Auxiliary diagnosis method for medical image diagnosis

Abstract

The invention provides an auxiliary diagnosis method for medical image diagnosis, and the method comprises the steps: obtaining a medical image deep learning network model which can be used for recognizing a pathological region, and carrying out the optimization training of a clinical medical image analysis scene for the medical image deep learning network model, thereby obtaining a test model, the diagnosis accuracy of the test model is improved, misdiagnosis and missed diagnosis are reduced, the efficient auxiliary diagnosis function is achieved in the mode of conducting primary diagnosis and secondary diagnosis on the basis of the test model, the diagnosis efficiency and quality are improved, and opportunities are made for treatment of patients.

Description

Technical Field

[0001] The present application relates to the field of smart medical care, and in particular to a method for medical imaging diagnosis and auxiliary diagnosis. Background Art

[0002] In the medical field, imaging diagnosis has always been an extremely important link. It involves the cross-integration of many medical disciplines, such as radiology, ultrasound, pathology and other fields. Medical imaging technology includes X-ray imaging, CT (computed tomography), MRI (magnetic resonance imaging), ultrasound imaging and other types. These medical images provide doctors with intuitive visual information on the internal structure and tissue condition of the human body, which helps in disease detection, diagnosis, treatment planning and disease monitoring.

[0003] Traditional medical imaging diagnosis mainly relies on the professional knowledge and experience of doctors. Doctors observe images directly with their naked eyes and rely on their learning and accumulation of various disease image features to determine whether there are abnormalities in the images and the type of disease that may correspond. For example, when analyzing lung CT images, doctors need to carefully check the lung texture, density, nodules and other features in the images, and make a comprehensive judgment based on clinical symptoms, patient medical history and other information. Since manual film reading is highly dependent on the doctor's subjective experience and professional knowledge level, different doctors may have different judgments on the same image, making it difficult to ensure the accuracy and consistency of the diagnosis. In addition, doctors are prone to visual fatigue when reading films for a long time. When faced with massive image data, they are inefficient and may delay the diagnosis of the disease.

[0004] With the continuous development of modern medical technology, medical imaging data is growing in massive amounts. How to analyze and diagnose these images efficiently and accurately has become a key issue that needs to be solved urgently, and the application of computer technology and artificial intelligence technology in this field is becoming increasingly extensive and in-depth. Although there are studies on AI models for different medical image recognition in the prior art, most of them are only for specific types of image recognition. For example, the lung nodule recognition model can only identify possible lung nodules in medical images, and lacks the ability to intelligently identify and deeply analyze the image content. In other words, the auxiliary technology of medical image recognition in the prior art lacks the ability to deeply mine and intelligently analyze image data, and cannot automatically extract key features in the image and conduct comprehensive correlation analysis. For example, it is difficult to accurately identify the complex relationship between the tiny lesion area and the surrounding tissue in the image, and cannot make full use of the information contained in a large amount of historical image data to assist doctors in completing the diagnosis of current medical images. Summary of the invention

[0005] The present application provides a method for auxiliary diagnosis of medical imaging, which obtains a test model by optimizing the training of the medical imaging deep learning network model in clinical scenarios, thereby improving the diagnostic accuracy of the test model while reducing misdiagnosis and missed diagnoses. The method can be applied to a variety of clinical scenarios, and can achieve efficient auxiliary diagnosis functions through primary and secondary diagnosis, thereby improving the efficiency and quality of diagnosis and gaining opportunities for patient treatment.

[0006] To achieve the above objectives, in a first aspect, this solution provides a medical imaging diagnosis auxiliary diagnosis method, comprising the following steps:

[0007] S1: Obtaining a deep learning network model for medical imaging:

[0008] A medical imaging deep learning network model for obtaining recognition results for identifying different pathological areas of medical images;

[0009] S2: Build a test model:

[0010] Constructing an initial test model that matches the deep learning network model of the medical image, initializing the state of each network node of the initial test model and setting a state threshold of each network node, inputting a medical image with regional features of at least one known pathological area into the initial test model to associate and update the regional features of the pathological area in the medical image with the corresponding network node, until the state of each network node is greater than the set state threshold, then outputting the output value of each pathological area of ​​the current medical image, updating the initial test model based on the output value of the target deep learning model corresponding to the pathological area and the output value of the initial test model to obtain a test model;

[0011] S3: One test:

[0012] Initialize the state of each network node of the test model and set the state threshold of each network node, input the medical image to be tested into the test model to obtain the output value of each pathological area, assign and update the state of the corresponding network node based on the output value of each pathological area, generate a first memory result based on the learning state of the network node when the state of the network node is greater than 0, and output the output value of the pathological area whose state of the network node is greater than the set state threshold as a diagnosis and recognition result;

[0013] S4: Secondary test:

[0014] An initial test model that matches the deep learning network model of medical images is constructed, the status of each network node of the initial test model is initialized and the status threshold of each network node is set, the medical image to be tested is input into the initial test model to traverse each network node of the initial test model, the pathological area of ​​the medical image to be tested is associated with each network node and the neurons of the network nodes are assigned with the information status of each network node, when the status of the network node in the associated path in the output value of the initial test model is less than 0, the corresponding network nodes are repeatedly assigned with the output value until the network node status corresponding to each pathological area of ​​the output value is greater than 1 and a second memory result is generated based on the learning state of the network node, and the second memory result is assigned and optimized using an artificial intelligence algorithm to obtain a secondary diagnosis result.

