Artificial intelligence quality control method for DR image of occupational pneumoconiosis

Abstract

The invention provides an artificial intelligence quality control method for a DR image of occupational pneumoconiosis, and relates to the technical field of image quality control, and the method comprises the steps: extracting pathological features in the DR image through constructing a sparse convolutional network, automatically distributing high weights for important features in a training process, and achieving the precise capture of the features; a graph attention network is introduced to enhance the spatial perception ability, and it is ensured that the sparse convolutional network can focus on a key area; real-time evaluation is carried out according to a preset quality standard, if a quality problem is found, a feedback mechanism is automatically triggered, related personnel are notified, and a solution is provided, so that DR image acquisition parameters are adjusted in time or image processing is optimized in time. Besides, the quality control standard and the analysis strategy are continuously updated according to the feedback result, so that the accuracy and efficiency of subsequent DR image analysis are improved, and the diagnosis quality of the DR image is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image quality control, and in particular to an artificial intelligence quality control method for occupational pneumoconiosis DR images. Background Art

[0002] Occupational pneumoconiosis is an occupational disease caused by long-term inhalation of harmful particles such as mining dust. Early diagnosis is crucial to controlling the disease. Currently, the use of digital radiographic DR images for the diagnosis of occupational pneumoconiosis has become a common method.

[0003] Existing DR image analysis methods mostly rely on manual judgment and lack effective quality control mechanisms. In practical applications, fluctuations in DR image quality, inaccurate feature extraction, and human factors in DR image analysis can lead to inconsistent diagnostic results. In addition, current DR image processing technology for occupational pneumoconiosis often fails to fully utilize its unique features and relevant clinical information, resulting in multiple problems in the analysis of DR images of occupational pneumoconiosis due to low quality, which affects the accuracy and efficiency of diagnosis. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the shortcomings of the existing technology, the present invention provides an artificial intelligence quality control method for occupational pneumoconiosis DR images, which solves the problem of how to improve the quality control of occupational pneumoconiosis DR images, thereby better analyzing and identifying the symptoms.

[0006] (II) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: an artificial intelligence quality control method for occupational pneumoconiosis DR images, comprising:

[0008] S1 extracts the unique pathological features in occupational pneumoconiosis DR images through a sparse convolutional network; the sparse convolutional network dynamically adjusts the sparsity of the convolution kernel by assigning weights;

[0009] S2 introduces a graph attention network to enhance the spatial capabilities of the sparse convolutional network for each pixel in the DR image, so as to accurately capture the features related to pneumoconiosis; by focusing on key areas, the accuracy of feature extraction is improved;

[0010] After feature extraction, S3 conducts real-time evaluation based on the extracted features and preset quality standards. When the DR image quality is lower than the standard, an alarm is automatically issued through the interactive interface to prompt the operator to make necessary adjustments to ensure that the DR image quality meets the diagnostic requirements.

[0011] S4 Develop quality control standards for DR images of occupational pneumoconiosis based on DR image characteristics and relevant clinical data;

[0012] S5 continuously optimizes DR image analysis and quality control through a real-time feedback mechanism. When quality issues are detected, it updates standards and adjusts analysis strategies in real time to improve the accuracy and efficiency of subsequent DR image analysis.

[0013] Preferably, a sparse convolutional network is used to extract the unique pathological features in the DR images of occupational pneumoconiosis, and a sparse convolutional network is first constructed. The core of this network is to identify and extract key features by learning DR image data. During the construction process, we first initialize the weights of the sparse convolutional network, and use random initialization to assign small random values ​​to the weights of the sparse convolutional network to break the symmetry of the weights and ensure that different neurons can learn different features; the weight values ​​are randomly drawn from a uniform distribution or a normal distribution, for example, a normal distribution with a mean of 0 and a standard deviation of 0.1 is used to initialize the weights; for sparse convolutional networks using the ReLU activation function, the He initialization method is used to adjust the standard deviation according to the number of nodes in the previous layer. The standard deviation is the square root of 2 divided by the number of nodes in the previous layer, so as to ensure that the gradient does not disappear or explode during training; for sparse convolutional networks using tanh or sigmoi The sparse convolutional network with d activation function uses Xavier initialization, which extracts weight values ​​from a uniform or normal distribution with a mean of 0 and adjusts its variance to 2 divided by the sum of the number of nodes in the previous and next layers to ensure that the variance of the activation values ​​and gradients of the previous and next layers remain unchanged during the propagation process; then, the sparse convolutional network architecture is designed, including an input layer, multiple sparse convolutional layers, a pooling layer, and a fully connected layer; each sparse convolutional layer consists of multiple convolution kernels, each of which is responsible for extracting different features in the image.

