using a cyclic generative adversarial network, this solution can perform domain transfer on ultrasound images, thereby eliminating the image quality differences between ultrasound images of different devices, improving the normalization degree of the target ultrasound image, and improving the consistency and comparability of ultrasound images through normalization processing; at the same time, the adversarial loss function used in the training process of the normalization network model takes into account losses in multiple dimensions, enabling the normalization network model to achieve better image processing effects for complex original ultrasound images and improving the image quality of the target ultrasound image; in particular, the target task loss in the adversarial loss function combines the normalization network model with specific target tasks in actual applications, which is beneficial to more accurately process ultrasound images in target tasks and has high practicality. Through the above solutions, reliable data support is provided for the automated analysis of ultrasound images, which helps to improve the accuracy of intelligent diagnosis systems and promote the sharing of imaging data and joint research among medical institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present disclosure and used together with the specification to explain the principles of the present disclosure.

[0011] To more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the following will briefly introduce the accompanying drawings required for describing the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0012] Figure 1 It is a flowchart of the method for processing the image quality of the ultrasound image according to the embodiment of the present disclosure; Figure 2 It is a flowchart of the training method of the normalization network model according to the embodiment of the present disclosure. Figure 3 It is a schematic diagram of the input and output of the standardized network model described in the embodiments of the present disclosure; Figure 4 It is a schematic structural diagram of an image quality processing device for ultrasonic images described in the embodiments of the present disclosure; Figure 5 It is a schematic structural diagram of an electronic device described in the embodiments of the present disclosure. Detailed implementation manners

[0013] In order to be able to more clearly understand the above-mentioned objects, features, and advantages of the present disclosure, the solutions of the present disclosure will be further described below. It should be noted that, without conflict, the embodiments of the present disclosure and the features in the embodiments may be combined with each other.

[0014] In the following description, many specific details are set forth in order to fully understand the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only a part of the embodiments of the present disclosure, rather than all the embodiments.

[0015] Due to differences in ultrasonic device models, processing algorithms, etc., the image quality of ultrasonic images is uneven. At the same time, with the wide application of artificial intelligence in medical image analysis, standardized and high-quality ultrasonic images are crucial for improving the reliability of automatic diagnosis algorithms.

[0016] Based on this, the present disclosure provides an image quality processing method, device, equipment, and medium for ultrasonic images. The ultrasonic image domain transformation method of the present disclosure based on a cyclic generative adversarial network is used to eliminate imaging differences between different devices, transform low-quality or device-specific original ultrasonic images into target ultrasonic images in a standardized target domain, realize the domain transformation of ultrasonic images, while retaining the structural consistency of the images, and improve the accuracy of ultrasonic image segmentation in downstream tasks through domain transformation, thereby improving the comparability of images, enhancing the generalization ability of intelligent diagnosis models, alleviating performance losses caused by insufficient algorithm robustness, and providing technical support for cross-device medical image sharing and remote consultation.

[0017] Embodiment 1: Figure 1 It is a flowchart of an image quality processing method for ultrasonic images provided by the embodiments of the present disclosure. This method can be executed by an image quality processing device, and the device can be implemented by software and / or hardware. Referring to Figure 1 , an image quality processing method may include the following steps.

[0018] S102, Obtain the original ultrasonic image to be processed.

[0019] In this embodiment, the original ultrasonic image to be processed can be collected by ultrasonic devices of any model. Since there are differences in imaging parameters, probe characteristics, and post-processing algorithms among ultrasonic devices of different brands and models, the ultrasonic images obtained for the same patient on different devices may show significant quality differences. Based on this, in this embodiment, it is necessary to standardize the image quality of the collected original ultrasonic image to generate a high-quality target ultrasonic image.

[0020] In one embodiment, the collected original ultrasonic image can also be preprocessed; the preprocessing may include at least one of the following: random cropping and scaling, horizontal flipping, random rotation and conversion to grayscale images, image normalization, registration, denoising, and image enhancement, etc. By preprocessing, the consistency of the original ultrasonic image is ensured, which is more conducive to the standardized network model to perform standardized processing on the original ultrasonic image.

[0021] S104, perform standardized processing on the original ultrasonic image through a pre-trained standardized network model to generate a target ultrasonic image.

[0022] Among them, the standardized network model adopts a cyclic generative adversarial network, and the adversarial loss functions used in the training process at least include: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the image after standardized processing in the target task.