[0015] Different from the prior art, the main contributions and innovations of the present invention are as follows:

[0016] 1. Use the convolutional neural network algorithm to build a deep learning network model for medical images, automatically extract key regional features such as lung nodules and blood vessels in lung CT images, and compare the recognition results with the pathological diagnosis results during training to optimize the parameters to ensure that the model's image recognition is highly consistent with the actual pathology. For example, it can accurately identify tumor size, shape, density and other features in tumor images, making the diagnosis more accurate.

[0017] 2. The constructed medical image information database stores rich correspondences between images and pathologies, deeply calculates and stores complex correlations between image regions, such as accurately calculating the weight relationship between the network nodes of cardiovascular disease regions and myocardial status regions to form correlation node pairs. When new images to be tested are input, historical cases can be quickly associated based on these stored correlation information, providing doctors with multi-dimensional and in-depth references to help them analyze the condition from a more comprehensive perspective.

[0018] 3. Construct a test model that matches the deep learning network model of medical imaging, and input a variety of medical images containing different regional features (such as tumor images, bone images, liver ultrasound images, breast X-ray images, etc.) into the initial test model for training. Through continuous iterative optimization training, the model continuously optimizes its own performance and has strong generalization capabilities. Whether it is facing common diseases or relatively rare disease imaging diagnosis, it can stably play an auxiliary role in complex and diverse medical diagnosis scenarios and provide effective support for the diagnosis of different diseases.

[0019] 4. Use the trained test model to conduct a secondary diagnosis on the corresponding area of ​​the image to be tested. Taking liver ultrasound images as an example, first initialize the node state of the test model in a test, input the image to obtain the output value of the pathological area and update the node state, generate the first memory result and output the first diagnosis recognition result; then initialize the model again in the second test, re-input the image-related area features and nodes and assign values ​​to update, generate the second memory result and optimize the artificial intelligence algorithm to obtain the secondary diagnosis result. During the operation of the test model, the first memory result generated by a test integrates the output value of the network node when processing the image, and records in detail the key information such as the node state change and the response to the features, reflecting the learning process and results of the model on the current image; the second memory result generated by the second test integrates the current and historical states of the network nodes. These memory results provide rich and reliable basis for diagnosis from different aspects, enhance the scientific nature of the diagnosis results, and help doctors make more accurate and reasonable diagnosis judgments.

[0020] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0022] Figure 1 It is an overall logical schematic diagram of the medical imaging diagnosis auxiliary diagnosis method provided by this scheme.

[0023] Figures 2 to 4 It is a diagram of the specific steps of the medical imaging diagnosis auxiliary diagnosis method provided by this scheme. DETAILED DESCRIPTION

[0024] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with one or more embodiments of this specification. Instead, they are merely examples of devices and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0025] It should be noted that: in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the steps included in the method may be more or less than those described in this specification. In addition, a single step described in this specification may be decomposed into multiple steps for description in other embodiments; and multiple steps described in this specification may be combined into a single step for description in other embodiments.

[0026] Embodiment 1

[0027] like Figure 1 to Figure 4 As shown, the embodiment of the present application provides a method for assisting medical image diagnosis, which improves the effectiveness of the medical image deep learning network model in complex medical scenarios by making it better adapt to the needs of actual medical diagnosis, including the following steps:

[0028] S1: Obtaining a deep learning network model for medical imaging:

[0029] A medical imaging deep learning network model for obtaining recognition results for identifying different pathological areas of medical images;

[0030] S2: Build a test model:

[0031] Constructing an initial test model that matches the deep learning network model of the medical image, initializing the state of each network node of the initial test model and setting a state threshold of each network node, inputting a medical image with regional features of at least one known pathological area into the initial test model to associate and update the regional features of the pathological area in the medical image with the corresponding network node, until the state of each network node is greater than the set state threshold, then outputting the output value of each pathological area of ​​the current medical image, updating the initial test model based on the output value of the target deep learning model corresponding to the pathological area and the output value of the initial test model to obtain a test model;

[0032] S3: One test:

[0033] Initialize the state of each network node of the test model and set the state threshold of each network node, input the medical image to be tested into the test model to obtain the output value of each pathological area, assign and update the state of the corresponding network node based on the output value of each pathological area, generate a first memory result based on the learning state of the network node when the state of the network node is greater than 0, and output the output value of the pathological area whose state of the network node is greater than the set state threshold as a diagnosis and recognition result;

[0034] S4: Secondary test:

[0035] An initial test model that matches the deep learning network model of medical images is constructed, the status of each network node of the initial test model is initialized and the status threshold of each network node is set, the medical image to be tested after a test is input into the initial test model to traverse each network node of the initial test model, the pathological area of ​​the medical image to be tested is associated with each network node and the neurons of the network node are assigned with the information status of each network node, when the status of the network node in the associated path in the output value of the initial test model is less than 0, the corresponding network nodes are repeatedly assigned with the output value until the status of the network node corresponding to each pathological area of ​​the output value is greater than 1, based on the learning status of the network node combined with the first memory result of the corresponding network node, a second memory result is updated and generated, and the second memory result is assigned and optimized using an artificial intelligence algorithm to obtain a secondary diagnosis result.

[0036] Regarding step S1:

[0037] In step S1, this solution first uses the convolutional neural network algorithm to conduct deep learning on various medical images to build a medical image deep learning network model, so as to obtain the recognition of various pathological areas in various medical images. The medical image deep learning network model has a strong feature extraction capability and can automatically learn the key features in medical images, such as identifying different areas such as lung nodules and blood vessels in lung CT images.

