Medical image report automatic quality control error correction system and method

Abstract

The invention relates to the technical field of medical image report error correction, in particular to a medical image report automatic quality control error correction system and method. The method comprises the steps of importing medical image data, analyzing the medical image data by using an open source tool, performing standardization processing and local contrast enhancement on an image, performing automatic segmentation of a target area through a UNet model, and calculating density features of a key area to obtain feature vectors; meanwhile, the report text data is subjected to word segmentation, standardization processing and keyword extraction steps, the word importance is evaluated by utilizing a BERT model, a semantic vector of the report is generated through a text model, finally, the consistency of the image and the report is detected by calculating the similarity between a feature vector and the semantic vector, and an automatic error correction program is started when the image and the report are inconsistent. And error types are analyzed and corrected. According to the method, medical image processing, natural language processing and deep learning technologies are combined, so that human errors are reduced, and the diagnosis efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image report error correction, and in particular to a medical image report automated quality control error correction system and method. Background Art

[0002] With the rapid development of medical imaging technology, equipment such as CT, MRI and X-ray have become important tools for disease diagnosis. However, in the process of medical image analysis, the quality of the generated diagnostic report directly affects clinical decision-making. Current reports mainly rely on the doctor's experience and manual input, and there are problems such as spelling errors, irregular grammar, inconsistent terminology, and mismatch between images and report content. These problems are particularly prominent under high workload and time pressure, which not only increases the risk of medical errors, but also affects the efficiency of medical services and the treatment effects of patients.

[0003] A Chinese invention patent application with publication number CN114927199A discloses a quality control management system and method for medical imaging reports and images, including: an administrator workstation for sampling and assigning quality control tasks, specifically: receiving the operation of the quality control administrator to create a quality control sampling task; obtaining the reports and images to be quality controlled from the database according to the sampling conditions in the quality control sampling task; assigning the reports and images to be quality controlled to quality control experts for quality control according to the allocation rules; the application uses cloud computing technology to build a cloud quality control system including an administrator workstation and a quality control expert workstation, which can realize the management and control of the quality of medical images and reports, and the above-mentioned quality control system improves the quality control efficiency and ensures the rationality and auditability of the quality control process, thereby providing perfect tools and means for the refined quality management of medical imaging authorities (including imaging departments, hospitals, industry associations and health administrative departments).

[0004] However, in actual production life, the manual review process is time-consuming and laborious, and it is difficult to detect potential errors in the report in a timely manner, further exacerbating the risk of medical errors. This not only reduces the accuracy of the diagnosis report, but also affects the efficiency of medical decision-making and the treatment effect of patients. The development of artificial intelligence technology provides technical support for the automated quality control and error correction of medical imaging reports. The breakthrough of deep learning in image feature extraction and natural language processing makes it possible to automatically extract features from imaging data and match diagnostic reports. By building an automated quality control and error correction system that combines intelligent analysis, error detection and feedback optimization, it can not only significantly improve the report quality and diagnostic efficiency, but also provide data support for the standardized management of medical institutions. Summary of the invention

[0005] The purpose of the present invention is to address the problems existing in the background technology and to propose a medical image report automated quality control and error correction system and method.

[0006] The technical solution of the present invention is a method for automatic quality control and error correction of medical image reports, comprising the following specific implementation steps:

[0007] S1. Import medical image data, analyze the medical image data, standardize the image, use contrast-limited adaptive histogram equalization method to enhance the local contrast of the image, then extract the target area, automatically perform pixel-level segmentation based on the context information of the target area, extract the key area, and obtain the feature vector based on it, and then construct a coupled information group <patient identification number, patient detection time, feature vector>;

[0008] S2, import the report text data corresponding to the medical imaging data, split the report text data into word sequences through word segmentation technology, generate a standardized word set, extract keywords, convert each keyword into an embedded vector representation, output a keyword vector set, and use the text model to generate a semantic vector, and construct a coupled information group <patient identification number, patient detection time, semantic vector> based on this;