[0023] In one example, the cyclic generative adversarial network adopted by the standardized network model may include: two generators and two discriminators, which are used to perform standardized processing on the low-quality original ultrasonic image to generate a high-quality target ultrasonic image.

[0024] After generating the target ultrasonic image, in this embodiment, the original ultrasonic image and the target ultrasonic image after standardized processing can be displayed through a GUI (Graphical User Interface) interface. Displaying the original ultrasonic image and the target ultrasonic image after standardized processing through the GUI interface facilitates doctors to compare the image effects before and after standardized processing; moreover, the GUI interface provides a parameter adjustment function, enabling doctors to adjust the migration intensity of the standardized network model during standardized processing to meet different clinical needs.

[0025] The image quality processing method for ultrasonic images provided by the embodiments of the present disclosure includes: obtaining an original ultrasonic image to be processed; performing normalization processing on the original ultrasonic image through a pre-trained normalization network model to generate a target ultrasonic image; wherein, the normalization network model adopts a cyclic generative adversarial network, and the adversarial loss function used in the training process at least includes: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the image after normalization processing in the target task.

[0026] Based on the normalization network model adopting a cyclic generative adversarial network, this embodiment can perform domain transfer on ultrasonic images, thereby eliminating the image quality differences between ultrasonic images of different devices, improving the normalization degree of the target ultrasonic image, and enhancing the consistency and comparability of ultrasonic images through normalization processing; at the same time, the adversarial loss function used in the training process of the normalization network model takes into account losses in multiple dimensions, enabling the normalization network model to achieve better image processing effects when facing complex original ultrasonic images and improving the image quality of the target ultrasonic image; in particular, the target task loss in the adversarial loss function combines the normalization network model with specific target tasks in actual applications, which is conducive to more accurately processing ultrasonic images in target tasks and has high practicality. The above solution provides reliable data support for the automated analysis of ultrasonic images, helps improve the accuracy of intelligent diagnosis systems, and promotes the sharing and joint research of imaging data among medical institutions.

[0027] For ease of understanding, the embodiments of the present disclosure are described in detail below.

[0028] In one embodiment, the normalization network model is a cyclic generative adversarial network, including: a first generator, a second generator, and a discriminator. In a possible specific example, the discriminator may include a first discriminator and a second discriminator.

[0029] For the task of quality normalization processing of ultrasonic images in this embodiment, it is understood as a domain transfer task of ultrasonic images, and the differences between ultrasonic images of different devices and different qualities are defined as a kind of "style" difference. Thus, a cyclic generative adversarial network is adopted as the normalization network model to provide a method of a cyclic consistency generative adversarial network based on depthwise separable convolution. Based on this, the domain transformation of ultrasonic images is realized while retaining the structural consistency of the images, and the accuracy of ultrasonic image segmentation in downstream tasks is improved by the domain transformation, alleviating the performance loss caused by insufficient algorithm robustness.

[0030] Specifically, the standardized network model in this embodiment uses a cycle generative adversarial network as the core architecture to construct a network model including two generators (the first generator G and the second generator F) and two discriminators (the first discriminator Dx and the second discriminator Dy).

[0031] The first generator G is used to convert low-quality or device-specific ultrasound images in the source domain into standardized ultrasound images in the target domain; the second generator F is used to convert the ultrasound images in the target domain back to the ultrasound images in the source domain to ensure that the converted images can be restored and maintain the consistency of image information, especially the consistency of the medical information expressed by the ultrasound images.

[0032] For the above first generator G and second generator F, in this embodiment, a large convolution kernel and an inverted bottleneck design can be adopted to increase the receptive field of the standardized network model, and convolutional inductive bias and residual connections are introduced to extract context information and retain image detail features. At the same time, the ResNet module is embedded in the middle layer of the generator, further enhancing the stability and expressive ability of the network in high-level abstract information extraction.

[0033] The first discriminator Dx and the second discriminator Dy are respectively used to determine whether the input images come from the true data distribution, so as to guide the two generators to learn the true image conversion mapping.

[0034] For the above first discriminator Dx and second discriminator Dy, in this embodiment, a dual-scale update rule and 1-Lipschitz continuous spectral normalization can be adopted to improve the stability of the standardized network model.

[0035] In another embodiment, the cycle generative adversarial network adopted by the standardized network model can further enhance the effect by optimizing the network structure. For example, techniques such as adopting a deeper network architecture, introducing an attention mechanism or multi-scale feature fusion can be used, and these methods can improve the expressive ability of the network and the feature extraction effect.

