Medical image-based physical examination report intelligent auxiliary interpretation method

Abstract

The application relates to the technical field of medical image processing, in particular to a medical image-based physical examination report intelligent auxiliary interpretation method. The method comprises the following steps: edge pixel points are obtained by detecting the edge of a subject organ region in a medical image; a plurality of connected domains with different pathological characteristics are determined according to the gray gradient of each edge pixel point; features are extracted from different scales for each connected domain to obtain a multi-scale feature vector of each connected domain; the matching degree of each connected domain and each target historical connected domain is determined according to the node similarity degree between the multi-scale feature vectors of each connected domain and each target historical connected domain and the performance characteristic nodes in a disease characteristic knowledge graph; the target text description mode recommended for each connected domain is determined according to the matching degree and the text description mode of the physical examination report of each target historical connected domain; and the interpretation content in the physical examination report of the medical image is generated according to the target text description mode of each connected domain. The application improves the interpretability of the physical examination report of the medical image.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method for intelligent auxiliary interpretation of physical examination reports based on medical images. Background Technology

[0002] With economic development and improved living standards, people are paying increasing attention to their health. Health checkups are an important way to understand the examinee's health status, detect early signs of disease, and identify potential health risks, making them essential for the aging population and those in a sub-healthy state. With the continuous development of medical imaging technology, imaging techniques are being used more and more widely in health checkup reports, especially CT, MRI, and X-rays, which can clearly show the condition of lesions within the human body.

[0003] However, medical images contain massive amounts of complex data, requiring specialized expertise for interpretation. Traditional manual interpretation of imaging reports is not only time-consuming but also susceptible to human error, leading to diagnostic inaccuracies. Therefore, how to use intelligent methods to assist doctors in analyzing medical images has become an important issue in medical image interpretation.

[0004] Currently, report generation models based on natural language processing technology mainly perform global analysis on the input image. However, since the prominence of small lesion areas in the overall image is insufficient, global analysis can greatly dilute the features of small lesion areas, easily leading to the erroneous classification of different types of lesions into the same type. This results in low precision and weak interpretability of the generated medical examination report. Summary of the Invention

[0005] To address the technical problems of low precision and weak interpretability in generated medical examination reports, the present invention aims to provide an intelligent auxiliary interpretation method for medical examination reports based on medical images. The specific technical solution adopted is as follows: This invention improves a method for intelligent auxiliary interpretation of medical examination reports based on medical images, the method comprising: Edge detection is performed on the examined organ region in medical images to obtain edge pixels; Based on the grayscale gradient of each edge pixel, multiple connected regions with different pathological features are determined; Features are extracted from each of the connected components at different scales to obtain multi-scale feature vectors for each of the connected components; The matching degree between each connected component and each target historical connected component is determined based on the node similarity between the multi-scale feature vectors of each connected component and each target historical connected component and the performance feature nodes in the disease feature knowledge graph. Based on the matching degree and the text description method of the physical examination report of each target historical connected domain, determine the recommended target text description method for each connected domain; Based on the target text description method of each of the connected domains, the interpretation content in the medical image examination report is generated.

[0006] According to the intelligent auxiliary interpretation method for medical examination reports based on medical images provided by the present invention, the step of determining multiple connected components with different pathological features based on the gray-level gradient of each edge pixel includes: The growth direction factor of each edge pixel on the same edge is determined based on the gray-level gradient of each edge pixel on the same edge. Based on the degree of difference in the growth direction factor of each edge pixel, the lesion feature parameters of each edge pixel are determined; Based on the similarity between the pathological feature parameters of each edge pixel, the edge pixels are merged to obtain multiple connected regions with different pathological features.

[0007] According to the intelligent auxiliary interpretation method for medical examination reports based on medical images provided by the present invention, the step of determining the lesion feature parameters of each edge pixel based on the degree of difference of the growth direction factor of each edge pixel includes: For each edge, the mean value of the growth direction factor of each edge pixel on the edge is calculated to obtain the mean value of the growth direction factor corresponding to the edge; For each edge pixel, the lesion feature parameters of the edge pixel are determined based on the difference between the growth direction factor of the edge pixel and the mean growth direction factor of the edge where the edge pixel is located.

[0008] According to the intelligent auxiliary interpretation method for medical examination reports based on medical images provided by the present invention, the step of extracting features from each of the connected components at different scales to obtain multi-scale feature vectors for each of the connected components includes: The image content within each of the connected components is input into a pre-trained multi-scale feature extraction model, which outputs the multi-scale feature vector of the connected component. The pre-trained multi-scale feature extraction model is specifically used for: The image content within the connected component is convolved to obtain a feature map; The feature map is divided into regions of different scales, and pooling is performed on each region to obtain the pooling results for each region; The pooling results of each region are combined to obtain the multi-scale feature vector of the connected domain.