[0038] Specifically, during the training process of the deep learning network model for medical images, a large number of medical image samples are input, such as different types of tumor images, bone images, etc., so that the convolutional neural network can learn the characteristic patterns of different pathological areas in the medical images, such as the shape, size, density and other characteristics of the tumor, as well as the structure, texture and other information of the bone. In addition, during the training process of the deep learning network model for medical images, the recognition results of each pathological area in various types of medical images are compared with the corresponding recognition data in the pathological diagnosis results to obtain error data, and the parameters of the convolutional neural network algorithm are optimized based on the error data, and saved as a deep learning network model for medical images after meeting the clinical application conditions. For example, if there is a deviation between the size of the tumor identified by the algorithm and the actual size in the pathological diagnosis results, the weights and thresholds and other parameters in the convolutional neural network will be adjusted according to this deviation to improve the accuracy of the algorithm.

[0039] The medical image deep learning network model provided by this solution is a multi-layer convolutional neural network, including convolutional layers, pooling layers, and fully connected layers connected in sequence. In the training process of the medical image deep learning network model, a large number of medical image samples with features of pathological regions annotated are used as training samples, such as lung CT images, liver MRI images, etc. Each sample image contains specific feature labels, such as the shape, size, density distribution, etc. of the tumor, so that the medical image deep learning network model can learn feature mapping relationships from them through forward propagation, wherein the convolution layer extracts local features in the image through filters, and the pooling layer further reduces the dimension of the features to reduce the computational complexity, and finally completes classification or regional segmentation through the fully connected layer, and the output of the medical image deep learning network model is the recognition result of each pathological region in the medical image.

[0040] In some embodiments, the features of the pathological area in the training sample are compared with the prediction results of the medical image deep learning network model, the error is calculated and the parameters of the medical image deep learning network model are optimized. Specifically, the optimization process is based on the error function, and the commonly used loss function is the cross entropy loss or the mean square error. Let y i is the real pathological area result. is the prediction result of the deep learning network model of medical images, and the loss function can be expressed as

[0041]

[0042] Where N is the total number of training samples, K is the number of classification categories, and y i k represents the true label of the i-th sample belonging to category k, is the predicted probability of the i-th sample corresponding to category k. Through the gradient descent algorithm, the parameters of the model (such as the weight and bias of the convolution kernel) are optimized and updated according to the loss function. The iterative update formula is:

[0043]

[0044] where w (t) represents the model parameters of the tth iteration, η is the learning rate, is the gradient of the loss function with respect to the parameters. Through multiple iterations, the model error is gradually reduced until the preset clinical application conditions are met. The goal of this optimization process is to make the recognition results of the medical image deep learning network model for medical images closer to the actual data of pathological diagnosis, thereby improving the accuracy and stability of the medical image deep learning network model.

[0045] Regarding step S2:

[0046] This solution builds a test model based on the deep learning network model of medical images, and then uses the test model to test the medical images to be tested. Although the deep learning network model of medical images constructed in step S1 can identify pathological areas, it is a relatively general model and needs to be optimized and adapted for specific application scenarios and needs by building a test model to provide verification and optimization for subsequent diagnosis.

[0047] Specifically, an initial test model matching the medical imaging deep learning network model is constructed, and the weights w of all network nodes of the initial test model are initialized. ij , set the state threshold θ of each network node j , and set the state of each network node to s j = 0. Through this initialization operation, it can be ensured that all network nodes of the initial test model are in a unified starting state to avoid interference caused by random states. The settings of the weights and state thresholds of the network nodes are usually predefined according to the network structure and training data. For example, the initial value of the weight adopts uniform distribution or Gaussian distribution, and the initial value of the state threshold is set to the median range of the activation function. The initial settings of the weights and thresholds will be continuously adjusted in the subsequent model training process.

[0048] In an embodiment of the present scheme, a medical image with known regional features of a pathological area is represented as a two-dimensional or three-dimensional pixel matrix, wherein each pixel in the pixel matrix represents the image information of the pathological area, and the regional features of the corresponding pathological area are extracted by performing a convolution operation on the pixel points.

[0049] Specifically, the input medical image I(x, y) is a two-dimensional or three-dimensional pixel matrix, and each pixel point I xy The convolution calculation formula for extracting features by convolution operation is:

[0050]

[0051] Among them, h ij is a point in the feature map, w mn is the weight of the convolution kernel, σ is the activation function (specifically the ReLU function), and b is the bias term. Through convolution calculation, the regional feature values ​​of different pathological areas in the image are extracted.

[0052] The regional features of the pathological area in the medical image are associated with the corresponding network nodes. Each network node in the initial test model is designed to process a specific type or aspect of information. Associating the extracted pathological area features with the corresponding network nodes means allocating or mapping the various feature values ​​of the regional features to the corresponding network nodes according to certain rules, and updating them as input values ​​of the network nodes. Once the association is established, the feature values ​​of the pathological area can be transmitted and processed along the connections between the nodes in the neural network. The network nodes operate on the input feature values ​​according to their own calculation rules (such as weighted summation, activation function processing, etc.), thereby gradually extracting higher-level feature information or performing tasks such as classification judgment.