[0036] In order to enable the standardized network model to be directly applied to the standardized processing of ultrasound images, it is necessary to train the standardized network model in advance. The parameters of the standardized network model need to be obtained through training. The purpose of training the standardized network model is to finally determine the parameters that meet the requirements. Refer to Figure 2 , this embodiment provides a training method for the standardized network model, including the following steps S202 - S214.

[0037] S202, obtain the preprocessed training data.

[0038] Specific embodiments may include: obtaining multiple candidate ultrasound images from different ultrasound devices; selecting low-quality ultrasound images in the source domain and high-quality ultrasound images in the target domain from the candidate ultrasound images; preprocessing the low-quality ultrasound images and high-quality ultrasound images, and determining the preprocessed low-quality ultrasound images and high-quality ultrasound images as training data; wherein, the preprocessing includes at least but is not limited to: image registration, image normalization, image denoising, and data augmentation.

[0039] In this embodiment, candidate ultrasound images from ultrasound devices of different brands and models can be collected to ensure that the training data has a wide distribution. When selecting images, ultrasound images with the same anatomical part and similar imaging angles can be selected from the candidate ultrasound images. For the selected ultrasound images, according to a preset image quality index value, low-quality ultrasound images in the source domain and high-quality ultrasound images in the target domain are established; wherein, the low-quality ultrasound images in the source domain are generally low-quality images or device-specific images with an image quality index value lower than the preset value, and the high-quality ultrasound images in the target domain are generally standardized high-quality images with an image quality index value not lower than the preset value.

[0040] Preprocess the low-quality ultrasound images and high-quality ultrasound images, wherein the preprocessing includes the following examples.

[0041] a) Image registration: Perform morphological registration on ultrasound images from different devices to reduce the deviation of anatomical structures.

[0042] b) Image normalization: Unify the gray-scale distribution of ultrasound images to reduce the contrast and brightness differences of different devices.

[0043] c) Image denoising: Use Gaussian filtering or wavelet transform for denoising to improve the quality of ultrasound images.

[0044] d) Data augmentation: Flip, rotate, and scale-transform the ultrasound images to enhance the generalization ability of the model.

[0045] The ultrasound images involved in the above preprocessing process include low-quality ultrasound images and high-quality ultrasound images, and the consistency of the input training data can be ensured through preprocessing.

[0046] S204. Construct a cyclic generative adversarial network. The cyclic generative adversarial network may include, for example: a first generator G, a second generator F, a first discriminator Dx, and a second discriminator Dy.

[0047] S206. Input the training data into the cycle generative adversarial network for training to obtain an adversarial loss function. The adversarial loss function at least includes: cycle consistency loss, identity mapping loss, least squares GAN loss, and target task loss. The target task loss is used to represent the accuracy of the normalized image in the target task.

[0048] This step S206 may include: generating a high-quality reconstructed image G(x) by the first generator G according to the input low-quality ultrasound image x; generating a low-quality reconstructed image F(y) by the second generator F according to the input high-quality ultrasound image y; calculating the adversarial loss function based on the low-quality ultrasound image x, the high-quality reconstructed image G(x), the high-quality ultrasound image y, and the low-quality reconstructed image F(y).

[0049] In the adversarial loss function, the cycle consistency loss is used to ensure that the input image passing through one generator should be close to the original input image after being reversely generated by the other generator, so as to ensure that important information will not be lost during the conversion process of the image through the generator and the reverse generator.

[0050] The cycle consistency loss L cycle (G,F) in this embodiment can refer to the following formula (1): (1) Wherein, G(x) represents the high-quality reconstructed image generated by the first generator G according to the input low-quality ultrasound image x; F(y) represents the low-quality reconstructed image generated by the second generator F according to the input high-quality ultrasound image y; ||.||1 represents the L1 loss metric, which is used to calculate the difference between the high-quality reconstructed image G(x) and the original low-quality ultrasound image x, and to calculate the difference between the low-quality reconstructed image F(y) and the original high-quality ultrasound image y.

[0051] In the adversarial loss function, the identity mapping loss is used to train the two generators so that when the input is a high-quality ultrasound image in the target domain, they can generate an output image identical to the input image, thereby maintaining the identity information of the image.

[0052] The identity mapping loss L identity (G) in this embodiment can refer to the following formula (2): (2) Among them, ||.||1 represents the L1 loss metric, ensuring that when the image already belongs to the target domain, the generated image is as close as possible to the original input image. That is, when the image already belongs to the high-quality ultrasound image in the target domain, the generated low-quality reconstructed image is as close as possible to the original input high-quality ultrasound image.

