Ultrasonic image quality control method and system

Abstract

The present application relates to the field of ultrasonic imaging, and in particular to a method and system for ultrasonic image quality control, the method comprising: obtaining ultrasonic image information; performing preliminary quality control on the ultrasonic image information to obtain a first image of preliminary quality; performing position information quality control on the first image of preliminary quality to obtain a second image with correct position information; and performing internal imaging quality control on the second image with correct position information to obtain a high-quality ultrasonic image. Through preliminary quality control, the present application can quickly identify some ultrasonic images with poor imaging quality, thereby reducing errors caused by subsequent inspection steps. Then, through position information quality control, it can preliminarily identify ultrasonic images with incorrect detection positions, thereby reducing invalid detections. Finally, through internal imaging quality control, it can obtain high-quality ultrasonic images, thereby obtaining more accurate and reliable detection results.

Description

Technical Field

[0001] The present application relates to the field of ultrasonic imaging, and in particular to a method and system for controlling ultrasonic imaging quality. Background Art

[0002] Due to the superficial location of the carotid artery, ultrasound is highly sensitive to it. Carotid color Doppler ultrasound can indirectly reflect systemic atherosclerosis by measuring the degree of carotid arteriosclerosis. Carotid ultrasound is highly effective in assessing patients at high risk for cardiovascular disease, providing a comprehensive understanding of their condition, guiding clinical treatment, and assessing prognosis. It has been widely used in clinical practice.

[0003] Ultrasound imaging, with its advantages of being non-invasive, low-cost, and highly real-time, has become the preferred imaging modality for the clinical diagnosis of numerous diseases. However, ultrasound imaging results are easily influenced by the physician's scanning technique and experience. Incorrect scanning planes can directly affect the diagnosis, while omissions can also lead to missed diagnoses.

[0004] With respect to the above-mentioned related technologies, the inventors believe that in order to obtain high-quality ultrasound images, the quality of the ultrasound images needs to be controlled. Summary of the Invention

[0005] In order to obtain high-quality ultrasound images, the present application provides an ultrasound image quality control method and system.

[0006] In a first aspect, the present application provides an ultrasound image quality control method, which adopts the following technical solution:

[0007] A method for controlling ultrasound image quality, comprising:

[0008] Obtaining ultrasound imaging information;

[0009] Performing preliminary quality control on the ultrasound image information to obtain a first image of preliminary quality;

[0010] Performing part information quality control on the first image that has initially qualified quality, to obtain a second image with correct part information;

[0011] Internal imaging quality control is performed on the second image with correct location information to obtain a high-quality ultrasound image.

[0012] By adopting the above technical solution, through preliminary quality control, some ultrasound images with poor imaging quality can be quickly identified, reducing the errors caused by subsequent inspection steps. Then, through location information quality control, ultrasound images with incorrect detection locations can be preliminarily identified, reducing invalid detections. Finally, through internal imaging quality control, high-quality ultrasound images can be obtained, and test results can be obtained based on them, making the test results more accurate and reliable.

[0013] Optionally, performing preliminary quality control on the ultrasound image information includes:

[0014] extracting text information and image information from the ultrasonic image information respectively;

[0015] Comparing and analyzing the text information with the actual text, and calculating the accuracy of the text recognition;

[0016] Performing clarity analysis on the image information.

[0017] By adopting the above technical solution, the problem of mismatch between ultrasound images and patients can be reduced through the extracted text information. By analyzing the clarity of the extracted image information, some ultrasound images with poor imaging quality can be quickly identified, reducing the errors caused by subsequent inspection steps.

[0018] Optionally, the performing clarity analysis on the image information includes:

[0019] Acquiring a histogram of the image information;

[0020] Calculating the distance between the histogram of the image information and a preset standard image histogram;

[0021] If the distance is higher than a preset threshold, the current image information is discarded; if the distance is lower than the preset threshold, the current image information is used as the first image with preliminary qualified quality.