[0009] According to the intelligent auxiliary interpretation method for medical examination reports based on medical images provided by the present invention, the step of determining the matching degree between each connected component and each target historical connected component based on the node similarity between the multi-scale feature vectors of each connected component and each target historical connected component and the performance feature nodes in the disease feature knowledge graph includes: For each of the connected components, a target historical connected component that shares common performance characteristics with the connected component is determined; For each of the common performance features, the difference in node similarity between the common performance features of the connected component and the target historical connected component is determined to obtain the node similarity difference; Each of the common performance features is taken as a target performance feature, and the representation ability of the target performance feature for the patient is determined based on the difference in the similarity between the nodes corresponding to each connected component that possesses the target performance feature. The matching degree between the connected component and the target historical connected component is determined based on the number of common performance features of the connected component and the target historical connected component, the difference in node similarity corresponding to each common performance feature, and the representational ability of each common performance feature to the patient.

[0010] According to the intelligent auxiliary interpretation method for medical examination reports based on medical images provided by the present invention, the step of determining the representational power of the target performance feature for the patient based on the difference in the similarity between the nodes corresponding to each connected component possessing the target performance feature includes: Determine the difference in similarity between nodes corresponding to any two connected components that possess the target performance characteristics; The average of the differences is used to determine the representativeness of the target performance feature for the patient.

[0011] The intelligent assisted interpretation method for medical examination reports based on medical images provided by the present invention further includes: A first historical connected component with a matching degree greater than or equal to a preset matching degree threshold and a second historical connected component with a matching degree greater than or equal to the preset matching degree threshold are determined. Based on the matching degree between each of the first historical connected components and each of the second historical connected components, the matching degree between the connected component and the target historical connected component is corrected to obtain the corrected matching degree; The step of determining the recommended target text description method for each connected domain based on the matching degree and the text description method of the physical examination report of each target historical connected domain includes: Based on the corrected matching degree and the text description method of the health check report of each target historical connected domain, the recommended target text description method for each connected domain is determined.

[0012] According to the intelligent assisted interpretation method for medical examination reports based on medical images provided by the present invention, the step of correcting the matching degree between the connected component and the target historical connected component based on the matching degree between each first historical connected component and each second historical connected component to obtain the corrected matching degree includes: The matching degree of each first historical connected component and each second historical connected component is summed to obtain the total matching degree; The corrected matching degree is obtained by multiplying the matching degree between the connected component and the target historical connected component by the sum of the matching degrees.

[0013] According to the intelligent assisted interpretation method for medical examination reports based on medical images provided by the present invention, the step of determining the recommended target text description method for each of the target historical connected domains based on the matching degree and the text description method of each target historical connected domain includes: For each of the connected components, determine the description level of the physical examination report for each of the target historical connected components of the connected component; For each text description method, the recommendation degree of the text description method for the connected component is determined based on the matching degree between the connected component and each of the target historical connected components using the text description method, and the description level of each of the target historical connected components using the text description method. Based on the recommendation degree of each of the aforementioned text description methods, the target text description method recommended for the connected component is determined.

[0014] According to the intelligent auxiliary interpretation method for physical examination reports based on medical images provided by the present invention, the step of determining the target text description method recommended for the connected component based on the recommendation degree of each of the aforementioned text description methods includes: According to the order of recommendation degree of each of the text description methods from large to small, a preset number of text description methods are selected as candidate text description methods; For each candidate text description method, the probability of the candidate text description method being replaced is determined based on the difference between the characters in the connected component using the candidate text description method and the characters in each similar historical connected component using the candidate text description method, as well as the recommendation degree of the candidate text description method. Based on the replacement probability of each of the candidate text description methods, a target text description method recommended for the connected component is determined.

[0015] The present invention has the following beneficial effects: by identifying multiple connected regions with different pathological features from the examined organ region of medical images, and extracting features from each connected region at different scales to obtain multi-scale feature vectors for each connected region, it is possible to capture the performance features of lesion regions at multiple scales, avoiding the problem that global analysis would greatly dilute the features of small lesion regions. Then, by combining the disease feature knowledge graph, it is possible to accurately determine the matching degree between each connected region and each target historical connected region. Based on the matching degree and the text description method of the physical examination report of each target historical connected region, it is possible to accurately determine the target text description method recommended for each connected region, thereby improving the accuracy and interpretability of the interpretation content in the physical examination report of medical images. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an intelligent assisted interpretation method for medical examination reports based on medical images, provided as an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the effect of preprocessed medical images provided in one embodiment of the present invention. Figure 3 This is a flowchart illustrating the process of determining multiple connected regions with different pathological features according to an embodiment of the present invention. Figure 4 This is a schematic diagram of a process for determining the matching degree between connected components according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating the process of determining a target text description method according to an embodiment of the present invention. Detailed Implementation

[0018] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent auxiliary interpretation method for medical examination reports based on medical images provided by this invention.