[0053] In the adversarial loss function, the least squares GAN loss (LSGAN Loss) is used to train two discriminators to better distinguish between real images and generated reconstructed images. To prevent instability during training, this embodiment can adopt a label smoothing strategy.

[0054] The least squares GAN loss in this embodiment can refer to the following formula (3): (3) Among them, D(x) represents the output of the first discriminator D; G(z) represents the output of the first generator G; represents the label smoothing factor, which is used to adjust the label of the real sample to 1 - and the label of the fake sample to to alleviate instability during training.

[0055] In the adversarial loss function, the target task loss is used to take the performance of the normalized image in a medical AI (Artificial Intelligence) task (such as lesion classification) as an auxiliary loss to optimize the image quality to improve diagnostic usability.

[0056] Traditional cycle generative adversarial networks only rely on pixel-level adversarial loss and cycle consistency loss. On this basis, this embodiment introduces an auxiliary loss function based on downstream tasks, that is, uses the accuracy of the normalized image in specific ultrasound diagnosis tasks (such as lesion classification, segmentation) as the optimization goal to improve the clinical usability of the generated image. In addition, this embodiment can adopt a multi-scale generator and residual connections to enhance the capture of detailed features, and combine a dual-scale update rule and 1-Lipschitz continuous spectral normalization to enhance the model stability, to ensure that the cycle generative adversarial network can be trained stably, and the converted image not only conforms to the characteristics of medical imaging but also keeps the diagnostic information unchanged.

[0057] The target task loss in this embodiment includes: (1)In the target task, determine the prediction mask and the real mask of the high-quality reconstructed image; calculate the L1 loss between the prediction mask and the real mask; where the L1 loss is used to represent the absolute difference between the prediction mask and the real mask, as shown in the following formula (4): (4) where N is the number of pixels in the high-quality reconstructed image, represents the predicted mask, and represents the ground truth mask.

[0058] (2) Calculate the Dice loss between the predicted mask and the ground truth mask; where the Dice loss is used to represent the similarity between the predicted mask and the ground truth mask. Specifically, the Dice loss measures the similarity between the predicted mask and the ground truth mask by calculating the ratio of the overlapping part of the predicted mask and the ground truth mask to their union; the Dice loss L Dice is calculated according to the following formula (5): (5) (3) Determine the weighted sum between the L1 loss and the Dice loss as the target task loss. The target task loss L Task-Specific is calculated according to the following formula (6): (6) where and represent the weight coefficients of the Dice loss and the L1 loss respectively, which are used to adjust the contributions of the Dice loss and the L1 loss in the overall loss.

[0059] The core of this embodiment is to use a downstream target task close to clinical applications (such as lesion detection, segmentation, or classification) as the loss function to guide the training of the cyclic generative adversarial network. This training method not only ensures the practicality of the cyclic generative adversarial network in the clinical environment but also bypasses the problem of how to define high-quality medical images in traditional methods, enabling the cyclic generative adversarial network to directly learn features highly relevant to clinical needs.

[0060] In this embodiment, the adversarial loss function is usually used in generative adversarial networks (GANs) to train the generator and the discriminator. The generator tries to generate as realistic images as possible by competing with the discriminator, while the discriminator aims to distinguish between real images and generated images.

[0061] Generally, the adversarial loss function is as shown in the following formula (7): (7) Combined with the cycle consistency loss, identity mapping loss, least squares GAN loss, and target task loss in the foregoing embodiments, the overall adversarial loss function in this embodiment can be composed of the cycle consistency loss, identity mapping loss, least squares GAN loss, and target task loss, and the formula is as follows: (8) Among them, and are the weight coefficients of the cyclic consistency loss and the identity mapping loss respectively; represents the least squares GAN loss between the first generator G and the first discriminator Dx; represents the least squares GAN loss between the second generator F and the second discriminator Dy; represents the downstream target task loss.

[0062] S208. Update the network parameters according to the adversarial loss function.

[0063] In this embodiment, the training data in step S202 can be divided into data for training and data for validation. Specifically, for example, 80% of the training data is used for training, and 20% of the training data is used for validation.

[0064] During the training process, each batch of training data contains unpaired ultrasound images from different devices, that is, low-quality ultrasound images and high-quality ultrasound images.

[0065] This embodiment can use the Adam optimizer for parameter optimization, and the initial learning rate is set to 0.0002 (for example only, not limited to this). First, update the two discriminators. Through the confrontation between real images and generated images, improve the classification ability of the discriminators; then update the two generators so that the converted images can not only deceive the discriminators but also meet the structural requirements of medical images.