[0022] By adopting the above technical solution, the boundary representation of tissues in ultrasound images is affected by the brightness of the image. The overall quality of ultrasound images is low, and there are often problems of being too bright or too dark. These have a certain degree of impact on image detection. Some even make it impossible to observe some key features with the naked eye, thereby affecting the accuracy of the subsequent detection process and causing misdiagnosis. By extracting the histogram of image information and calculating it with the preset standard image histogram, unqualified images are discarded to obtain relatively high-quality images, which is convenient for further detection and extraction of important features.

[0023] Optionally, performing part information quality control on the first image that has preliminary qualified quality includes:

[0024] detecting whether features of the part to be diagnosed exist in the first image;

[0025] performing blood vessel site recognition on the first image containing features of the site to be diagnosed;

[0026] determining whether the identified blood vessel portion is a standard section;

[0027] If it is a non-standard section, the vascular position is standardized.

[0028] By adopting the above technical solution, during the actual operation, due to the technician's error, the input image may not be the image of the specified part, which may easily lead to misdiagnosis in subsequent judgments. The input image may also not be a standard cross-section, in which case standardization processing is required.

[0029] Optionally, determining whether the identified blood vessel portion is a standard section includes:

[0030] Based on the clinical definition of standard sections, we can make a priori knowledge, that is, if a section is deformed or stretched to resemble a standard section, it is still considered a standard section.

[0031] Collect standard section and non-standard section images as a dataset and assign corresponding labels. According to prior knowledge, the standard section includes deformed standard section images.

[0032] Use deep learning based classification method to do binary classification.

[0033] By adopting the above technical solution, by collecting and training a dataset of standard and non-standard cross-section images, and using a deep learning-based classification method to perform binary classification, the determination of whether the vascular site is a standard cross-section can be further reduced. Furthermore, by defining standard cross-sections, such as including deformed and stretched cross-sections as standard cross-sections, the requirements for standard cross-sections are lowered, preventing even slightly inferior images from being considered non-standard, thereby reducing misjudgments.

[0034] Optionally, performing internal imaging quality control on the second image having correct part information includes:

[0035] determining whether there is a clearly visible tunica intima in the second image;

[0036] If present, the area covered by the clearly visible intima-media membrane was quantified;

[0037] The intima-media image is repaired according to the time sequence information.

[0038] By adopting the above technical solution, in the second image in which the endo-media membrane is clearly visible, the coverage area of ​​the endo-media membrane is quantified, thereby judging the imaging quality of the second image. At the same time, the endo-media membrane image can be repaired according to the time series information, so that the endo-media membrane image better meets the requirements of subsequent detection.

[0039] Optionally, repairing the intima-media image according to the time series information includes:

[0040] Acquiring serial ultrasound images;

[0041] Splitting the sequence of ultrasound images into a plurality of ultrasound image frames arranged in time sequence;

[0042] A generative adversarial network is used to perform image generation training on a plurality of ultrasound image frames arranged in time sequence.

[0043] By adopting the above technical solution, the sequential ultrasound image is composed of a series of ultrasound image frames arranged in time sequence. When information is missing in a certain frame due to noise or artifacts, it can be repaired in other frames.

[0044] Optionally, after repairing the intima-media image according to the time series information, the method further includes: using a classified deep learning convolutional neural network to determine whether a plaque exists.

[0045] By adopting the above technical solution, after repairing the intima-media image, it is possible to continue to determine whether there is a plaque, and the detection result can be provided to the doctor as a reference.

[0046] Optionally, after obtaining a high-quality ultrasound image, the method further includes: generating a report based on the ultrasound image.

[0047] By adopting the above technical solution, a more intuitive report can be generated based on indicators such as the clarity quality of the input image, the quality of the detection site, and the imaging quality of the intima-media membrane.

[0048] In a second aspect, the present application also provides an ultrasound image quality control system, which adopts the following technical solution:

[0049] An ultrasound image quality control system, comprising:

[0050] Ultrasonic image information acquisition module: used to acquire ultrasonic image information;

[0051] A preliminary quality control module is used to perform preliminary quality control on the ultrasound image information to obtain a first image with preliminary qualified quality;

[0052] A part information quality control module is configured to perform part information quality control on the first image that has initially passed quality control, to obtain a second image with correct part information;

[0053] Internal imaging quality control module: used to perform internal imaging quality control on the second image with correct location information to obtain a high-quality ultrasound image.