[0019] Please see Figure 1 The diagram illustrates a flowchart of an intelligent assisted interpretation method for medical examination reports based on medical images, according to an embodiment of the present invention, comprising the following steps: Step 102: Perform edge detection on the examined organ region in the medical image to obtain edge pixels.

[0020] In one embodiment, medical images can be images obtained by acquiring images of a living organism using techniques such as CT (Computed Tomography), MRI (Nuclear Magnetic Resonance Imaging), or X-rays (X-rays).

[0021] In one embodiment, the area of ​​the organ being examined may be the area corresponding to organs such as the heart, spleen, liver, lungs, or kidneys.

[0022] In one embodiment, medical image data for several patients can be retrieved from the hospital's imaging system database. This medical image data can be stored in DICOM (Digital Imaging and Communications in Medicine) format, containing both medical images and patient metadata. The patient metadata may include the patient's identifier, examination time, and physician diagnostic information.

[0023] In one embodiment, since medical images do not all originate from the same type of imaging device, there will be significant differences in image format and size. Furthermore, besides the examined organ region, the images may contain other irrelevant device information. Therefore, image cropping and scaling methods can be used to standardize the acquired raw medical images into a uniform digital image format. Noise reduction and enhancement processing is then performed on the uniform format medical images to obtain preprocessed medical images. Step 104 and subsequent steps are then performed based on the preprocessed medical images. The uniform format can be a 512*512 pixel JPG image. The effect of the preprocessed medical image is as follows: Figure 2 As shown.

[0024] It is understandable that medical images may contain non-examined tissue regions besides the examined organ region. For example, if the examined organ region is the lung region, other unrelated tissue regions such as bones or the heart may also exist in the medical image, interfering with the accuracy of image recognition. Since different tissues have different absorption characteristics for detection rays, there are differences in grayscale values ​​between different tissues in medical images. Therefore, the examined organ region can be determined from the medical image based on grayscale values, eliminating the influence of unrelated regions. Thus, in one embodiment, a region growing algorithm can be used to identify and separate the examined organ region in the medical image, and then edge detection can be performed on the examined organ region to obtain edge pixels.

[0025] In one embodiment, an edge detection operator can be used to process the pixels in the examined organ region to obtain the edge pixels and their grayscale gradients. For example, the edge detection operator could be the Sobel operator.

[0026] Step 104: Determine multiple connected regions with different pathological features based on the gray-level gradient of each edge pixel.

[0027] In one embodiment, the lesion feature parameters of each edge pixel can be determined based on the gray-level gradient of each edge pixel, and then the edge pixels can be merged to obtain multiple connected regions with different pathological features based on the similarity between the lesion feature parameters of each edge pixel.

[0028] Step 106: Extract features from each connected component at different scales to obtain multi-scale feature vectors for each connected component.

[0029] Among them, the multi-scale feature vector is the feature vector obtained by combining feature vectors extracted from different scales.

[0030] In one embodiment, the image content within each connected component is input into a pre-trained multi-scale feature extraction model, which outputs a multi-scale feature vector of the connected component.

[0031] In one embodiment, the multi-scale feature extraction model can be a model based on the SPP (Spatial Pyramid Pooling) algorithm.

[0032] Step 108: Determine the matching degree between each connected component and each target historical connected component based on the node similarity between the multi-scale feature vectors of each connected component and each target historical connected component and the performance feature nodes in the disease feature knowledge graph.

[0033] In this context, the target historical connected domain refers to a historical connected domain that shares common performance characteristics with the connected domains in the medical images to be interpreted. Historical connected domains are connected domains in the medical images of patients with existing medical examination reports. Each node in the disease feature knowledge graph represents a different performance characteristic. Performance characteristics can include disease type and symptoms. For example, the disease type could be pneumonia or emphysema, and the symptoms could be shortness of breath.

[0034] In one embodiment, multiple corresponding performance feature nodes can be determined from the disease feature knowledge graph based on the multi-scale feature vector. The node similarity between the multi-scale feature vector and the performance feature nodes in the disease feature knowledge graph can be calculated. Based on the node similarity between each connected component and each target historical connected component and the corresponding performance feature nodes in the disease feature knowledge graph, the matching degree between each connected component and each target historical connected component can be determined.

[0035] Step 110: Based on the matching degree and the text description method of the health check report of each target historical connected domain, determine the recommended target text description method for each connected domain.

[0036] In one embodiment, the recommendation degree of various text description methods is determined based on the matching degree of the connected components and the text description methods of the health check reports of each target historical connected component. Then, based on the recommendation degree of various text description methods, the target text description method recommended for the connected component is determined.

[0037] In one embodiment, the text description method with the highest recommendation among various text description methods can be directly determined as the target text description method. In another embodiment, a preset number of text description methods can be selected as candidate text description methods in descending order of their recommendation levels. Then, based on the character situation when using various candidate text description methods in connected components, the replacement probability of each candidate text description method is determined, and the candidate text description method with the lowest replacement probability is selected as the target text description method.