[0009] S3, extracting the feature vector of the medical image data and the semantic vector of the corresponding report text data, calculating the similarity between the feature vector and the semantic vector, and comparing the similarity with the set similarity threshold: if the similarity exceeds the set similarity threshold, it is considered that the content of the report matches the content of the image; otherwise, it is considered that there is an inconsistency between the content of the image and the content of the report, identifying the difference, and executing S4;

[0010] S4, start the automatic error correction program, input the feature vector, the semantic vector and the difference part into the pre-trained automatic error correction generation model, the automatic error correction generation model repairs the input semantic vector, obtains the corrected semantic vector and the corrected text report, and compares the similarity between the corrected semantic vector and the feature vector with the set similarity threshold, and outputs the comparison result;

[0011] S5. Generate an error correction report and visualize the error correction report through an interactive interface, allowing the user to review the error correction report and make a final confirmation. The user can choose to accept or reject the modification suggestions on their own, and a medical imaging report error correction confirmation report is generated and saved accordingly.

[0012] Preferably, the key area extraction process is as follows:

[0013] S21. Analyze medical imaging data and extract the imaging matrix I(x,y) through open source tools.

[0014] Among them, the image matrix I(x,y) represents the distribution of two-dimensional pixel grayscale values, which preliminarily describes the quality and structural details of medical image data;

[0015] S22. Standardize the image to eliminate differences between different scanning devices or imaging environments, and linearly convert the grayscale value range from device-specific values ​​to a unified standard range:

[0016]

[0017] In the formula, I norm (x, y) represents the normalized grayscale value; (x, y) represents the position coordinates of a pixel in the medical imaging data; I min and I max Respectively represent the minimum and maximum grayscale values ​​of the image matrix I(x,y);

[0018] S23. A contrast-limited adaptive histogram equalization method is used to enhance the local contrast of the image. The grayscale distribution of the local area in the image is adjusted to highlight important details while avoiding over-processing caused by global contrast adjustment. The calculation formula of the contrast-limited adaptive histogram equalization method is:

[0019]

[0020] In the formula, g new Represents the gray value of the processed image; clip() means to ensure that each gray value g new In the limited range [G min , G max ]Inside; G min and G max Respectively represent the minimum and maximum grayscale range of the image; CDF(g) represents the cumulative distribution function of the image grayscale value g; CDF min Represents the minimum cumulative distribution value; N clip Indicates the threshold set for limiting contrast;

[0021] S24, extract edge information from the image through the Canny edge detection algorithm, identify the contour of the lesion or organ, and based on the set high threshold T high and low threshold T low , determine the edge pixels and obtain the target area;

[0022]

[0023] In the formula, G(x,y) represents the pixel gradient value, that is, the edge strength; G x and G y Represents the gradient of the image in the x direction and the gradient in the y direction

[0024] If the gradient strength G(x,y) of a pixel is greater than the high threshold T high, then the pixel is considered to be an edge pixel; if the gradient strength G(x,y) of a pixel is lower than the high threshold T high , but connected to high gradient pixels belonging to the edge and whose gradient intensity is higher than the low threshold T low , and the pixel is also marked as an edge pixel; other pixels are marked as non-edge pixels, and the target area is obtained accordingly;

[0025] S25. Use the pre-trained UNet model to automatically segment the target area, automatically perform pixel-level segmentation based on the context information of the target area, and extract the key area in the target area.

[0026] Preferably, the feature vector extraction process is: using the image model to extract the features in the key area of ​​the medical image data, and obtaining the feature vector F = {F1, F2, ..., F i , ..., F m}.