[0054] By adopting the above technical solution, the preliminary quality control module can quickly screen out some ultrasound images with poor imaging quality, reducing the errors caused by subsequent inspection steps. The position information quality control module can preliminarily screen out ultrasound images with incorrect detection positions, reducing invalid detections. The internal imaging quality control module can obtain high-quality ultrasound images and obtain test results based on them, making the test results more accurate and reliable.

[0055] To sum up, through preliminary quality control, this application can quickly screen out some ultrasound images with poor imaging quality, reduce the errors caused by subsequent inspection steps, and then through location information quality control, it can preliminarily screen out ultrasound images with incorrect detection locations, reduce invalid detections, and finally through internal imaging quality control, it can obtain high-quality ultrasound images, and obtain test results based on them, so that the test results are more accurate and reliable, ensuring the quality of the ultrasound images to be tested, improving the efficiency of ultrasound inspection business execution, ensuring the standardization and standardization of ultrasound images, and improving the accuracy of neck examinations. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of the ultrasound image quality control method of this application.

[0057] Figure 2 This is the block diagram of the ultrasound image quality control system of this application.

[0058] Description of reference numerals:

[0059] 1. Ultrasound image information acquisition module; 2. Preliminary quality control module; 3. Location information quality control module; 4. Internal imaging quality control module. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1-2 It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0061] The present application embodiment discloses a method for controlling the quality of ultrasound images. Figure 1 , the method comprises the following steps:

[0062] S01: Acquire ultrasound image information;

[0063] S02: performing preliminary quality control on the ultrasound image information to obtain a first image of preliminary quality;

[0064] S03: performing part information quality control on the first image that has preliminarily qualified quality, to obtain a second image with correct part information;

[0065] S04: Perform internal imaging quality control on the second image with correct location information to obtain a high-quality ultrasound image.

[0066] In an embodiment of the present application, performing preliminary quality control on the ultrasound image information includes:

[0067] extracting text information and image information from the ultrasonic image information respectively;

[0068] Comparing and analyzing the text information with the actual text, and calculating the accuracy of the text recognition;

[0069] Performing clarity analysis on the image information.

[0070] Specifically, the integrity of the ultrasound image information is first analyzed, including but not limited to the presence of various textual information. For example, medical images produced by the same device usually contain various textual information, such as device information, patient name, age, ID number, etc. These textual information are first extracted through detection technology and compared with the input standard textual information to avoid misdiagnosis caused by mismatch between the image and the patient.

[0071] Among them, the above detection technology can use OCR text recognition technology, first analyzing the position information of each text in the ultrasound image of different brands of ultrasound equipment, identifying and extracting the text at the corresponding position in the image based on the position information, and then outputting the extracted text information.

[0072] The standard text information indicates the actual text information of each image, such as the patient's name, age or ID number. The recognized text information is compared and analyzed with the actual standard text information, and the accuracy of text recognition is statistically analyzed to determine whether the ultrasound image information to be tested can correspond one-to-one with the patient, thereby reducing misdiagnosis.

[0073] The resolution of ultrasound images is usually very low, and there are often problems of being too bright or too dark. These have a certain degree of impact on image detection. Some even make it impossible for the naked eye to observe some key features, thus affecting the accuracy of the subsequent detection process and causing misdiagnosis. By extracting the histogram of image information and calculating it with the preset standard image histogram, unqualified images are discarded. This process belongs to the preliminary quality control of the image. The image quality is evaluated only from the brightness and contrast of the image. This step is also very necessary because only relatively high-quality images are convenient for further detection and important feature extraction.