[0053] At the same time, the state of the network node is updated and adjusted in real time according to the input feature value and the pre-set rules. The node state is an internal representation of the network node in the process of processing information. It reflects the degree of response or activation of the node to the input information. When the node state S j When it exceeds 1, it indicates that the network node has been fully activated and can further process the features of the pathological area. Through this dynamic adjustment, the network node can gradually learn the feature distribution of the input image.

[0054] In other words, this solution associates the state of each network node in the initial test model with the image information of each pathological area in the medical image based on the state of each network node in the initial test model and the image information of each pathological area in the medical image, and uses the information state of each network node in the initial test model to assign values ​​to the neurons of each network node in the initial test model. In this way, the test model can adjust its own neuron state according to the image information and gradually learn the characteristics of the image.

[0055] Correspondingly, the characteristic values ​​of the regional features of the pathological area in the medical image are mapped to the corresponding network nodes as the input values ​​of the current network nodes, and the states of the network nodes are updated until the state of each network node is greater than 1. In this scheme, if the state S of the associated network node is j If it is less than 0, the following formula is used to update the status of the network node:

[0056]

[0057] Among them, η is the learning rate, which is used to control the update amplitude, and h i is the eigenvalue of the corresponding regional feature.

[0058] Then, this scheme compares the state of the network node with the set state threshold θ j , if the state S j Exceeding the set state threshold θ j, then the output value of the current pathological area is output, where the output value corresponds to the diagnosis result of the current pathological area, judging whether there is a lesion or abnormal area. j Does not exceed the set state threshold θ j The output value of the network node remains unchanged. This output rule can be achieved through the logic function o j =II(s j >θ j ) is implemented, where II represents the indicator function and the output value O j It is the binary form of the diagnosis result.

[0059] In addition, this scheme additionally includes the step of determining whether the state of each network node in the associated path of each pathological area in the above-mentioned output value is less than 0. If so, assigning a value to each pathological area of ​​the above-mentioned output value until the state of each network node is greater than 0; and using the output value of each network node to obtain the first memory result of each network node, wherein the first memory result can record some key information of the test model in the process of processing the image, which is used for subsequent model optimization and adjustment.

[0060] Specifically, the first memory result synthesizes the output values ​​of each network node in the process of processing the image. These output values ​​include the learning and understanding of the nodes on the characteristics of different pathological regions in the image. It records key information such as the state changes of each node and the response to the characteristics when the test model processes the image. This information reflects the learning process and results of the model on the current image. For example, for a complex image containing multiple pathological features, the first memory result can record the node response of the model to different feature areas (such as nodules of different sizes and densities, and changes in surrounding tissues, etc.), including which nodes are sensitive to specific features and which nodes have undergone large state changes during the processing process. These recorded key information is crucial for subsequent model optimization and adjustment. During the training process of the test model, by analyzing the first memory results of a large number of medical images, the advantages and disadvantages of the test model in processing different types of pathological features can be found. For example, if it is found that the test model always produces a low state or unstable output on certain network nodes when processing images of a specific disease, the parameters of these network nodes (such as weights, biases, etc.) can be adjusted in a targeted manner, or the structure and training methods of the test model can be improved, thereby improving the accuracy and generalization ability of the test model. The first memory results provide a basis for the continuous improvement of the model, enabling the model to continuously adapt to different medical imaging data and provide more reliable diagnostic assistance.

[0061] In order to further optimize the recognition accuracy of the test model, this solution selects the trained target deep learning model corresponding to the pathological area to verify and optimize the output results of the initial test model. The target deep learning model uses its own more accurate weights and feature relationships to analyze the states of each node obtained after the initial test model processes the medical image to determine whether it accurately reflects the pathological features in the medical image. For example, for a lung CT image, the initial test model may misjudge the boundary of a lung nodule. Through the correction of the target deep learning model, the initial test model can gradually adjust the weights and thresholds, and finally accurately identify the morphology and location of the nodule.

[0062] In some embodiments, the target deep learning model corresponding to the pathological area is a convolutional neural network built with more precise weights and feature relationships, and is trained using medical imaging images that mark the pathological area to ensure that the target deep learning model can accurately learn the characteristics of the pathological area. For example, for a target deep learning model dedicated to lung nodules, the training data will focus on the various morphological, size, density and other characteristic information of lung nodules in lung images, and will be annotated with accurate diagnostic information such as nodule boundaries and properties.

[0063] The target deep learning model provided by this solution also uses a convolutional neural network structure similar to that in the deep learning network model of medical images, including convolutional layers, pooling layers, fully connected layers, etc. The convolutional layers extract local features of the image, the pooling layers reduce the feature dimensions, and the fully connected layers perform operations such as classification or regional segmentation, thereby achieving accurate recognition and analysis of pathological areas. During the training process, the model's predicted results will also be compared with the actual results of pathological diagnosis, the error will be calculated, and algorithms such as gradient descent will be used to optimize the model's parameters (such as the weight and bias of the convolution kernel, etc.) to continuously improve the model's recognition accuracy and accuracy of pathological area features. For example, by adjusting the weights, the model can more accurately determine the boundaries and properties of lung nodules and improve the ability to detect small lesions.

[0064] In this solution, the same medical image is input into the target deep learning model to obtain the output value. Based on the difference between the output value of the target deep learning model corresponding to the pathological area and the output value of the initial test model, the weights and bias items of the initial test model are adjusted by the back propagation algorithm to obtain the test model, so that it is consistent with the result of the target deep learning model. In some embodiments, the initial test model is set to be trained using the gradient descent method, and the loss function is set Among them, j is the output of the test model, m j is the output of the target model, and the loss value is reduced through iterative optimization, thereby improving the performance of the initial test model.