[0027] Preferably, the keyword extraction process is as follows:

[0028] S41. Use TF-IDF weight method to extract keywords and set TF-IDF threshold TF-IDF min , only keep TF-IDF(w i ')≥TF-IDF min Keywords, get the initial keyword set K TI ={k1, k2, ..., k i , ..., k s}, k i represents the i-th keyword, s represents the number of initial keywords;

[0029] The calculation formula based on TF-IDF weight is as follows:

[0030] TF-IDF(w i ')=TF(w i ')×IDF(w i ');

[0031]

[0032] In the formula, TF-IDF(w i ') indicates word w i 'TF-IDF weight; w i ' represents the standardized word segmentation result W' = {w1', w'2, ..., w i ',...,w' n} the i-th normalized word in TF(w i ') indicates word w i'Word frequency in the report text data; n i Word w i 'Number of times in the report text data; Indicates the total number of words in the report text data; IDF(w i ') indicates word w i '; N represents the total number of all medical reports; n wi' Word w i 'Number of reports;

[0033] S42. Use the BERT-based sentence embedding model to further evaluate the importance of words and generate an attention score A(k i ), set the attention score threshold A min , if A(k i )<A min , then the semantic importance of the word is considered insufficient, and the word is removed to obtain the optimized keyword set K DL ={k1, k2, ..., k i , ..., k t}, t represents the number of optimized keywords;

[0034] The attention score formula is:

[0035] A(k i )=Softmax(W q ·Embed(k i )·(W k ) T );

[0036] In the formula, A(k i ) represents the initial keyword k i The attention score; Softmax() represents the Softmax function, that is, the normalization function, which maps the score to [0,1]; W q and W k represents the query and key weight matrix in the attention mechanism; Embed(k i ) represents the initial keyword k i The embedding vector of k ) T Represents the matrix W k The transposed matrix of

[0037] S43, supplement and optimize the keyword set by combining the medical field dictionary and regular expression rules to generate a final keyword set K = {k1, k2, ..., k i , ..., k h};

[0038] Among them, h represents the final number of keywords.

[0039] Preferably, the text model and the image model share a target embedding dimension m, and during the training process, the dimensionality and semantic consistency of the feature vector F and the semantic vector S are constrained by contrastive learning or cross-modal loss.

[0040] Preferably, the similarity between the feature vector and the semantic vector is calculated as follows:

[0041] S61. Extract the coupling information group according to the patient identification number ID <ID,T i , F> and coupling information group <ID,T i , S>, match the feature vector F with the semantic vector S, and get the coupling information group<F,S> ;

[0042] Where ID represents the patient identification number; T i It represents the time when the patient's medical imaging data is obtained, that is, the patient's detection time; F represents the feature vector; S represents the semantic vector;

[0043] S62, according to the coupling information group<F,S> , calculate the similarity between the feature vector F and the semantic vector S, the calculation formula is as follows:

[0044]

[0045] In the formula, sim(F,S) represents the similarity between the feature vector F and the semantic vector S; F·S represents the dot product of the feature vector F and the semantic vector S; F i and S i They respectively represent the component of the i-th dimension of the feature vector F and the component of the i-th dimension of the semantic vector S; ||F|| and ||S|| represent the modulus of the feature vector F and the modulus of the semantic vector S, respectively.

[0046] Preferably, the components of the automatic error correction generation model include:

[0047] The encoder is used to process the difference part ΔS of the input medical report semantic vector S. The encoder extracts the context information of the text through the self-attention mechanism to capture the semantic connection between the various parts of the report;

[0048] The decoder is used to generate the corrected semantic vector S' and the text report Text' based on the image feature vector F and the context information output by the encoder. The decoder also uses the self-attention mechanism and uses cross attention to associate the semantic information between the image and the report;

[0049] The self-attention mechanism is used to enable the automatic error correction generation model to pay attention to the information of all other words in the sequence when processing each word, thereby improving the ability to understand the context;

[0050] The cross-attention mechanism is used to generate a text report Text' with high consistency to the image information by combining the image features F with the reported semantic vector S.