[0038] Step 112: Generate the interpretation content of the medical image examination report based on the target text description method of each connected domain.

[0039] In one embodiment, the target text descriptions corresponding to each connected region in a medical image are integrated to generate interpretation content for a medical examination report, thereby generating the medical examination report document. The interpretation content may include descriptions of the morphology, location, size, and density of normal and diseased organs. For example: A nodule of about 1.5 cm in the upper lobe of the right lung, with smooth edges and uniform density; a ground-glass opacity of about 2 cm in the lower lobe of the left lung, with uneven density.

[0040] The aforementioned intelligent auxiliary interpretation method for medical examination reports based on medical images identifies multiple connected regions with different pathological features from the examined organ regions in medical images. Features are extracted from each connected region at different scales to obtain multi-scale feature vectors for each region. This allows for the capture of the manifestation features of lesion areas at multiple scales, avoiding the problem of global analysis significantly diluting the features of small lesion areas. Then, combined with a disease feature knowledge graph, the matching degree between each connected region and each target historical connected region can be accurately determined. Based on the matching degree and the text description methods of the medical examination reports for each target historical connected region, the recommended target text description method can be accurately determined for each connected region. This improves the accuracy and interpretability of the interpretation content in the medical examination report, ensuring high consistency in the interpretation effect of medical examination reports from different patients and improving the standardization of medical examination reports. Furthermore, the fully automated auxiliary interpretation of medical examination reports based on medical images avoids the inefficiency caused by manual interpretation, improving the efficiency of medical examination report generation.

[0041] In one embodiment, see Figure 3 Step 104, based on the grayscale gradient of each edge pixel, determines multiple connected components with different pathological features, including the following steps: Step 302: Determine the growth direction factor of each edge pixel on the same edge based on the gray-level gradient of each edge pixel on the same edge.

[0042] In one embodiment, the grayscale gradients of each edge pixel on the same edge can be normalized based on their grayscale gradients to obtain the growth direction factor of each edge pixel on the same edge. The formula is as follows: in, Indicates the first on the same edge The growth direction factor of each edge pixel. Indicates the first on the same edge The grayscale gradient of each edge pixel. This represents the minimum grayscale gradient of each edge pixel on the same edge. This represents the maximum value of the grayscale gradient of each edge pixel on the same edge.

[0043] Step 304: Determine the lesion feature parameters of each edge pixel based on the degree of difference in the growth direction factor of each edge pixel.

[0044] Step 306: Based on the similarity between the lesion feature parameters of each edge pixel, merge each edge pixel to obtain multiple connected regions with different pathological features.

[0045] In one embodiment, for each edge pixel, the similarity between the edge pixel and the pathological feature parameters of each edge pixel in its neighborhood is calculated. Edge pixels with similarity greater than or equal to a preset similarity threshold are merged to obtain multiple connected domains with different pathological features. The neighborhood can be an 8-neighborhood.

[0046] In the above embodiments, the growth direction factor of each edge pixel on the same edge is determined based on the gray-level gradient of each edge pixel on the same edge. Since the growth direction of edge pixels in the same pathological region is similar, the lesion feature parameters of each edge pixel can be accurately measured based on the degree of difference in the growth direction factors of each edge pixel. Since the lesion feature parameters of pixels in the same pathological region show feature similarity, multiple connected regions with different pathological features can be accurately determined based on the similarity between the lesion feature parameters of each edge pixel, thereby achieving accurate division of connected regions.

[0047] In one embodiment, determining the lesion feature parameters of each edge pixel based on the degree of difference in the growth direction factors of each edge pixel includes: calculating the mean value of the growth direction factors of each edge pixel on each edge to obtain the mean value of the growth direction factors corresponding to the edge; and determining the lesion feature parameters of each edge pixel based on the difference between the growth direction factor of the edge pixel and the mean value of the growth direction factors of the edge on which the edge pixel is located.

[0048] In one embodiment, the lesion feature parameters of edge pixels can be determined according to the following formula: in, Indicates the first edge The lesion feature parameters of each edge pixel. Indicates the first edge The growth direction factor of each edge pixel. This represents the mean of the growth direction factor at the edge.

[0049] In the above embodiments, since the growth direction of edge pixels in the same pathological region is similar, the pathological feature parameters of each edge pixel can be accurately determined based on the difference between the growth direction factor of the edge pixel and the mean growth direction factor of the edge where the edge pixel is located.

[0050] In one embodiment, features are extracted from each connected component at different scales to obtain multi-scale feature vectors for each connected component. This includes: inputting the image content within each connected component into a pre-trained multi-scale feature extraction model and outputting multi-scale feature vectors for the connected components. The pre-trained multi-scale feature extraction model is specifically used to: perform convolution processing on the image content within the connected components to obtain feature maps; divide the feature maps into regions at different scales and perform pooling processing on each region to obtain pooling results for each region; and combine the pooling results for each region to obtain multi-scale feature vectors for the connected components.