[0065] The reason why this scheme uses the output results of the target deep learning model to correct the initial test model is that the target deep learning model is usually highly optimized for specific pathological areas. For example, the lung nodule-specific model performs well in processing lung nodule-related images, but may have poor adaptability to other pathological areas (such as liver lesions, bone diseases, etc.). The constructed test model is trained by inputting a variety of medical images containing characteristics of different regions (such as tumor images, bone images, liver ultrasound images, breast X-ray images, etc.). It has stronger generalization ability and can stably play an auxiliary role in complex and diverse medical diagnosis scenarios. It is suitable for the diagnosis of various diseases.

[0066] In actual clinical scenarios, a test model that comprehensively considers multiple factors and pathological areas is needed. After comparison and optimization with the target deep learning model, the initial test model can better adapt to specific clinical application needs, integrating the advantages of the target deep learning model into a more general test model, so that it can not only accurately identify specific pathological areas, but also provide more comprehensive and accurate auxiliary diagnosis results in the overall medical imaging diagnosis. At the same time, by optimizing the initial test model based on the target deep learning model, the function of the test model can be more flexibly adjusted and expanded according to different clinical needs and emerging pathological conditions. For example, when a new type of disease emerges or an existing model needs to be improved, the test model can be updated by adjusting some features of the target deep learning model or retraining the target model, rather than relying entirely on a single target deep learning model. This approach can better cope with the changing diagnostic needs in the medical field and improve the practicality and adaptability of the model.

[0067] In addition, this solution uses multiple medical images containing different regional features as input into the initial test model for continuous training to improve the generalization ability of the test model. Through this iterative optimization, the test model can adapt to different types of medical image inputs and improve its applicability in complex medical scenarios.

[0068] Regarding step S3: Testing the test model using the collected medical images to be tested is a key step in evaluating the recognition ability of the test model and optimizing the diagnostic accuracy.

[0069] In the step of “initializing the state of each network node of the test model and setting the state threshold of each network node”, each network node of the test model is initialized and the state threshold is set, and the state s of all network nodes are initialized. j Assigning a value of 0 ensures that each test starts from a unified initial state. This initialization method avoids the interference of historical states on the current test and ensures the comparability of test results and the scientific nature of analysis.

[0070] In the step of "inputting the medical image to be tested into the test model to obtain the output value of each pathological area, and assigning and updating the state of the corresponding network node based on the output value of each pathological area", the medical image to be tested is input into the test model to obtain the regional features of each pathological area, and the regional features are extracted as feature values ​​f k =[f1,f2,…,f n ] is used as the output value of each pathological area and is assigned to the state of the corresponding network node according to the association rule of the network node. The update formula of the state of the network node is:

[0071]

[0072] where w jk is the weight of network node j and regional feature k in the test model, f k is the characteristic value of the regional feature, corresponding to the output value, is the initial state of the network node. Through this assignment operation, the characteristics of the pathological area are effectively mapped to the network node, providing a basis for subsequent testing.

[0073] Specifically, this solution inputs the medical image to be tested into the above test model, and assigns the state of each network node according to the output value of each pathological area of ​​the medical image to be tested, and obtains the state of each network node. For example, for liver ultrasound images, the corresponding network node state is assigned according to the regional recognition results such as the morphology and echo of the liver in the image.

[0074] In the step of "generating a first memory result based on the learning state of the network node when the state of the network node is greater than 0", the state of each network node is judged. When the state of the network node is less than 0, the state of the network node is cumulatively updated using the output value of the corresponding pathological area. The formula is as follows:

[0075]

[0076] Where η is the learning rate, which controls the update step. represents the state of the network node after the t+1th iteration. The update process continues. jk is the weight of network node j and regional feature k in the test model, f k is the characteristic value of the regional feature;

[0077] When the state of the network node is greater than 0, the learning state of the network node is used to generate the first memory result m j , the formula is as follows:

[0078]

[0079] Where α is the memory update coefficient, It is the historical memory value of the network node.

[0080] Such an update mechanism in this scheme can enable the test model to more accurately identify the features in the medical images to be tested. The first memory result integrates the output values ​​of each network node in the process of processing medical images, records key information such as the state changes of each network node when the test model processes medical images and the response to the features, reflects the learning process and results of the test model on the current image, and provides a basis for model optimization and adjustment in the secondary test.

[0081] In the step of "outputting the output value of the pathological region whose state of the network node is greater than the set state threshold as a diagnosis and recognition result", the node state s of the network node is j With the state threshold θ j Compare, if the node status s j Greater than the state threshold θ j , then the corresponding pathological area in the medical image to be tested meets the recognition conditions, and the output value of the current pathological area is output. For example, it can be determined whether the area belongs to normal tissue or has abnormal characteristics; if the node state s j Not greater than the state threshold θ j , then the pathological area is considered not to meet the identification conditions.

[0082] This judgment process of this scheme is realized through a logical judgment function, such as

[0083] o j =II(s j >θ j ),

[0084] Among them, j Represents the output binary judgment result, 1 means it meets the conditions, and 0 means it does not meet the conditions. Through comprehensive analysis of all node states, the specific diagnosis results of each pathological area in the image to be tested can be determined.