[0051] Preferably, the training process of the automatic error correction generation model is as follows:

[0052] S81, collecting training data, where the training data includes a medical image, an image feature vector F, a corresponding medical report, and a semantic vector of the corresponding medical report;

[0053] S82, by comparing the image feature vector F with the semantic vector of the corresponding medical report, the training model learns how to generate a text report matching the image content;

[0054] S83. Define the loss function:

[0055] Where Loss represents the total loss value of the automatic error correction generation model, that is, the difference between the generated text report Text' and the target report. The smaller the value, the more consistent the semantics and content of the generated text report Text' with the target report; T represents the total number of words in the report, that is, the number of words in the generated text report Text'; P θ (y t |x1,x2,...,x t-1 ) represents the conditional probability of the t-th word predicted by the generated model under the model parameter θ, and the probability distribution of the t-th word is predicted based on the previous t-1 words; x1, x2, ..., x n Indicates the first t-1 words that the model has generated when generating the current report, that is, the context information output by the encoder;

[0056] S84, using the gradient descent optimization algorithm Adam optimizer to minimize the loss function and update the parameters of the model so that the quality of the generated text report is gradually improved;

[0057] S85. Train the model until the maximum number of iterations is reached or the model performance meets the requirements, obtain the automatic error correction generation model, and deploy the automatic error correction generation model to the automatic error correction module.

[0058] Preferably, the image model is a convolutional neural network and the text model is a Transformer network.

[0059] The technical solution of the present invention is a medical image report automatic quality control and error correction system, which is used to execute the above-mentioned medical image report automatic quality control and error correction method, comprising:

[0060] A medical image data processing module is used to extract medical image data from imaging equipment and extract feature vectors;

[0061] The report parsing module is used to perform word segmentation, grammatical analysis and semantic parsing on the generated medical report text and generate a semantic vector;

[0062] Image-report semantic matching module, used to check the semantic consistency between medical image features and report content;

[0063] Automatic error correction module, which combines the predefined medical terminology library and AI model to identify spelling errors, grammatical problems and semantic mismatches in reports and provide correction suggestions;

[0064] The user interaction module is used to provide an interactive interface for users, allowing them to review the error correction suggestions and make final confirmations, and generate quality control reports;

[0065] The system database is used to store data generated during the automated quality control and error correction process of medical image reports.

[0066] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0067] The present invention designs an automated quality control and error correction system and method for medical image reports, which provides a complete closed-loop process from error discovery, automatic analysis to generating correction reports. It can significantly improve the automation level of medical image and report quality control, reduce errors in diagnosis, and ensure the accuracy and completeness of diagnosis reports. At the same time, through intelligent feature extraction and semantic analysis, it improves the efficiency and quality of diagnosis, and provides innovative technical support for modern medical image analysis:

[0068] (1) Automatic consistency verification of image and report content. By extracting the key area feature vectors of medical image data and the semantic vectors of report text and performing similarity calculation, it automatically detects whether the content between medical images and reports matches, thereby improving the reliability of medical image diagnosis reports and reducing the risk of human omissions and errors.

[0069] (2) Intelligent error correction: When an inconsistency between an image and a report is found, the error type is automatically analyzed, including image description errors, missing information, and language expression errors. A pre-trained automatic error correction generation model is used to automatically generate a repaired report, thereby improving quality control and error correction efficiency while reducing the frequency of manual intervention.

[0070] (3) Efficient data processing and feature extraction: Through standardized processing and CLAHE enhancement technology, equipment differences are effectively eliminated and image quality is improved, ensuring more accurate feature extraction of key areas. Combining the Canny algorithm with the UNet model, key areas in medical images are accurately extracted, greatly improving the accuracy and efficiency of image information processing;

[0071] (4) Deep intelligence of report semantic analysis: The BERT model and Transformer network are used to segment, standardize and extract semantics from report text data, achieving high-dimensional semantic representation. Combined with the TF-IDF weighting method and BERT-based attention scoring, it can not only extract core keywords but also generate optimized semantic vectors, ensuring that the semantic expression of the report is more comprehensive and accurate.