[0066] S210. Determine whether the cyclic generative adversarial network meets the training stop condition; among them, the training stop condition includes: the number of training times reaches the maximum value (such as 200 rounds) and / or the adversarial loss function converges. If not, return to the step of inputting the training data into the cyclic generative adversarial network for training. If so, execute the following step S212.

[0067] S212. Stop the training of the cyclic generative adversarial network, and determine the cyclic generative adversarial network at the time of stopping training as the standardized network model.

[0068] After multiple trainings, when it is determined that the cyclic generative adversarial network meets the training stop condition, it means that the quality of the images generated by the trained cyclic generative adversarial network is stable and can be used for the standardization processing of ultrasound images to improve the image quality. In this case, the cyclic generative adversarial network at the time of stopping training is determined as the standardized network model.

[0069] Refer to Figure 3For the standardized network model after training, in this embodiment, the original ultrasound images to be migrated (such as ultrasound images from different devices) can be input into the standardized network model, and after being transformed by the first generator G of the standardized network model, the target ultrasound images after standardized processing are obtained.

[0070] Compared with traditional methods, in this embodiment, not only a new generator and discriminator are combined to improve the image generation quality, but also aiming at the problem that it is difficult to objectively evaluate the quality of ultrasound images, the accuracy of the downstream target task is introduced as the loss function (i.e., the target task loss) to optimize the training process of the cyclic generative adversarial network, so as to significantly improve the domain migration effect and the usability of the downstream target task while ensuring the consistency of image content. This method provides strong technical support for the cross-device generality and diagnostic value of ultrasound images.

[0071] Based on the above embodiments, the image quality processing method for ultrasound images provided by the present disclosure may further include: evaluating the network performance of the standardized network model according to preset test metrics; where the test metrics at least include: structural similarity index, peak signal-to-noise ratio, and classification accuracy of the target task.

[0072] In this embodiment, the structural similarity index (SSIM) is used to evaluate the structural consistency between the converted high-quality target ultrasound image corresponding to the original ultrasound image in the source domain and the high-quality ultrasound image in the target domain.

[0073] The peak signal-to-noise ratio (PSNR) is used to measure the quality of the generated target ultrasound image.

[0074] The classification accuracy of the downstream target task is used to verify whether the target ultrasound image after standardized processing improves the accuracy of the target task through the diagnostic model.

[0075] In addition, the test metrics may further include an expert score; the expert score is used to evaluate the visual quality and clinical usability of the generated target ultrasound image by a doctor in the ultrasound department.

[0076] In summary, the image quality processing method for ultrasound images provided by the present disclosure includes: obtaining the original ultrasound images to be processed; performing standardized processing on the original ultrasound images through a pre-trained standardized network model to generate target ultrasound images; where the standardized network model uses a cyclic generative adversarial network, and the adversarial loss functions used in the training process at least include: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the images after standardized processing in the target task.

[0077] This solution is based on a standardized network model that uses a cyclic generative adversarial network (CycleGAN) and can perform domain transfer on ultrasound images, thereby eliminating the differences in image quality between ultrasound images from different devices, improving the standardization of target ultrasound images, enhancing the consistency and comparability of ultrasound images through standardization processing, providing reliable data support for the automated analysis of ultrasound images, helping to improve the accuracy of intelligent diagnosis systems, and facilitating the sharing and joint research of imaging data among medical institutions.

[0078] Moreover, due to differences in imaging parameters, probe types, and post-processing algorithms among ultrasound devices of different brands and models, there are obvious device specificities in ultrasound images. To address this issue, the present disclosure performs standardization processing on the original ultrasound images through a standardized network model. Among them, the standardized network model is based on a cyclic generative adversarial network and improves the generator and discriminator as well as optimizes the adversarial loss function. Through the standardized network model, low-quality or device-specific original ultrasound images can be converted into standardized high-quality target ultrasound images, improving the consistency of ultrasound images and making them easier to compare and analyze across devices.

[0079] Regarding the standardized network model in the present disclosure, the adversarial loss function used in its training process takes into account losses in multiple dimensions, enabling the standardized network model to achieve better image processing effects when facing complex original ultrasound images and improving the image quality of target ultrasound images. In particular, the medical downstream target task loss is incorporated into the adversarial loss function. By adding the target task loss to optimize the adversarial loss function during training, the standardized network model can be combined with specific target tasks in actual applications, which is beneficial for more accurately processing ultrasound images in target tasks and better realizing high-quality ultrasound image style transfer, and has high practicality.