[0085] Take a liver ultrasound image as an example. If the characteristic f of a pathological area in the liver ultrasound image k Showing obvious echo enhancement and boundary blur, the test model maps its features to network node j and calculates the node status s according to the node state. j With the threshold value θ j The output diagnosis result is a high-risk area for liver fibrosis. In addition, the first memory result m j The characteristic changes of the pathological area can be stored to provide support for trend analysis of future images. Through this cyclic input, state update and output judgment process, the test model can make full use of the characteristic information of the image to be tested, improve the recognition accuracy of the model, and show better reliability and applicability in actual diagnostic scenarios.

[0086] Regarding step S4: the initial test model is used to perform a second test on the corresponding area in the medical image to be tested after the first test, in order to confirm and optimize the potential lesion area in the medical image to be tested through a more in-depth analysis. It should be noted that in step S4, the initial test model is rebuilt and a series of operations are performed to further analyze and diagnose the medical image to be tested from different angles to improve the accuracy and reliability of the diagnosis results.

[0087] In the step of "constructing an initial test model that matches the medical imaging deep learning network model, initializing the state of each network node of the initial test model and setting the state threshold of each network node", the weights w of each network node of the initial test model are set. ij and state threshold θ j , and the state of the network node s j Initialized to 0, this initialization provides a unified starting point for the secondary diagnosis operation of the model, ensuring the consistency and reliability of the secondary diagnosis process.

[0088] In the step of "inputting the medical image to be tested after one test into the initial test model to traverse each network node of the initial test model, associating the pathological area of ​​the medical image to be tested with each network node and assigning the neurons of the network node with the information state of each network node", the medical image to be tested after one test is segmented into multiple pathological areas, each pathological area is feature extracted to obtain the feature value of the corresponding regional feature, and the feature value is mapped to the network node of the initial test model through the feature allocation rule, which is used to update the input value in each network node of the initial test model, so that the initial test model can accurately assign neurons and update input values ​​according to the regional features, wherein the feature allocation rule is:

[0089]

[0090] where w j k represents the weighted connection strength between network node j and regional feature k, is the initial state of network node j. This formula reflects the priority allocation effect of network weights on features and maps regional features to network nodes through linear combinations.

[0091] It should be noted that the initial test model traverses all network nodes one by one to check whether there are any missed state updates. For each network node j, the state of the network node is adjusted according to its associated regional characteristics, and the iterative update state is as follows:

[0092]

[0093] Where η is the learning rate, which controls the state update step of each iteration, and the function Represents the gradient of state update. This update process ensures that the model can dynamically adjust the state of network nodes and gradually strengthen the learning of regional features.

[0094] In the step of "when the state of the network node in the associated path in the output value of the initial test model is less than 0, repeatedly assigning the corresponding network nodes with the output value", it is determined whether the state of the network node in the associated path in the output value of the initial test model is less than 0. If so, the current output value continues to be assigned to the network node by accumulation until the state of the network node of each regional information in the above output value is greater than 1. If not, the output value of each network node is used to obtain the second memory result of each network node. At this time, the training is stopped and the output value of the updated initial test model is used to perform the first correction on the input end of the above initial test model. In this way, the output result of the model is continuously optimized, and the input end is corrected to improve the accuracy of the model.

[0095] In some embodiments, it is continuously and repeatedly determined whether the state of the network node in the associated path in the output value of the initial test model is less than 0. If so, the current output value is continuously assigned to the network node by accumulation until the state of the network node of each area information in the above output value is greater than 1. This setting takes into account that in actual operation, there may be new situations that cause the node state to change again. In the case of the node state, it is necessary to determine again whether the state of the network node in the associated path is less than 0. If so, continue to assign and update, so as to more strictly guarantee the accuracy and stability of the output results, and ensure that the state of each network node always meets the preset diagnostic requirements of the model.

[0096] In the step of "updating and generating a second memory result based on the learning state of the network node in combination with the first memory result of the corresponding network node", the update formula of the second memory result is as follows:

[0097]

[0098] Where β is the memory update coefficient, is the first memory result, s j Is the current state of the network node.

[0099] The secondary diagnosis of this scheme retains the learning process and result information of the model recorded in the first memory result when processing the image before, and incorporates new information of the current state, so that the second memory result can more comprehensively reflect the comprehensive information of the network nodes in the entire secondary diagnosis process, thereby providing a richer and more accurate basis for subsequent diagnosis.

[0100] Specifically, the second memory result is the comprehensive memory information after the model deeply learns and analyzes the image during the secondary diagnosis process. In the process of generating the second memory result, the model reconstructs the initial test model, re-inputs the image to be tested, and performs a series of node association, state update and other operations, especially for the case where the network node state in the association path is less than 0, the node state is repeatedly assigned until it is greater than 1. This process enables the model to more comprehensively capture the pathological feature information in the medical image to be tested. This information is integrated into the second memory result, reflecting the model's deeper understanding and learning status of the image features of the medical image to be tested.

[0101] In the step of "optimizing the second memory result by using an artificial intelligence algorithm to assign values ​​to obtain a secondary diagnosis result", an artificial intelligence algorithm is used to use the second memory result to adjust the model to judge and evaluate the pathological area of ​​the medical image to be tested.

[0102] The artificial intelligence algorithms mentioned in this solution are used to analyze, process and optimize the methods and technologies of the second memory results. For example, the data mining and pattern recognition technologies in the machine learning algorithm are used to conduct in-depth analysis of the information in the second memory results to further optimize the diagnosis results and improve the model's ability to identify abnormalities in the image area, thereby obtaining more accurate secondary diagnosis results.