[0051] The multi-scale feature extraction model includes convolutional layers, SPP (Spatial Pyramid Pooling) layers, and fully connected layers. Convolutional layers process the image content within the connected region to obtain a feature map. Then, the SPP layer divides the feature map into regions of different scales and performs pooling on each region, yielding pooling results for each region. The fully connected layer combines these pooling results to obtain a multi-scale feature vector for the connected region. For example, different scale region divisions could be performed at scales of 4×4, 2×2, and 1×1. Finally, combining the pooling results from each region yields the multi-scale feature vector, enabling the extraction of feature vectors of the same length from input images of different sizes.

[0052] In the above embodiments, the feature map is divided into regions of different scales by a multi-scale feature extraction model, and pooling is performed on each region to obtain the pooling results of each region. The pooling results of each region are combined to obtain the multi-scale feature vector of the connected component. This can extract feature vectors of the same length from input images of different sizes, and can capture the performance features of lesion areas from multiple scales. This avoids the problem that global analysis will greatly dilute the features of small lesion areas, and improves the accuracy of interpretation.

[0053] In one embodiment, see Figure 4 Step 108, based on the node similarity between the multi-scale feature vectors of each connected component and each target historical connected component and the performance feature nodes in the disease feature knowledge graph, determines the matching degree between each connected component and each target historical connected component, including the following steps: Step 402: For each connected component, determine the target historical connected components that share common performance characteristics with the connected components.

[0054] Step 404: For each common performance feature, determine the difference in node similarity between the common performance features of the connected component and the target historical connected component, and obtain the node similarity difference.

[0055] In one embodiment, the difference between the node similarity between the connected component and the target historical connected component under each common performance feature can be calculated to obtain the node similarity difference corresponding to each common performance feature.

[0056] Step 406: Each common performance feature is taken as the target performance feature. Based on the difference in the similarity between the nodes corresponding to each connected component that has the target performance feature, the representation ability of the target performance feature for the patient is determined.

[0057] Step 408: Determine the matching degree between the connected domain and the target historical connected domain based on the number of common performance features of the connected domain and the target historical connected domain, the difference in node similarity corresponding to each common performance feature, and the representation ability of each common performance feature for the patient.

[0058] In one embodiment, the matching degree between the connected component and the target historical connected component is positively correlated with the number of common representational features. The matching degree between the connected component and the target historical connected component is positively correlated with the representativeness of each common representational feature for the patient. The matching degree between the connected component and the target historical connected component is negatively correlated with the difference in node similarity corresponding to each common representational feature.

[0059] In one embodiment, for each common performance feature, the ratio between the representation ability of the common performance feature to the patient and the difference in the similarity between the nodes corresponding to the common performance feature can be calculated. Then, the ratios corresponding to each common performance feature are summed. Based on the product of the summation result and the number of common performance features, the matching degree between the connected component and the target historical connected component is determined.

[0060] In one embodiment, the matching degree between the connected component and the target historical connected component can be determined according to the following formula: in, Indicates the first The connected component and the first The degree of matching between the target historical connected components. Indicates the first The connected component and the first The number of common performance features of the target historical connected domains. Indicates the first A common feature reflects a patient's ability to express themselves. Indicates the first The connected component in the th order The degree of similarity between nodes that share common performance characteristics. Indicates the first The target historical connected components are in the first The degree of similarity between nodes that share common performance characteristics. Indicates the first The connected component and the first The first target historical connected domain The difference in the degree of similarity between nodes corresponding to common performance features. This represents a minimum hyperparameter, used to avoid the denominator equal to 0. In this embodiment... The value is 0.01. In specific applications, implementers can set it according to the specific circumstances.

[0061] In the above embodiments, based on the common performance features of the two connected components, the greater the number of common performance features and the more similar the nodes of each common performance feature are (i.e., the smaller the differences), the higher the matching degree between the two connected components. Furthermore, when calculating the similarity, the representational ability of each performance feature for the patient being compared needs to be considered; the stronger the representational ability, the higher the matching degree between the two connected components. Therefore, based on the number of common performance features between the connected component and the target historical connected component, the differences in node similarity corresponding to each common performance feature, and the representational ability of each common performance feature for the patient, the matching degree between the connected component and the target historical connected component can be accurately determined.

[0062] In one embodiment, determining the representational power of a target performance feature for a patient based on the difference in the similarity between nodes corresponding to each connected component possessing the target performance feature includes: determining the difference in the similarity between nodes corresponding to any two connected components possessing the target performance feature; and averaging the differences to determine the representational power of the target performance feature for the patient.