[0014] Preferably, during the training process, a cross entropy loss function is used to evaluate the difference between the network prediction result and the actual result; the gradient of the loss function with respect to the sparse convolutional network parameters is calculated through the back-propagation algorithm, and the sparse convolutional network weights are updated using the gradient descent method; specifically, the gradient of the loss function with respect to the output of the output layer is first calculated, and then the gradient of the loss function with respect to the weight of each convolution kernel is calculated through the chain rule; for each convolution kernel, the gradient indicates the direction and magnitude of the increase or decrease in the weight so as to reduce the loss during the next forward propagation; a term proportional to the absolute value of the weight is added to the loss function through L1 regularization, so as to prompt the sparse convolutional network to automatically reduce the weights of unimportant features during the training process; the convolution kernel that improves the diagnostic accuracy, that is, the convolution kernel that can identify the specific pathological features of pneumoconiosis, is given a high weight; and the weight of the convolution kernel that is not very helpful for diagnosis will be reduced; by dynamically adjusting the weight, the sparse convolutional network is more focused on the characteristics of occupational pneumoconiosis DR images.

[0015] Preferably, a graph attention network is introduced to enhance the spatial perception ability of the sparse convolutional network. The graph attention network creates a pixel-level graph structure in which each pixel is regarded as a node in the graph; the edges between nodes are defined based on the spatial relationship between pixels, thereby forming a comprehensive spatial connection graph in the sparse convolutional network; the sparse convolutional network can capture the spatial dependency between pixels. In the graph attention network, each node dynamically adjusts its weight according to its relationship with other nodes; different attention coefficients are assigned according to the role of each node in the overall feature extraction; the calculation of the attention coefficient reflects the contribution of each pixel to the overall feature extraction, so that the sparse convolutional network can automatically identify and focus on the feature areas related to pneumoconiosis; first, an attention mechanism is defined, and the attention coefficient of each node relative to all other nodes is calculated in the form of a softmax function; the attention coefficient is used to weight the features of the nodes, thereby generating a weighted feature representation, highlighting the areas that are more critical for diagnosing pneumoconiosis.

[0016] Preferably, after the feature extraction process is completed, the real-time quality assessment stage is automatically entered; according to the quality standards previously developed in cooperation with imaging experts, the quality standards take into account all key requirements for the diagnosis of occupational pneumoconiosis, including quality standards such as DR images should be able to clearly display fine structures of no less than 1 mm, when evaluating the hilar area, the CNR of blood vessels to surrounding lung parenchyma should not be less than 35:1, and DR images should clearly display nodules with a diameter of 2 mm, to evaluate the extracted features; the extracted features are compared with the preset quality standards; by calculating the deviation between the features and the standards, it is possible to analyze in real time whether the DR images meet the minimum requirements for diagnosis; if the analysis results show that certain quality parameters of the DR images do not meet the preset standards, an alarm will be triggered immediately; ensure that relevant personnel take timely action and notify relevant personnel of quality problems through visual prompts or sound warnings on the interactive interface; real-time feedback allows relevant personnel to respond quickly and adjust relevant parameters of DR image acquisition according to the alarm information; relevant personnel take various measures to improve the quality of DR images, including retaking DR images to obtain clearer images, adjusting the exposure time of the X-ray machine to optimize contrast, or enhancing image processing to improve recognizability.

[0017] Preferably, after the DR image analysis is completed, a real-time monitoring phase is entered to perform a quality assessment on the analysis results; the DR image features obtained by the analysis are carefully compared with the preset quality standards, and each analyzed DR image is scored to quantify the gap between it and the standards; when it is detected that the DR image quality does not meet the preset standards, for example, when it is found that the accuracy of feature extraction is insufficient or there are results that affect the diagnosis, a feedback mechanism is immediately triggered; the feedback mechanism quickly conveys the quality problem and its detailed information to the relevant personnel through email, text message or system notification; the relevant personnel who receive the notification take measures promptly according to the detailed report, including adjusting the parameters of the DR imaging equipment, such as exposure time, voltage or focal size, or optimizing image processing, such as enhancing denoising or contrast adjustment functions; once adjustments are made, the DR image will be recaptured or processed and the quality assessment will be performed again; if the problem is solved, the adjustment will be recorded as a reference for similar problems in the future. If the problem persists, feedback will continue until a suitable solution is found; according to the feedback results, its quality control standards and analysis strategies are automatically updated to adapt to new data and conditions to achieve continuous improvement and optimization.