[0080] The method for processing the image quality of ultrasound images provided by the embodiments of the present disclosure can be widely applied to various medical imaging tasks, such as cross-device ultrasound image standardization, ultrasound imaging big data analysis, telemedicine, and artificial intelligence-assisted diagnosis, etc. It has broad application value in aspects such as ultrasound imaging standardization, telemedicine, and AI-assisted diagnosis, providing important technical support for medical imaging data sharing and intelligent diagnosis.

[0081] Some application examples based on the present disclosure can refer to the following scenarios.

[0082] Cross-device standardization of ultrasound imaging: Applying the present disclosure can better solve the problem of inconsistent ultrasound image quality between different devices, making the ultrasound images of different devices have similar visual characteristics and improving diagnostic consistency.

[0083] Telemedicine and Data Sharing: The application of the present disclosure can promote the sharing of ultrasound image data between different hospitals and regions, and improve the feasibility of remote consultation.

[0084] Enhancement of AI Diagnostic Model: Since the AI diagnostic system is highly sensitive to the quality of training data, the use of the present disclosure can eliminate device differences, improve the generalization ability of the standardized network model, and make it applicable to more ultrasound devices.

[0085] In summary, the image quality processing method for ultrasound images provided by the embodiments of the present disclosure has the following advantages: 1) It can better eliminate the imaging differences between ultrasound devices and improve the standardization of ultrasound images.

[0086] 2) Maintain the consistency of medical information and enhance clinical usability. Traditional image enhancement or style transfer methods may change the key information in medical ultrasound images and affect clinical diagnosis. The present disclosure introduces cycle consistency loss and perceptual loss in the training process of the standardized network model to ensure that the converted ultrasound images retain important medical features while optimizing the visual quality, guaranteeing reliability in actual clinical applications.

[0087] 3) Combine downstream task optimization to improve the applicability of medical AI diagnosis. Traditional image style transfer methods only focus on the visual similarity of images and ignore the medical application value. This case introduces the accuracy of the medical downstream target task as an auxiliary loss (i.e., target task loss) to directly optimize the artificial intelligence diagnosis performance during the training process, enabling the standardized target ultrasound images to improve the generalization ability of the medical AI diagnostic model and enhance the accuracy of disease detection and classification.

[0088] 4) No paired data is required, expanding the model's applicable range. Compared with traditional super-resolution or image registration methods that require paired data for supervised learning, the present disclosure is based on an unsupervised cycle generative adversarial network as the standardized network model, only requiring the collection of ultrasound images from different devices as training data without strict pairing, which greatly reduces the data collection difficulty and makes the method applicable to more devices and application scenarios.

[0089] 5) Improve the feasibility of telemedicine and medical image data sharing. The image standardization method of the present disclosure can make the ultrasound images of different medical institutions tend to be consistent in quality, facilitating cross-device diagnosis by doctors in the telemedicine system, enhancing the sharing value of the medical image database, and increasing the generality of ultrasound image AI training data.

[0090] 6) Adapt to a variety of ultrasound applications and improve the stability of clinical diagnosis. The present disclosure can be widely applied to multiple medical imaging scenarios such as breast ultrasound, abdominal ultrasound, and cardiac ultrasound, and can improve the imaging stability of different ultrasound applications, enabling doctors and AI diagnosis systems to obtain stable image quality on different devices and reducing the misdiagnosis rate caused by device differences.

[0091] 7) End-to-end automatic processing, suitable for actual clinical workflows. The standardized network model in the present disclosure adopts an end-to-end automatic processing mode, without the need for doctors to manually adjust parameters, and can automatically identify input images and complete the standardized processing, which is suitable for integration with hospital PACS systems and improves the work efficiency of doctors.

[0092] Based on this, the present disclosure has advantages such as high efficiency, medical consistency, AI friendliness, wide application range, and remote medical support in the standardization of ultrasound image quality, provides important technical support for medical institutions, artificial intelligence imaging diagnosis systems, and ultrasound image data sharing, helps to promote the further development of ultrasound medical imaging in the field of intelligent medicine, can effectively improve the diagnostic accuracy and efficiency, and the present disclosure has strong clinical application potential and versatility.

[0093] Embodiment 2: Figure 4 It is a schematic structural diagram of an image quality processing device for an ultrasound image provided by an embodiment of the present disclosure. This device can be used to implement the image quality processing method for an ultrasound image provided in the above embodiment, and this device can be implemented by software and / or hardware. Referring to Figure 4 An image quality processing device may include the following modules.