[0074] Specifically, the performing clarity analysis on the image information includes:

[0075] Acquiring a histogram of the image information;

[0076] Calculating the distance between the histogram of the image information and a preset standard image histogram;

[0077] Specifically, the preset standard image histogram can be obtained in the following way: first, select a batch of high-quality images, use traditional image processing methods to obtain the two-dimensional pixel matrix of the high-quality images, construct the grayscale histogram of the high-quality images, perform statistical analysis on these grayscale histograms, and obtain the standard image histogram as a reference object for the algorithm.

[0078] If the distance is higher than a preset threshold, the current image information is discarded; if the distance is lower than the preset threshold, the current image information is used as the first image with preliminary qualified quality.

[0079] Specifically, if the distance is lower than a preset threshold, a histogram equalization algorithm may be used to adjust the brightness and contrast of the image, thereby achieving image enhancement and optimization.

[0080] Among them, the preset threshold can be set manually. On the one hand, the effect of the algorithm can be subjectively evaluated by experienced doctors. On the other hand, the original image and the optimized image can be put into the subsequent algorithm for detection to compare and verify the effect.

[0081] In the embodiment of the present application, performing part information quality control on the first image having preliminary qualified quality includes:

[0082] detecting whether features of the part to be diagnosed exist in the first image;

[0083] Specifically, the device needs to input an image of a certain part before it can perform detection. However, during actual operation, due to the technician's error, the input image may not be the image of the specified part, which may easily lead to misdiagnosis in subsequent judgments. Therefore, this step is required to perform automated judgment.

[0084] For example, a deep learning-based multi-classification approach can be used. During the training process, medical ultrasound images of various locations, such as leg vascular ultrasound, carotid artery ultrasound, and cardiac ultrasound, are selected and assigned different labels. Classic classification networks, such as VGG, ResNet, and GoogleNet, are then used to train these images for classification. Alternatively, traditional machine learning classification methods, such as Bayesian models, ensemble algorithms, and clustering, can be used to classify images. The trained model is then tested on a test set, and the accuracy and confusion matrix of the test set are calculated to evaluate the model's performance.

[0085] performing blood vessel site recognition on the first image containing features of the site to be diagnosed;

[0086] Specifically, a target detection network can be used to identify the ultrasound image of the vascular area and automatically extract the predicted ROI (region of interest). At the same time, the corresponding manual ROI is manually extracted in advance, and an advanced target detection network, such as Yolov5 and Faster R-CNN, is used to train the target detection model. The degree of overlap between the predicted ROI and the manual ROI is then used as an evaluation criterion. Once the training is completed, it can be used as a target detection network model to identify the vascular area of ​​the first image containing the characteristics of the area to be diagnosed.

[0087] Determining whether the identified blood vessel portion is a standard section can further reduce the possibility of errors in subsequent detection steps;

[0088] Specifically, determining whether the identified blood vessel portion is a standard section includes:

[0089] Based on the clinical definition of standard sections, we can make a priori knowledge, that is, if a section is deformed or stretched to resemble a standard section, it is still considered a standard section.

[0090] Collect standard section and non-standard section images as a dataset and assign corresponding labels. According to prior knowledge, the standard section includes deformed standard section images.

[0091] A deep learning-based classification method is used for binary classification. The classification network can be selected from VGG, ResNet, GoogleNet, etc., and an attention mechanism can be added to the model's feature extraction. The trained network is tested on a test set, and the binary classification accuracy and confusion matrix are calculated to evaluate the model's performance. When the model's performance meets the preset standard, the identification of the blood vessel can be determined to be a standard section.

[0092] If it is a non-standard section, the vascular position is standardized.

[0093] Specifically, the input image may not be a standard section, in which case standardization is required. Based on prior knowledge, a generative adversarial network is used to generate standardized images, and non-standard images are converted into corresponding standard images without losing information. Generative adversarial networks (GANs) are widely used in image generation. In this application, the data set is from standard and non-standard section images of the same patient and the same period. They are used as input to the model and put into a supervised GAN network for learning, thereby training a model that can automatically generate images corresponding to standard sections with non-standard section images as input. Among them, the performance of the network is evaluated during the training process through methods such as FID score. In addition, after the model is generated, professional doctors can also make subjective evaluations of the generated images, and the quality of the model can be evaluated based on the doctors' feedback.