[0018] (III) Beneficial effects

[0019] The present invention provides an artificial intelligence quality control method for occupational pneumoconiosis DR images, which has the following beneficial effects:

[0020] 1. The present invention combines sparse convolutional networks with graph attention networks to accurately extract the unique pathological features in occupational pneumoconiosis DR images, thereby reducing the visual errors caused by manual judgment and improving the consistency and accuracy of diagnosis.

[0021] 2. The present invention can evaluate image quality in real time, automatically trigger an alarm, and promptly notify the operator when the DR image quality does not meet the standard, thereby ensuring that DR image acquisition and processing meet diagnostic requirements and reducing the risk of misdiagnosis due to DR image quality issues.

[0022] 3. The present invention uses feedback and dynamic adjustment strategies to continuously update quality control standards based on real-time analysis results and quality assessments, thereby improving the quality of subsequent DR images and the efficiency and accuracy of analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] Embodiment 1:

[0026] A large amount of occupational pneumoconiosis DR imaging data was collected and standardized, including adjusting image size, normalizing pixel values, enhancing contrast, etc., to ensure the consistency and comparability of the data; a sparse convolutional network was designed and constructed. The sparse convolutional network identified and extracted key features by learning DR imaging data, and automatically assigned weights according to the importance of the features.

[0027] First, the weights of the sparse convolutional network are initialized using random initialization or transfer learning methods; the architecture of the sparse convolutional network is designed, including an input layer, multiple sparse convolutional layers, a pooling layer, a fully connected layer, etc.; each sparse convolutional layer is composed of multiple convolutional kernels, each of which is responsible for extracting different features in the DR image; during the training process, the sparse convolutional network is prompted to automatically reduce the weights of unimportant features during the training process through L1 regularization or sparsity induction technology to achieve the sparsity of the convolutional kernel; the network is trained using a DR image dataset of occupational pneumoconiosis, and during the training process, the network weights are updated using the backpropagation algorithm and the gradient descent method; the validation set is used to evaluate the network. Estimate the performance of the network and adjust the parameters of the sparse convolutional network, such as learning rate and regularization coefficient, according to the evaluation results to improve the generalization ability. When training the sparse convolutional network, the cross entropy loss function is used to evaluate the difference between the network prediction results and the actual results. The cross entropy loss function measures the difference between two probability distributions, one of which comes from the predicted output of the sparse convolutional network and the other from the actual result. The cross entropy loss function is defined as L(y,p)=-[ylog(p)+(1-y)log(1-p)], where y is the true result, which takes the value of 0 or 1; p is the probability that the sparse convolutional network predicts 1.

[0028] Next, the gradient is calculated through the back-propagation algorithm, and the sparse convolutional network weights are updated to reduce the prediction error; the back-propagation algorithm updates the parameters by calculating the gradient of the loss function with respect to the sparse convolutional network parameters, thereby minimizing the loss function; the input data passes through the sparse convolutional network, and each layer calculates the output until the last layer calculates the prediction result; the loss function is used to calculate the difference between the predicted result and the actual result; starting from the output layer, the gradient of the loss function with respect to the sparse convolutional network parameters is calculated, and these gradients are propagated backward layer by layer; the sparse convolutional network parameters are updated according to the gradient using the gradient descent method or other optimization algorithms.

[0029] Take a simple neural network as an example. The network has an input layer, a hidden layer and an output layer. Assume that the hidden layer has a neuron whose output z=w1x+b1, where w1 is the weight, b1 is the bias, and x is the input. The activation function of the neuron is the sigmoid function σ, and the output Assuming that the final loss function L is calculated based on the predicted output p of the output layer and the true result y, the gradient of the loss function L with respect to w1 can be obtained by the chain rule Calculate, where is the gradient of the loss function with respect to the output of the output layer, is the gradient of the output layer output with respect to the output of the activation function, is the gradient of the activation function with respect to its input, is the gradient of the neuron input with respect to the weight; by calculating (1-σ(z)) and Multiplying these gradients gives Once the gradient is calculated, we can use gradient descent Update the weights, where η is the learning rate.

[0030] In order to enhance the spatial perception ability of the network, a graph attention network is introduced; the graph attention network constructs a pixel-level graph structure, with each pixel as a node in the graph; based on the sparse convolutional network, a pixel-level graph structure is constructed, with each pixel as a node in the graph; the edges between nodes are defined according to the spatial relationship between pixels, such as using a Gaussian kernel to calculate the similarity between pixels; the attention coefficient of each node is calculated through the graph attention network to reflect the contribution of each pixel to the overall feature extraction; the attention coefficient is applied to the feature vector to enhance the weight of the feature area related to pneumoconiosis; the enhanced features are fused with the original features to form a richer feature representation.