[0072] (5) Cross-modal data fusion and consistency analysis: Through contrastive learning and cross-modal loss, the dimensionality and semantic consistency of the image feature vector and the text semantic vector are ensured, and the content is automatically verified through similarity threshold detection, which effectively reduces the inconsistency problem in multimodal data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a system architecture diagram of a medical image report automated quality control and error correction system proposed by the present invention;

[0074] Figure 2 This is a method flow chart of an automated quality control and error correction method for medical image reports proposed by the present invention. DETAILED DESCRIPTION

[0075] Embodiment 1, as Figure 1 As shown, the present invention proposes an automated quality control and error correction system for medical image reports, comprising: a medical image data processing module, a report parsing module, an image-report semantic matching module, an automatic error correction module, a user interaction module and a system database.

[0076] The medical image data processing module is used to extract medical image data from imaging devices (including but not limited to MRI, CT or X-ray) and extract feature vectors;

[0077] The report parsing module is used to perform word segmentation, grammatical analysis and semantic parsing on the generated medical report text to generate a semantic vector;

[0078] The image-report semantic matching module performs semantic consistency checks on medical image features and report content;

[0079] The automatic error correction module combines a predefined medical terminology library and AI models to identify spelling errors, grammatical problems, and semantic mismatches in reports and provide correction suggestions;

[0080] The user interaction module provides an interactive interface for users, allowing them to review the error correction suggestions and make final confirmations, and generate quality control reports;

[0081] The system database is used to store data generated during the automated quality control and error correction process of medical image reports.

[0082] Embodiment 2, as Figure 2 As shown, the present invention proposes an automated quality control and error correction method for medical image reports, which is applied to an automated quality control and error correction system for medical image reports proposed in Example 1, and its specific implementation steps are as follows:

[0083] S1. The medical image data processing module imports medical image data in DICOM format (DICOM format contains not only the image itself, but also the scanning device parameters and patient metadata), and performs the following operations:

[0084] S11. Analyze medical imaging data and extract the imaging matrix I(x,y) through open source tools (including but not limited to pydicom);

[0085] Among them, the image matrix I(x,y) represents the distribution of two-dimensional pixel grayscale values, which preliminarily describes the quality and structural details of medical image data;

[0086] S12. Standardize the image to eliminate differences between different scanning devices or imaging environments, and linearly convert the grayscale value range from device-specific values ​​to a unified standard range:

[0087]

[0088] In the formula, I norm (x, y) represents the normalized grayscale value; (x, y) represents the position coordinates of a pixel in the medical imaging data; I min and I max Respectively represent the minimum and maximum grayscale values ​​of the image matrix I(x,y);

[0089] S13. Contrast Limited Adaptive Histogram Equalization (CLAHE) is used to enhance the local contrast of the image. The grayscale distribution of the local area in the image is adjusted to highlight important details while avoiding over-processing caused by global contrast adjustment. The CLAHE process is as follows:

[0090]

[0091] In the formula, g newRepresents the gray value of the image after processing by the CLAHE method; clip() means to ensure that each gray value g new In the limited range [G min , G max ]Inside; G min and G max Respectively represent the minimum and maximum grayscale range of the image; CDF(g) represents the cumulative distribution function of the image grayscale value g; CDF min Represents the minimum cumulative distribution value; N clip Indicates the threshold set for limiting contrast;

[0092] S14, extract edge information from the image through the Canny edge detection algorithm, identify the contour of the lesion or organ, and based on the set high threshold T high and low threshold T low , determine the edge pixels and obtain the target area;

[0093]

[0094] In the formula, G(x,y) represents the pixel gradient value, that is, the edge strength; G x and G y Represents the gradient of the image in the x direction and the gradient in the y direction

[0095] If the gradient strength G(x,y) of a pixel is greater than the high threshold T high , then the pixel is considered to be an edge pixel; if the gradient strength G(x,y) of a pixel is lower than the high threshold T high , but connected to high gradient pixels belonging to the edge and whose gradient intensity is higher than the low threshold T low , and the pixel is also marked as an edge pixel; other pixels are marked as non-edge pixels, and the target area is obtained accordingly;