[0094] In the embodiment of the present application, performing internal imaging quality control on the second image having correct part information includes:

[0095] determining whether there is a clearly visible tunica intima in the second image;

[0096] Specifically, deep learning object detection networks, such as Yolov5 and Faster R-CNN, can be used to obtain the maximum ROI using non-maximum suppression. High-performance object detection models can ensure the clarity of the intima-media layer in the extracted ROI and determine the continuity of the intima-media layer based on the size of the ROI. In other words, the larger the ROI, the clearer and more continuous the intima-media layer is in the image.

[0097] Furthermore, because the intima-media is located within the vessel wall, this application enables simultaneous lumen detection. This multi-task learning strategy prioritizes the network encoder's attention to the intima-media region, thereby improving model performance. The output of this application is a single ROI window, which is used to initially determine the presence of a clear, continuous intima-media region. The trained network is then tested on a test set, and the degree of agreement between the ROI predicted by the target detection network and the standard is measured, which can be used to refine the target detection network model.

[0098] If present, the area covered by the clearly visible intima-media membrane was quantified;

[0099] Specifically, after performing preliminary quality control of the intima-media region, it is necessary to further quantify the explicit coverage of the intima-media in the image. In layman's terms, this refers to the extent of the clearer intima-media region in the image.

[0100] For example, an object detection network can be used to obtain all ROIs. Each ROI represents the region containing the intima-media membrane. After the network extracts the ROIs, it calculates the sum of the horizontal coordinates of the union of the ROI regions. The ROIs selected for the severity classification can be outlined by a professional physician, and the ROI's horizontal coordinate range should be roughly consistent with the intima-media membrane length. This ensures that the ROIs detected by the model can be used to reliably estimate the intima-media membrane length by calculating the horizontal coordinate length. The trained network is then tested on a test set, and the consistency of the ROIs predicted by the object detection network with the standard is measured to further refine the object detection network model.

[0101] The intima-media image is repaired according to the time sequence information.

[0102] Specifically include:

[0103] Acquiring serial ultrasound images;

[0104] Splitting the sequence of ultrasound images into a plurality of ultrasound image frames arranged in time sequence;

[0105] A generative adversarial network is used to perform image generation training on a plurality of ultrasound image frames arranged in time sequence.

[0106] Specifically, sequential ultrasound images consist of a series of ultrasound image frames arranged in time. If information is missing in a frame due to noise or artifacts, it can be restored in other frames. Therefore, using sequential ultrasound images as input, a Generative Adversarial Network (GAN) is trained for image generation, resulting in a clearer image containing more complete information. The network's performance can be evaluated during training using methods such as the FID score. Furthermore, after the model is generated, doctors can subjectively evaluate the generated images. The model's performance can be assessed based on their feedback, allowing for further adjustments.

[0107] In an embodiment of the present application, after repairing the intima-media image according to the time series information, the method further includes: using a classified deep learning convolutional neural network to determine whether there is a plaque.

[0108] Specifically, the dataset uses ultrasound images with and without plaques, labeled with the presence or absence of plaques. A classification-based deep learning convolutional neural network, such as VGG, GoogleNet, and ResNet, is trained. The trained network is then tested on a test set, and the accuracy and confusion matrix of the binary classification are calculated to evaluate model performance. After the intima-media images are restored, the presence of plaques is determined, and the test results can be provided to physicians for reference.

[0109] In an embodiment of the present application, after obtaining a high-quality ultrasound image, it also includes: generating a report based on the ultrasound image, which can generate a more intuitive report based on indicators such as the clarity quality of the input image, blood vessel quality, detection site quality, and intima-media imaging quality.

[0110] Specifically, universal quality control indicators can be obtained based on existing standard image data using clustering-based or manually formulated methods. This technology uses various image quality information, vascular quality, intima-media imaging quality, and other information obtained by other technologies as model input, generates models through text, and generates reports.