[0103] In some embodiments, the artificial intelligence algorithm corrects and improves the output results based on information such as the strength of association between the features of different pathological regions recorded in the second memory results, the response of network nodes to the features, etc., so that the final secondary diagnosis results will more accurately reflect the actual pathological conditions in the image, improve the reliability and accuracy of the diagnosis results, and thus achieve the ultimate goal of the scheme to provide accurate output results. In other words, the second memory result, as a carrier of intermediate information, provides key data support and decision-making basis for the optimization of output results. This scheme can integrate the current state and historical state of the network node by assigning the second memory result to generate a diagnostic memory with continuity and stability. By processing the second memory result, the model can further optimize the reliability of the diagnosis results and improve the ability to identify abnormalities in the image area.

[0104] In other embodiments, the present solution additionally includes the step of constructing a medical imaging information database. The present solution constructs a medical imaging information database by using the feature information output by the medical imaging deep learning network model as the medical imaging information, establishes the association between the regional features of the medical imaging and the network nodes, and stores them in a structured manner as a database resource that can be queried and referenced. Furthermore, by establishing a medical imaging information database, similar medical images can be matched when the medical image to be tested is input, so as to assist medical personnel in making a diagnosis in combination with the diagnostic results of the similar medical images.

[0105] Specifically, a medical imaging deep learning network model is used to identify each medical image to obtain the identification results of each pathological area, a region identifier corresponding to the current pathological area is generated based on the identification results, associated node pairs are obtained based on the region identifier, and the region identifiers and associated node pairs are stored in the form of structured data as a medical imaging information library.

[0106] In the step of "using the medical image deep learning network model to identify each medical image to obtain the identification results of each pathological area, and generating the regional identification corresponding to the current pathological area based on the identification results", the intermediate layer of the pathological area identification model is used to extract the regional features of the pathological area of ​​each medical image as the identification result, and the regional identification of the current pathological area is generated based on the connection weight area between the regional features and the network nodes corresponding to the current pathological area.

[0107] This solution uses the recognition results of each pathological area by the medical image deep learning network model to generate a regional identifier for each pathological area to ensure that the characteristics of different pathological areas can be accurately stored and distinguished. Specifically, the regional characteristics of each pathological area can be represented as a vector f i =[f i1 ,f i2 ,...,f im ], where f i represents the feature of the i-th pathological region, m is the dimension of the region feature, and the region identifier of the current pathological region is generated based on the region feature and the connection weight between the corresponding network nodes. For example, for cardiac imaging, the model may identify multiple pathological regions such as cardiac chambers, heart valves, and blood vessels, and generate specific region identifiers for distinction and management in the medical imaging information database.

[0108] In the step of "obtaining associated node pairs based on region identification", the weight values ​​of the network nodes in the current pathological region and other network nodes are calculated, and other network nodes with large weight values ​​are selected to form associated node pairs with the network nodes in the current pathological region.

[0109] Specifically, by calculating the weight relationship R between the network node in the current pathological area and other network nodes ij =w ij ·f j , the region identifier of the node can be associated with the feature vector of its associated node, where R ij is the relationship weight between network node i and network node j, reflecting the relevance and importance of the network node with other network nodes. For example, in cardiac imaging, assuming that the weight w between the network nodes of the vascular region and the cardiac cavity region identified by the medical imaging deep learning network model is ij If it is higher, it may indicate that vascular lesions have a significant impact on cardiac chamber function. By recording this relationship, it can provide relevant basis for subsequent diagnosis.

[0110] This scheme obtains other associated network nodes in the corresponding network node in the deep learning network model of medical images based on the region identification, and then associates the region identification and the corresponding weights of other network nodes with the neuron state characteristics in the other associated network nodes to form associated node pairs. This helps to establish a complex relationship model between pathological areas. For example, cardiovascular lesions may be associated with the state of the myocardium. Through the storage of this associated information, a more comprehensive reference can be provided for diagnosis.

[0111] In the step of "storing the region identifiers and associated node pairs in the form of structured data as a medical imaging information library", the region identifiers and corresponding region features are standardized and stored to facilitate subsequent query and retrieval.

[0112] When a new medical image to be tested is input, the relevant reference images and diagnostic results can be quickly located by searching for similar region identifiers and associated node pairs in the medical image information database. Specifically, the feature f of the medical image to be tested is input. test Then, through the cosine similarity formula:

[0113]

[0114] The similarity S between the image and each region in the medical image database is calculated to match the closest historical data. The matching results can be associated with historical cases. For example, for an input cardiac image, if its vascular features are highly similar to a known case in the database, the diagnosis results of the case can be combined to provide auxiliary suggestions for doctors.