[0094] An image acquisition module 310, configured to acquire an original ultrasound image to be processed; A standardization processing module 320, configured to perform standardization processing on the original ultrasound image through a pre-trained standardization network model to generate a target ultrasound image; Wherein, the standardization network model adopts a cyclic generative adversarial network, and the adversarial loss function used in the training process at least includes: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the image after standardization processing in the target task.

[0095] In one embodiment, the standardization network model includes: a first generator, a second generator, and a discriminator.

[0096] In one embodiment, the image quality processing device includes a model training module, which is used to execute the training process of the standardization network model; The training process of the standardization network model includes: Obtain the preprocessed training data; Construct a cyclic generative adversarial network; Input the training data into the cyclic generative adversarial network for training to obtain an adversarial loss function; Update the network parameters according to the adversarial loss function; Determine whether the cyclic generative adversarial network satisfies the training stop condition; wherein, the training stop condition includes: the number of training times reaches the maximum value and / or the adversarial loss function converges; If not, return to the step of inputting the training data into the cyclic generative adversarial network for training; If satisfied, stop the training of the cyclic generative adversarial network, and determine the cyclic generative adversarial network at the time of stopping training as the standardized network model.

[0097] In one embodiment, the model training module is used for: Obtain multiple candidate ultrasound images from different ultrasound devices; Select low-quality ultrasound images in the source domain and high-quality ultrasound images in the target domain from the candidate ultrasound images; Preprocess the low-quality ultrasound images and the high-quality ultrasound images, and determine the preprocessed low-quality ultrasound images and high-quality ultrasound images as training data; wherein, the preprocessing at least includes: image registration, image normalization, image denoising, and data augmentation.

[0098] In one embodiment, the model training module: Generate a high-quality reconstructed image by the first generator according to the input low-quality ultrasound image; Generate a low-quality reconstructed image by the second generator according to the input high-quality ultrasound image; Calculate the adversarial loss function based on the low-quality ultrasound image, the high-quality reconstructed image, the high-quality ultrasound image, and the low-quality reconstructed image.

[0099] In one embodiment, the target task loss includes: In the target task, determine the predicted mask and the ground truth mask of the high-quality reconstructed image; Calculate the L1 loss between the predicted mask and the ground truth mask; wherein, the L1 loss is used to represent the absolute difference between the predicted mask and the ground truth mask; Calculate the Dice loss between the predicted mask and the ground truth mask; wherein, the Dice loss is used to represent the similarity between the predicted mask and the ground truth mask; Determine the weighted sum between the L1 loss and the Dice loss as the target task loss.

[0100] In one embodiment, the image quality processing device further includes a performance evaluation module, which is configured to: Evaluate the network performance of the standardized network model according to preset test metrics; wherein, the test metrics at least include: structural similarity index, peak signal-to-noise ratio, and classification accuracy of the target task.

[0101] In one embodiment, the image quality processing device further includes an image display module, which is configured to: Display the original ultrasound image and the target ultrasound image after the normalization process through a GUI interface.

[0102] The device provided in this embodiment has the same implementation principle and the same technical effects as those of the foregoing method embodiment. For a brief description, for the parts not mentioned in the device embodiment, reference may be made to the corresponding content in the foregoing method embodiment.

[0103] Embodiment Three: Figure 5 This is a schematic structural diagram of an electronic device provided in an embodiment of the present disclosure. As Figure 5 shown, the electronic device 400 includes one or more processors 401 and a memory 402.

[0104] The processor 401 may be a central processing unit (CPU) or other forms of processing units with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 400 to perform desired functions.

[0105] The memory 402 may include one or more computer program products, and the computer program products may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 401 may run the program instructions to implement the ultrasound image quality processing method of the embodiments of the present disclosure described above and / or other desired functions. Various contents such as input signals, signal components, and noise components may also be stored in the computer-readable storage medium.

[0106] In one example, the electronic device 400 may further include: an input device 403 and an output device 404, and these components are interconnected through a bus system and / or other forms of connection mechanisms (not shown).

[0107] In addition, the input device 403 may further include, for example, a keyboard, a mouse, and so on.

[0108] The output device 404 may output various information to the outside, including the determined distance information, direction information, etc. The output device 404 may include, for example, a display, a speaker, a printer, a communication network, and remote output devices connected thereto, and so on.