[0063] In one embodiment, the representativeness of a target performance characteristic for a patient can be determined according to the following formula: in, in, Indicates the first A common feature of a patient's expressive ability. Indicates ownership of the first The first common monetization characteristic The similarity of nodes corresponding to each connected component. Indicates ownership of the first The first common monetization characteristic The similarity of nodes corresponding to each connected component. It represents the difference in the similarity between nodes corresponding to any two connected components. Indicates ownership of the first The number of connected components that share a common representation feature.

[0064] In the above embodiments, the difference between the similarity between nodes corresponding to any two connected components that have the target performance feature is determined, and the average value of each difference is calculated to determine the representation ability of the target performance feature for the patient. This can accurately measure the representation ability of each common performance feature for the patient.

[0065] In one embodiment, the method further includes: determining a first historical connected domain with a matching degree greater than or equal to a preset matching degree threshold, and a second historical connected domain with a matching degree greater than or equal to a preset matching degree threshold; adjusting the matching degree between the connected domain and the target historical connected domain based on the matching degree between each first historical connected domain and each second historical connected domain to obtain a adjusted matching degree. Determining the recommended target text description method for each connected domain based on the matching degree and the text description method of the medical examination reports for each target historical connected domain includes: determining the recommended target text description method for each connected domain based on the adjusted matching degree and the text description method of the medical examination reports for each target historical connected domain.

[0066] The preset matching degree threshold can be set according to the actual situation. For example, the preset matching degree threshold can be set to 0.75. That is, historical connected components with a matching degree greater than or equal to 0.75 with the connected component are determined as the first historical connected component (type A connected component), and historical connected components with a matching degree greater than or equal to 0.75 with the target connected component are determined as the second historical connected component (type B connected component).

[0067] In one embodiment, a correction weight can be calculated based on the matching degree between each first historical connected component and each second historical connected component. The matching degree between the connected component and the target historical connected component can then be corrected based on the correction weight to obtain the corrected matching degree.

[0068] In the above embodiments, a first historical connected domain with a matching degree greater than or equal to a preset matching degree threshold and a second historical connected domain with a matching degree greater than or equal to a preset matching degree threshold are determined respectively. Based on the matching degree between each first historical connected domain and each second historical connected domain, the matching degree between the connected domain and the target historical connected domain is corrected, which can obtain a more accurate matching degree.

[0069] In one embodiment, the matching degree between a connected component and a target historical connected component is corrected based on the matching degree between each first historical connected component and each second historical connected component to obtain a corrected matching degree. This includes: summing the matching degrees between each first historical connected component and each second historical connected component to obtain a total matching degree; and multiplying the matching degree between the connected component and the target historical connected component by the total matching degree to obtain the corrected matching degree.

[0070] In one embodiment, the corrected matching degree can be determined according to the following formula: in, Indicates the first The connected component and the first The corrected matching degree between the target historical connected components. Indicates the first The connected component and the first The degree of matching between the target historical connected components before correction. This indicates normalization processing. The th term represents the matching degree between each first historical connected component and each second historical connected component. 1 match degree. This represents the total matching degree. This represents the number of matching degrees between each first historical connected component and each second historical connected component.

[0071] In the above embodiments, the matching degree of each first historical connected domain and each second historical connected domain is summed to obtain the total matching degree. The matching degree between the connected domain and the target historical connected domain is multiplied by the total matching degree to obtain the corrected matching degree, which can obtain a more accurate matching degree.

[0072] In one embodiment, see Figure 5 Step 110, based on the matching degree and the text description methods of the health check reports of each target's historical connected components, determines the recommended target text description methods for each connected component, including the following steps: Step 502: For each connected component, determine the description level of the health check report of each target historical connected component.

[0073] The descriptive grade is used to characterize the severity of the lesion area as assessed in the physical examination report. For example, pneumonia is classified into mild pneumonia, moderate pneumonia, moderate-to-severe pneumonia, and severe pneumonia based on its severity level.

[0074] Step 504: For each text description method, determine the recommendation degree of the text description method for the connected component based on the matching degree between the connected component and each target historical connected component using the text description method, and the description level of each target historical connected component using the text description method.

[0075] In one embodiment, the recommendation degree of the text description method for a connected component is positively correlated with the matching degree between the connected component and each target historical connected component using the text description method. The recommendation degree of the text description method for a connected component is positively correlated with the description level of each target historical connected component using the text description method.

[0076] In one embodiment, for each text description method, the matching degree between the connected component and each target historical connected component using that text description method is multiplied by the description level of each target historical connected component using that text description method to obtain the multiplication result. Then, the average of the multiplication results corresponding to each target historical connected component using that text description method is calculated to obtain the recommendation degree of that text description method for the connected component.

[0077] In one embodiment, the recommendation degree of the text description method for connected components can be determined according to the following formula: in, Indicates the first The recommendation level of each text description method for connected components. Representing connected components and using the first The first of the text description methods The corrected matching degree between the target historical connected components. Indicates the use of the first The first of the text description methods The description level of the health check report for each target historical connected domain. The target history of each connected component is represented by the first connected component. The number of target historical connected domains in each text description method.