[0031] After feature extraction is completed, the real-time quality assessment stage begins. Based on the diagnostic requirements of occupational pneumoconiosis and the advice of imaging experts, a series of quality standards are formulated, including image clarity, contrast, resolution, and recognizability of pathological features. Real-time analysis of the extracted features is used to assess whether the DR image meets these standards. If the DR image quality does not meet the preset standards, this method will automatically trigger an alarm mechanism and notify the operator through an interactive interface prompt or a sound warning.

[0032] Design a real-time feedback mechanism to monitor the results of DR image analysis in real time and compare them with the preset quality standards. Once a quality problem is found in the DR image analysis, the feedback mechanism will be triggered to immediately notify the relevant personnel and provide specific information about the problem, including the type of problem and suggested solutions. Based on the feedback information, the relevant personnel can quickly adjust the DR image acquisition parameters or optimize the image processing algorithm to solve the quality problem.

[0033] Design an intuitive user interface that enables operators to easily monitor the status of DR-affected equipment, receive alerts, and make adjustments as needed, and train operators so they understand how this approach works, how to operate it, and how to make adjustments based on alert information.

[0034] Embodiment 2:

[0035] In this embodiment, an artificial intelligence quality control method for occupational pneumoconiosis DR images is applied in the actual operation process of the hospital imaging department; the implementation of this method not only improves the quality of DR images, but also provides strong support for doctors' clinical decision-making.

[0036] In the imaging department of a large hospital, after preliminary technical training and equipment debugging, an artificial intelligence quality control method for occupational pneumoconiosis DR images was officially put into use; whenever a patient undergoes a DR examination, the DR imaging equipment will automatically transmit the image data to this method; after the data is transmitted, the sparse convolutional network immediately performs preliminary processing on the image, extracts features from the DR image, and identifies pathological changes related to occupational pneumoconiosis; this process is real-time, ensuring the timeliness of the diagnosis.

[0037] Once feature extraction is complete, the graph attention network evaluates each pixel and uses its attention mechanism to enhance features related to pneumoconiosis, including unextended shoulder blades and insufficient breathing. In this process, this method can automatically identify key areas in DR images, such as tiny nodules or venous shadows in the lungs, which are important indicators for diagnosing pneumoconiosis. This step significantly improves the accuracy of DR image analysis of occupational pneumoconiosis and provides a more reliable reference for doctors' subsequent judgments.

[0038] After completing the feature analysis, this method evaluates the image in real time according to the preset DR image quality standards; if it is found that the DR image quality does not meet the diagnostic requirements, this method automatically makes adjustments by contacting relevant personnel; this real-time feedback mechanism effectively reduces the risk of misdiagnosis due to DR image quality problems, while also saving doctors' time and improving diagnostic efficiency.

[0039] In clinical applications, this method is not only used for routine DR examinations, but can also be used to analyze historical data and clinical cases to develop personalized quality control standards. For example, for patients in certain high-risk occupations, this method can perform specific analysis and evaluation based on their DR imaging characteristics and medical history. This flexible approach enables the DR image quality control method to play an effective role in different scenarios.

[0040] This method significantly improves the early diagnosis rate of occupational pneumoconiosis by improving the quality of DR images. At the same time, based on the continuous optimization of the real-time feedback mechanism, this method can continuously adapt to new clinical needs and provide efficient technical support for the hospital's imaging diagnosis.

[0041] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence quality control method for occupational pneumoconiosis DR images, characterized in that: The following steps are involved: S1 extracts the unique pathological features in occupational pneumoconiosis DR images through a sparse convolutional network; the sparse convolutional network dynamically adjusts the sparsity of the convolution kernel by assigning weights; S2 introduces a graph attention network to enhance the spatial capabilities of the sparse convolutional network for each pixel in the DR image, so as to accurately capture the features related to pneumoconiosis; by focusing on key areas, the accuracy of feature extraction is improved; After feature extraction, S3 conducts real-time evaluation based on the extracted features and preset quality standards. When the DR image quality is lower than the standard, an alarm is automatically issued through the interactive interface to prompt the operator to make necessary adjustments to ensure that the DR image quality meets the diagnostic requirements. S4 Develop quality control standards for DR images of occupational pneumoconiosis based on DR image characteristics and relevant clinical data; S5 continuously optimizes DR image analysis and quality control through a real-time feedback mechanism. When quality issues are detected, it updates standards and adjusts analysis strategies in real time to improve the accuracy and efficiency of subsequent DR image analysis.