[0111] In the training set, the reports are provided by professional physicians in a standardized format containing the required information. The model generates more intuitive reports by inputting image clarity quality, blood vessel quality, detection site quality, and intima-media imaging quality indicators.

[0112] The present application also discloses an ultrasound image quality control system. Figure 2 , the system comprises:

[0113] Ultrasonic image information acquisition module: used to acquire ultrasonic image information;

[0114] A preliminary quality control module is used to perform preliminary quality control on the ultrasound image information to obtain a first image with preliminary qualified quality;

[0115] A part information quality control module is configured to perform part information quality control on the first image that has initially passed quality control, to obtain a second image with correct part information;

[0116] Internal imaging quality control module: used to perform internal imaging quality control on the second image with correct location information to obtain a high-quality ultrasound image.

[0117] The above are all preferred embodiments of the present application and are not intended to limit the scope of protection of this application. Unless otherwise specified, any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features. In other words, unless otherwise specified, each feature is merely an example of a series of equivalent or similar features.

Claims

1. A method for controlling ultrasound image quality, characterized in that: include: Obtaining ultrasound imaging information; Performing preliminary quality control on the ultrasound image information to obtain a first image of preliminary quality; Performing part information quality control on the first image that has initially qualified quality, to obtain a second image with correct part information; performing internal imaging quality control on the second image having correct location information to obtain a high-quality ultrasound image; The preliminary quality control of the ultrasound image information includes: extracting text information and image information from the ultrasound image information respectively; comparing and analyzing the text information with the actual text, and calculating the accuracy of text recognition; and performing clarity analysis on the image information; Performing part information quality control on the first image that has initially qualified quality includes: detecting whether there are features of the part to be diagnosed in the first image; identifying the blood vessel part in the first image that has the features of the part to be diagnosed; determining whether the identified blood vessel part is a standard section; and if it is a non-standard section, standardizing the blood vessel part; Performing internal imaging quality control on the second image with correct location information includes: determining whether a clearly visible intima-media is present in the second image; if so, quantifying the coverage area of ​​the clearly visible intima-media; and repairing the intima-media image according to the time series information; After repairing the intima-media image according to the time series information, the method also includes: using a classified deep learning convolutional neural network to determine whether there is a plaque.

2. The ultrasound image quality control method according to claim 1, wherein: The performing clarity analysis on the image information includes: Acquiring a histogram of the image information; Calculating the distance between the histogram of the image information and a preset standard image histogram; If the distance is higher than a preset threshold, the current image information is discarded; if the distance is lower than the preset threshold, the current image information is used as the first image with preliminary qualified quality.

3. The ultrasound image quality control method according to claim 1, wherein: The determining whether the identified blood vessel portion is a standard section includes: Based on the clinical definition of standard sections, we can make a priori knowledge, that is, if a section is deformed or stretched to resemble a standard section, it is still considered a standard section. Collect standard section and non-standard section images as a dataset and assign corresponding labels. According to prior knowledge, the standard section includes deformed standard section images. Use deep learning based classification method to do binary classification.

4. The ultrasound image quality control method according to claim 1, wherein: The repairing of the intima-media image according to the time sequence information includes: Acquiring serial ultrasound images; Splitting the sequence of ultrasound images into a plurality of ultrasound image frames arranged in time sequence; A generative adversarial network is used to perform image generation training on a plurality of ultrasound image frames arranged in time sequence.

5. The ultrasound image quality control method according to claim 1, characterized in that: After obtaining a high-quality ultrasound image, the method further includes: generating a report according to the ultrasound image.

6. An ultrasound image quality control system, the system using the ultrasound image quality control method according to any one of claims 1 to 5, characterized in that: include: Ultrasonic image information acquisition module: used to acquire ultrasonic image information; A preliminary quality control module is used to perform preliminary quality control on the ultrasound image information to obtain a first image with preliminary qualified quality; A part information quality control module is configured to perform part information quality control on the first image that has initially passed quality control, to obtain a second image with correct part information; Internal imaging quality control module: used to perform internal imaging quality control on the second image with correct location information to obtain a high-quality ultrasound image.

Images