[0078] Step 506: Determine the target text description method to be recommended for the connected component based on the recommendation degree of each text description method.

[0079] In the above embodiments, for each connected component, the description level of the medical examination report of each target historical connected component of the connected component is determined. For each text description method, based on the matching degree between the connected component and each target historical connected component using the text description method, and the description level of each target historical connected component using the text description method, the recommendation degree of the text description method for the connected component can be accurately determined. Based on the recommendation degree of each text description method, the target text description method recommended for the connected component can be accurately determined, thereby accurately obtaining the target text description method corresponding to each connected component in the medical image.

[0080] In one embodiment, determining the target text description method for the connected component based on the recommendation degree of each text description method includes: selecting a preset number of text description methods as candidate text description methods in descending order of their recommendation degree; for each candidate text description method, determining the likelihood of the candidate text description method being replaced based on the difference between the characters in the connected component using the candidate text description method and the characters in each similar historical connected component using the candidate text description method, as well as the recommendation degree of the candidate text description method; and determining the target text description method for the connected component based on the likelihood of the various candidate text description methods being replaced.

[0081] The preset number can be set according to actual needs. For example, the preset number can be set to 3. That is, the top 3 text description methods are selected as candidate text description methods in descending order of recommendation.

[0082] In one embodiment, the probability of a candidate text description being replaced is positively correlated with the difference between the characters in a connected component using the candidate text description and the characters in similar historical connected components using the candidate text description. The probability of a candidate text description being replaced is negatively correlated with the recommendation level of the candidate text description.

[0083] In one embodiment, the likelihood of a candidate text description being replaced can be determined by the ratio of the sum of the differences between the characters in a connected component using the candidate text description and the characters in each similar historical connected component using the candidate text description, to the recommendation score of the candidate text description. The formula is as follows: in, Indicates the first The possibility of replacing a candidate text description method. Indicates the first The recommendation degree of each candidate text description method for connected components. To indicate connected components, use the first... Characters representing candidate text description methods. The first term of the connected component The first similar historical connected component uses the first Characters representing candidate text description methods. To indicate connected components, use the first... The characters of the candidate text description method and the first The first similar historical connected component uses the first Differences between characters in the candidate text description methods. This represents the number of similar historical connected components within a connected component. To indicate connected components, use the first... The characters of the candidate text description method and each similar historical connected component are used with the first... The sum of the differences between characters in the candidate text description methods.

[0084] In one embodiment, the candidate text description method with the lowest probability of being replaced among various candidate text description methods can be determined as the target text description method recommended for the connected component.

[0085] In the above embodiments, since the rationality among multiple connected component descriptions should also be considered when recommending text description methods, the probability of a candidate text description method being replaced is determined by calculating the likelihood of text description duplication among connected components when changing the description method. The higher the recommendation level of a certain candidate text description method for a connected component, and the lower the difference in text description duplication across the entire medical examination report, the lower the probability of that text description method being replaced. Therefore, based on the differences between the characters in a connected component using the candidate text description method and the characters in similar historical connected components using the candidate text description method, as well as the recommendation level of the candidate text description method, the probability of a candidate text description method being replaced can be accurately determined. Finally, based on the probability of various candidate text description methods being replaced, the target text description method recommended for the connected component can be accurately determined.

Claims

1. A method for intelligent auxiliary interpretation of physical examination reports based on medical images, characterized in that, The method includes: Edge detection is performed on the examined organ region in medical images to obtain edge pixels; Based on the grayscale gradient of each edge pixel, multiple connected regions with different pathological features are determined; Features are extracted from each of the connected components at different scales to obtain multi-scale feature vectors for each of the connected components; The matching degree between each connected component and each target historical connected component is determined based on the node similarity between the multi-scale feature vectors of each connected component and each target historical connected component and the performance feature nodes in the disease feature knowledge graph. Based on the matching degree and the text description method of the physical examination report of each target historical connected domain, determine the recommended target text description method for each connected domain; Based on the target text description method of each of the connected domains, the interpretation content in the medical image examination report is generated.

2. The intelligent auxiliary interpretation method for physical examination reports based on medical images according to claim 1, characterized in that, The step of determining multiple connected regions with different pathological features based on the gray-level gradient of each edge pixel includes: The growth direction factor of each edge pixel on the same edge is determined based on the gray-level gradient of each edge pixel on the same edge. Based on the degree of difference in the growth direction factor of each edge pixel, the lesion feature parameters of each edge pixel are determined; Based on the similarity between the pathological feature parameters of each edge pixel, the edge pixels are merged to obtain multiple connected regions with different pathological features.