[0109] Of course, for simplicity, Figure 5 only some of the components related to the present disclosure in the electronic device 400 are shown, and components such as a bus, an input / output interface, and so on are omitted. In addition, according to specific application scenarios, the electronic device 400 may further include any other appropriate components.

[0110] Furthermore, this embodiment also provides a computer-readable storage medium, and the storage medium stores a computer program, and the computer program is used to execute the above-mentioned image quality processing method of the ultrasonic image.

[0111] A computer program product of an ultrasonic image quality processing method, device, electronic device, and medium provided by an embodiment of the present disclosure includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the method described in the foregoing method embodiments. For specific implementation, reference can be made to the method embodiments, and details are not described herein again.

[0112] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover a non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0113] The above are only specific embodiments of the present disclosure, enabling those skilled in the art to understand or implement the present disclosure. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to these embodiments described herein, but rather will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for processing the image quality of an ultrasonic image, characterized in that, The method includes: Obtaining an original ultrasound image to be processed; Performing normalization processing on the original ultrasound image through a pre-trained normalization network model to generate a target ultrasound image; Wherein, the normalization network model adopts a cyclic generative adversarial network, and the adversarial loss function used in the training process at least includes: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the image after normalization processing in the target task.

2. The method according to claim 1, wherein The normalization network model includes: a first generator, a second generator, and a discriminator.

3. The method according to claim 2, wherein The training process of the normalization network model includes: Obtaining preprocessed training data; Constructing a cyclic generative adversarial network; Inputting the training data into the cyclic generative adversarial network for training to obtain an adversarial loss function; Updating network parameters according to the adversarial loss function; Judging whether the cyclic generative adversarial network meets the training stop condition; wherein, the training stop condition includes: the number of training times reaches the maximum value and / or the adversarial loss function converges; If not, return to the step of inputting the training data into the cyclic generative adversarial network for training; If so, stop the training of the cyclic generative adversarial network, and determine the cyclic generative adversarial network at the time of stopping training as the normalization network model.

4. The method according to claim 3, wherein The obtaining of the preprocessed training data includes: Obtaining multiple candidate ultrasound images from different ultrasound devices; Selecting low-quality ultrasound images in the source domain and high-quality ultrasound images in the target domain from the candidate ultrasound images; Performing preprocessing on the low-quality ultrasound images and the high-quality ultrasound images, and determining the preprocessed low-quality ultrasound images and high-quality ultrasound images as training data; wherein, the preprocessing at least includes: image registration, image normalization, image denoising, and data augmentation.

5. The method according to claim 3, wherein The inputting the training data into the cyclic generative adversarial network for training to obtain an adversarial loss function includes: Generating a high-quality reconstructed image by the first generator according to the input low-quality ultrasound image; Generating a low-quality reconstructed image by the second generator according to the input high-quality ultrasound image; Calculating an adversarial loss function based on the low-quality ultrasound image, the high-quality reconstructed image, the high-quality ultrasound image, and the low-quality reconstructed image.

6. The method according to claim 5, wherein The target task loss includes: Determining a predicted mask and a ground truth mask of the high-quality reconstructed image in the target task; Calculating the L1 loss between the predicted mask and the ground truth mask; wherein, the L1 loss is used to represent the absolute difference between the predicted mask and the ground truth mask; Calculating the Dice loss between the predicted mask and the ground truth mask; wherein, the Dice loss is used to represent the similarity between the predicted mask and the ground truth mask; Determining the weighted sum between the L1 loss and the Dice loss as the target task loss.

7. The method according to claim 3, characterized in that, The method further includes: Evaluate the network performance of the standardized network model according to preset test metrics; wherein, the test metrics at least include: structural similarity index, peak signal-to-noise ratio, and classification accuracy of the target task.

8. The method according to claim 1, wherein The method further includes: Display the original ultrasound image and the target ultrasound image after standardization processing through the GUI interface.

9. An image quality processing device for ultrasonic images, characterized in that, The device includes: An image acquisition module, configured to acquire an original ultrasound image to be processed; A standardization processing module, configured to perform standardization processing on the original ultrasound image through a pre-trained standardized network model to generate a target ultrasound image; Wherein, the standardized network model adopts a cyclic generative adversarial network, and the adversarial loss function used in the training process at least includes: cyclic consistency loss, identity mapping loss, least squares GAN loss, and target task loss; the target task loss is used to represent the accuracy of the image after standardization processing in the target task.

10. An electronic device, characterized in that, The electronic device includes: A processor; A memory for storing executable instructions that can be executed by the processor; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the method described in any one of the above claims 1-8.

Images