2. According to claim 1, an artificial intelligence quality control method for occupational pneumoconiosis DR images is characterized by: The method of extracting the unique pathological features in occupational pneumoconiosis DR images by using a sparse convolutional network first constructs a sparse convolutional network to identify and extract key features by learning DR image data; During the training process, the sparse convolutional network automatically assigns weights according to the importance of the features. Each convolution kernel of the sparse convolutional network will try to extract different features in the image. Through the back propagation algorithm, the weight of each convolution kernel is adjusted according to the result of the loss function. Convolution kernels that can improve diagnostic accuracy, that is, convolution kernels that can identify pathological features specific to pneumoconiosis, are given high weights; on the contrary, the weights of convolution kernels that are not very helpful for diagnosis are reduced, thereby achieving the sparsity of the convolution kernels; By dynamically adjusting the weights, the sparse convolutional network can be more focused on the characteristics of occupational pneumoconiosis DR images.

3. The artificial intelligence quality control method for occupational pneumoconiosis DR images according to claim 1 is characterized by: The sparse convolutional network extracts features from the DR image through each convolution kernel through forward propagation, and passes these features to subsequent network layers; the sparse convolutional network evaluates the difference between the predicted result and the actual result through a predefined loss function; when the loss function calculates the prediction error, the back propagation algorithm starts to work, and determines the adjustment weight to reduce the prediction error by calculating the gradient of the loss function with respect to the convolution kernel; each layer of the sparse convolutional network is back-propagated until the convolution kernel of the first layer is reached; for each convolution kernel, the gradient indicates the direction and magnitude of the increase or decrease of the weight so as to reduce the loss in the next forward propagation; the sparse convolutional network can automatically focus on the features that are important to the DR image of pneumoconiosis, and this dynamic adjustment of the weight is iterative. As the training proceeds, the sparse convolutional network gradually learns a more effective feature extraction strategy.

4. The artificial intelligence quality control method for occupational pneumoconiosis DR images according to claim 1 is characterized by: The graph attention network is introduced to enhance the spatial perception ability of the sparse convolutional network. The graph attention network constructs a pixel-level graph structure, in which each pixel is a node in the graph, and the edges between the nodes represent the spatial relationship between the pixels. The sparse convolutional network can capture the spatial dependency between pixels, thus more accurately understanding the spatial information in DR images. In the graph attention network, each node dynamically adjusts its importance weight according to its relationship with other nodes, so that the sparse convolutional network focuses on the characteristic areas related to pneumoconiosis. The sparse convolutional network calculates the attention coefficient of each node, which reflects the contribution of each pixel to the overall feature extraction; it enables the sparse convolutional network to automatically focus on the key areas in the DR image. The graph attention network handles the spatial variability between different DR images and adapts to the DR image data of different patients.

5. The artificial intelligence quality control method for occupational pneumoconiosis DR images according to claim 1 is characterized by: After feature extraction is completed, real-time quality assessment will begin. The extracted features will be compared with the pre-set quality standards. The quality standards are formulated based on the diagnostic requirements of occupational pneumoconiosis and the advice of imaging experts. The extracted features will be analyzed in real time to assess whether the DR images meet the standards. When the DR image quality does not meet the preset standards, the alarm mechanism will be automatically triggered, and relevant personnel will be notified through interactive interface prompts or sound warnings. Relevant personnel will adjust DR image acquisition parameters or take other measures in a timely manner based on the alarm information, including reshooting, adjusting exposure time, or enhancing image processing.

6. The artificial intelligence quality control method for occupational pneumoconiosis DR images according to claim 1 is characterized by: The quality control standard for DR images of occupational pneumoconiosis is formulated by collecting a large amount of DR image data of occupational pneumoconiosis and the clinical diagnosis results corresponding to the DR images; determining which DR image features are directly related to the diagnosis of pneumoconiosis through analysis; combining the directly related features with the clinical diagnosis results to determine the minimum requirements for DR image quality, including the clarity, contrast, resolution of DR images and the recognizability of pathological features; considering the factors that affect the quality during the DR image acquisition process, proposing corresponding requirements in the standard, and integrating these requirements into a quality control standard.

7. The artificial intelligence quality control method for occupational pneumoconiosis DR images according to claim 1 is characterized by: The real-time feedback mechanism monitors the results of DR image analysis in real time and compares them with the preset quality standards; once quality problems are found in the image analysis, the feedback mechanism will be triggered; The feedback mechanism notifies relevant personnel immediately and provides specific information about the problem, including the type of problem and suggested solutions. Based on the feedback information, relevant personnel can quickly adjust DR image acquisition parameters or optimize image processing to solve quality problems. Based on the feedback results, the quality control standards and analysis strategies are automatically updated; through continuous learning and adaptation, the accuracy and efficiency of subsequent DR image analysis are improved.

Images