3. The intelligent auxiliary interpretation method for physical examination reports based on medical images according to claim 2, characterized in that, The step of determining the lesion feature parameters of each edge pixel based on the degree of difference in the growth direction factor of each edge pixel includes: For each edge, the mean value of the growth direction factor of each edge pixel on the edge is calculated to obtain the mean value of the growth direction factor corresponding to the edge; For each edge pixel, the lesion feature parameters of the edge pixel are determined based on the difference between the growth direction factor of the edge pixel and the mean growth direction factor of the edge where the edge pixel is located.

4. The intelligent assisted interpretation method for physical examination reports based on medical images according to claim 1, characterized in that, The step of extracting features from each of the connected components at different scales to obtain multi-scale feature vectors for each of the connected components includes: The image content within each of the connected components is input into a pre-trained multi-scale feature extraction model, which outputs the multi-scale feature vector of the connected component. The pre-trained multi-scale feature extraction model is specifically used for: The image content within the connected component is convolved to obtain a feature map; The feature map is divided into regions of different scales, and pooling is performed on each region to obtain the pooling results for each region; The pooling results of each region are combined to obtain the multi-scale feature vector of the connected domain.

5. The intelligent auxiliary interpretation method for physical examination reports based on medical images according to claim 1, characterized in that, The step of determining the matching degree between each connected component and each target historical connected component based on the node similarity between the multi-scale feature vectors of each connected component and each target historical connected component and the performance feature nodes in the disease feature knowledge graph includes: For each of the connected components, a target historical connected component that shares common performance characteristics with the connected component is determined; For each of the common performance features, the difference in node similarity between the common performance features of the connected component and the target historical connected component is determined to obtain the node similarity difference; Each of the common performance features is taken as a target performance feature, and the representation ability of the target performance feature for the patient is determined based on the difference in the similarity between the nodes corresponding to each connected component that possesses the target performance feature. The matching degree between the connected component and the target historical connected component is determined based on the number of common performance features of the connected component and the target historical connected component, the difference in node similarity corresponding to each common performance feature, and the representational ability of each common performance feature to the patient.

6. The intelligent auxiliary interpretation method for physical examination reports based on medical images according to claim 5, characterized in that, The step of determining the representational power of the target performance feature for the patient based on the difference in the similarity between the nodes corresponding to each connected component possessing the target performance feature includes: Determine the difference in similarity between nodes corresponding to any two connected components that possess the target performance characteristics; The average of the differences is used to determine the representativeness of the target performance feature for the patient.

7. The intelligent auxiliary interpretation method for physical examination reports based on medical images according to claim 5, characterized in that, The method further includes: A first historical connected component with a matching degree greater than or equal to a preset matching degree threshold and a second historical connected component with a matching degree greater than or equal to the preset matching degree threshold are determined. Based on the matching degree between each of the first historical connected components and each of the second historical connected components, the matching degree between the connected component and the target historical connected component is corrected to obtain the corrected matching degree; The step of determining the recommended target text description method for each connected domain based on the matching degree and the text description method of the physical examination report of each target historical connected domain includes: Based on the corrected matching degree and the text description method of the health check report of each target historical connected domain, the recommended target text description method for each connected domain is determined.

8. The intelligent auxiliary interpretation method for physical examination reports based on medical images according to claim 7, characterized in that, The step of correcting the matching degree between the connected component and the target historical connected component based on the matching degree between each of the first historical connected components and each of the second historical connected components to obtain the corrected matching degree includes: The matching degree of each first historical connected component and each second historical connected component is summed to obtain the total matching degree; The corrected matching degree is obtained by multiplying the matching degree between the connected component and the target historical connected component by the sum of the matching degrees.

9. The intelligent assisted interpretation method for medical examination reports based on medical images according to any one of claims 1 to 8, characterized in that, The step of determining the recommended target text description method for each connected component based on the matching degree and the text description method of the physical examination report of each target historical connected component includes: For each of the connected components, determine the description level of the physical examination report for each of the target historical connected components of the connected component; For each text description method, the recommendation degree of the text description method for the connected component is determined based on the matching degree between the connected component and each of the target historical connected components using the text description method, and the description level of each of the target historical connected components using the text description method. Based on the recommendation degree of each of the aforementioned text description methods, the target text description method recommended for the connected component is determined.

10. The intelligent auxiliary interpretation method for physical examination reports based on medical images according to claim 9, characterized in that, The step of determining the target text description method recommended for the connected component based on the recommendation degree of each of the text description methods includes: According to the order of recommendation degree of each of the text description methods from large to small, a preset number of text description methods are selected as candidate text description methods; For each candidate text description method, the probability of the candidate text description method being replaced is determined based on the difference between the characters in the connected component using the candidate text description method and the characters in each similar historical connected component using the candidate text description method, as well as the recommendation degree of the candidate text description method. Based on the replacement probability of each of the candidate text description methods, a target text description method recommended for the connected component is determined.

Images