[0115] Those skilled in the art should understand that the technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A medical imaging diagnosis auxiliary diagnosis method, characterized in that: The following steps are involved: S1: Obtaining a deep learning network model for medical imaging: A medical imaging deep learning network model for obtaining recognition results for identifying different pathological areas of medical images; S2: Build a test model: Constructing an initial test model that matches the deep learning network model of the medical image, initializing the state of each network node of the initial test model and setting a state threshold of each network node, inputting a medical image with regional features of at least one known pathological area into the initial test model to associate and update the regional features of the pathological area in the medical image with the corresponding network node, until the state of each network node is greater than the set state threshold, then outputting the output value of each pathological area of ​​the current medical image, updating the initial test model based on the output value of the target deep learning model corresponding to the pathological area and the output value of the initial test model to obtain a test model; S3: One test: Initialize the state of each network node of the test model and set the state threshold of each network node, input the medical image to be tested into the test model to obtain the output value of each pathological area, assign and update the state of the corresponding network node based on the output value of each pathological area, generate a first memory result based on the learning state of the network node when the state of the network node is greater than 0, and output the output value of the pathological area whose state of the network node is greater than the set state threshold as a diagnosis and recognition result; S4: Secondary test: An initial test model that matches the deep learning network model of medical images is constructed, the status of each network node of the initial test model is initialized and the status threshold of each network node is set, the medical image to be tested after a test is input into the initial test model to traverse each network node of the initial test model, the pathological area of ​​the medical image to be tested is associated with each network node and the neurons of the network node are assigned with the information status of each network node, when the status of the network node in the associated path in the output value of the initial test model is less than 0, the corresponding network nodes are repeatedly assigned with the output value until the status of the network node corresponding to each pathological area of ​​the output value is greater than 1, based on the learning status of the network node combined with the first memory result of the corresponding network node, a second memory result is updated and generated, and the second memory result is assigned and optimized using an artificial intelligence algorithm to obtain a secondary diagnosis result.

2. The medical imaging diagnosis auxiliary diagnosis method according to claim 1, characterized in that: Additional steps include: Building a medical imaging information database: Using a medical imaging deep learning network model to identify each medical image to obtain identification results for each pathological area, generating a region identifier corresponding to the current pathological area based on the identification results, obtaining associated node pairs based on the region identifiers, and storing the region identifiers and associated node pairs in the form of structured data as a medical imaging information database.

3. The medical imaging diagnosis-aided diagnosis method according to claim 1, characterized in that: In step S2, the medical image with regional features of the known pathological area is represented as a two-dimensional or three-dimensional pixel matrix, wherein each pixel in the pixel matrix represents the image information of the pathological area, and the regional features of the corresponding pathological area are extracted by performing a convolution operation on the pixel points.

4. The medical imaging diagnosis auxiliary diagnosis method according to claim 1, characterized in that: In step S2, the characteristic values ​​of the regional features of the pathological area in the medical image are mapped to the corresponding network nodes as the input values ​​of the current network nodes, and the states of the network nodes are updated until the state of each network node is greater than 1. If the state S of the associated network node is j If it is less than 0, the following formula is used to update the status of the network node: Among them, η is the learning rate, which is used to control the update amplitude, and h i is the eigenvalue of the corresponding regional feature.

5. The medical imaging diagnosis-aided diagnosis method according to claim 1, characterized in that: In step S2, the target deep learning model based on the corresponding pathological area is a convolutional neural network trained for the pathological area. The same medical image is input into the target deep learning model to obtain an output value. The difference between the output value of the target deep learning model based on the corresponding pathological area and the output value of the initial test model is used to adjust the weights and bias items of the initial test model through the back propagation algorithm to obtain a test model.

6. The medical imaging diagnosis-aided diagnosis method according to claim 1, characterized in that: In step S3, the medical image to be tested is input into the test model to obtain the regional features of each pathological area, and the regional features are extracted as feature values, which are assigned to the states of the corresponding network nodes according to the association rules of the network nodes. The update formula is: where w jk is the weight of network node j and regional feature k in the test model, f k is the characteristic value of the regional feature, corresponding to the output value, is the initial state of the network node.

7. The medical imaging diagnosis-aided diagnosis method according to claim 1, characterized in that: In step S3, the state of each network node is determined. When the state of the network node is less than 0, the state of the network node is cumulatively updated using the output value of the corresponding pathological area. The formula is as follows: Where η is the learning rate, represents the state of the network node after the t+1th iteration, w jk is the weight of network node j and regional feature k in the test model, f k is the characteristic value of the regional feature; When the state of the network node is greater than 0, the learning state of the network node is used to generate the first memory result m j , the formula is as follows: Where α is the memory update coefficient, It is the historical memory value of the network node.

8. The medical imaging diagnosis-aided diagnosis method according to claim 1, characterized in that: In step S4, the medical image to be tested is segmented into multiple pathological regions, and each pathological region is feature extracted to obtain a feature value of the corresponding region feature, and the feature value is mapped to the network node of the initial test model through the feature allocation rule, wherein the feature allocation rule is: where w j k represents the weighted connection strength between network node j and regional feature k, is the initial state of network node j; The initial test model goes through all network nodes one by one. For each network node j, the state of the network node is adjusted according to its associated regional characteristics. The iterative update state is as follows: Where η is the learning rate, which controls the state update step of each iteration, and the function Represents the gradient of the state update.

9. The medical imaging diagnosis auxiliary diagnosis method according to claim 1, characterized in that: In step S4, determine whether the state of the network node in the associated path in the output value of the initial test model is less than 0. If so, continue to assign the current output value to the network node by accumulation until the state of the network node of each regional information in the above output value is greater than 1. If not, use the output value of each network node to obtain the second memory result of each network node. At this time, stop training and use the output value of the updated initial test model to make the first correction to the input end in the above initial test model.

10. The medical imaging diagnosis auxiliary diagnosis method according to claim 1, characterized in that: In step S4, the calculation formula of the second memory result is: Where β is the memory update coefficient, is the first memory result, s j Is the current state of the network node.

Images