[0096] S15. Automatically segment the target area using the pre-trained UNet model, and automatically perform pixel-level segmentation based on the context information of the medical image to extract the key areas (including but not limited to lesions and organs) in the medical image;

[0097] It should be noted that the UNet model is a deep learning architecture designed specifically for image segmentation tasks. It adopts a unique U-shaped structure and consists of an encoder (Contracting Path) and a decoder (Expanding Path), which are tightly connected through skip connections. The encoder part gradually extracts the features of the input image through a series of convolutional layers, ReLU activation functions, and maximum pooling layers, while reducing the spatial dimension of the feature map and increasing the number of channels. The decoder part gradually restores the spatial resolution of the feature map through upsampling operations, and splices the feature map in the encoder with the feature map in the decoder through skip connections to retain more detailed information. This structure enables the UNet model to achieve high-precision image segmentation while maintaining computational efficiency.

[0098] S16, using an image model (a convolutional neural network is used in this embodiment) to extract features in key areas of the medical image data, and obtain a feature vector F={F1, F2, ..., F i , ..., F m};

[0099] Among them, F i represents the component of the ith dimension of the feature vector F; m represents the dimension of the feature vector F;

[0100] S17, obtain the patient identification number ID to which the medical imaging data belongs, and construct a coupling information group <ID,T i , F>, and the coupling information group <ID,T i , F>Save to the system database;

[0101] Where ID represents the patient identification number; T i It indicates the time when the medical imaging data of the patient is obtained, that is, the patient detection time.

[0102] S2. The report parsing module imports the report text data corresponding to the medical imaging data (including but not limited to free text description and structured data) and processes it. The specific implementation process is as follows:

[0103] S21. Decompose the report text data into words or phrases, combine the medical field dictionary and the pre-trained BERT model, split the sentence into individual words or phrases, and obtain the word sequence after word segmentation: W = {w1, w2, ..., w i , ..., w n};

[0104] Among them, W represents the word sequence after word segmentation; w i Represents the i-th word or phrase in the text;

[0105] S22. After word segmentation is completed, the word sequence W after word segmentation is normalized to obtain the normalized word segmentation result W' = {w1', w2',..., wi',..., wn'}, where wi' represents the i-th normalized word in the normalized word segmentation result; i ',..., w n} i ' represents the i-th normalized word in the normalized word segmentation result;

[0106] Among them, the normalization process includes but is not limited to:

[0107] Synonym processing: including but not limited to mapping terms such as "coronary heart disease" and "heart disease" to a unified word;

[0108] Abbreviation expansion: including but not limited to expanding "COPD" to "chronic obstructive pulmonary disease";

[0109] Redundancy removal: removing unimportant stop words (including but not limited to "de", "shi");

[0110] S23. From the normalized word segmentation result W' = {w1', w2',..., wi',..., wn'}, extract keywords with medical significance. The keywords reflect the core content of the report text data, including but not limited to lesion description, anatomical location, and diagnosis conclusion. The extraction process is as follows: i ',..., w n} to extract keywords with medical significance. The keywords reflect the core content of the report text data, including but not limited to lesion description, anatomical location, and diagnosis conclusion. The extraction process is as follows:

[0111] S2301. Use the TF-IDF weight method to extract keywords. Set the TF-IDF threshold TF-IDF, and only retain keywords where TF-IDF(wi') ≥ TF-IDF to obtain the initial keyword set K = {k1, k2,..., ki,..., ks}, where ki represents the i-th keyword and s represents the number of initial keywords; min , only retain TF-IDF(w i ') ≥ TF-IDF min of the keywords to obtain the initial keyword set K TI = {k1, k2,..., k i ,..., k s}, k i represents the i-th keyword, s represents the number of initial keywords;

[0112] The calculation formula based on the TF-IDF weight is as follows:

[0113] TF-IDF(wi i ') = TF(wi i ') × IDF(wi i ');

[0114]

[0115] In the formula, TF-IDF(wi i ') represents the TF-IDF weight of the word wi i '; wii ' represents the i-th standardized word in the standardized word segmentation result; TF(w i ') indicates word w i 'Word frequency in the report text data; n i Word w i 'Number of times in the report text data; Indicates the total number of words in the report text data; IDF(w i ') indicates word w i '; N represents the total number of all medical reports; n wi' Word w i 'Number of reports;

[0116] S2302: In order to capture the semantic association of keywords, the importance of words is further evaluated using the BERT-based sentence embedding model. Through the attention mechanism, the model generates an attention score A(k i ), set the attention score threshold A min , if A(k i )<A min , then the word is considered to be of insufficient semantic importance and is removed, and the optimized keyword set K is obtained. DL ={k1, k2, ..., k i , ..., k t}, t represents the number of optimized keywords;

[0117] The attention score formula is:

[0118] A(k i )=Softmax(W q ·Embed(k i )·(W k ) T );

[0119] In the formula, A(k i ) represents the initial keyword k i The attention score; Softmax() represents the Softmax function, that is, the normalization function, which maps the score to [0,1]; W q and W k represents the query and key weight matrix in the attention mechanism; Embed(k i ) represents the initial keyword k i The embedding vector of k ) T Represents the matrix W k The transposed matrix of

[0120] S2303: Supplement and optimize the keyword set by combining the medical field dictionary (including but not limited to SNOMED CT or a custom word library) and regular expression rules to generate a final keyword set K = {k1, k2, ..., k i , ..., k h};

[0121] Among them, h represents the final number of keywords;

[0122] S2304: Convert each keyword into an embedded vector representation and output a keyword vector set E k ={e1, e2, ..., e i , ..., e h};

[0123] Among them, e i Indicates keyword k i The embedding vector of

[0124] S24, use the text model (the present embodiment uses the Transformer network) for end-to-end training, automatically learn how to combine the embedding vectors of each keyword to generate a more accurate report semantic vector S = {S1, S2, ..., S i , ..., S m};

[0125] Among them, the semantic vector S represents the deep semantic information of the report text data; S i Represents the component of the i-th dimension of the semantic vector S;

[0126] It should be noted that the text model and the image model share a target embedding dimension m. During the training process, the dimension and semantic consistency of the feature vector F and the semantic vector S are constrained through contrastive learning or cross-modal loss.

[0127] S25. Obtain the patient identification number ID to which the report text data belongs, and construct a coupling information group <ID,T i , S>, and the coupling information group <ID,T i , S>Save to the system database.

[0128] S3, extracting the feature vector F in the key area of ​​the medical image data and the semantic vector S of the corresponding medical image report text data, and performing image-report consistency detection, the specific implementation process of which is as follows:

[0129] S31, according to the patient identification number ID, obtain the coupling information group from the system database <ID,T i, F> and coupling information group <ID,T i , S>, match the feature vector F with the semantic vector S, and get the coupling information group<F,S> ;

[0130] S32, according to the coupling information group<F,S> , calculate the similarity between the feature vector F and the semantic vector S, the calculation formula is as follows:

[0131]

[0132] In the formula, sim(F,S) represents the similarity between the feature vector F and the semantic vector S; F·S represents the dot product of the feature vector F and the semantic vector S; F i and S i They represent the component of the i-th dimension of the feature vector F and the component of the i-th dimension of the semantic vector S respectively; ||F|| and ||S|| represent the modulus of the feature vector F and the modulus of the semantic vector S respectively;

[0133] S33, compare sim(F,S) with the set similarity threshold θ t To compare:

[0134] If sim(F,S)>θ t , it is considered that the image and the report are semantically consistent, the content of the report matches the content of the image, and a matching pass message is sent to the user interaction module, and the feature vector F and the semantic vector S are transmitted to the user interaction module;

[0135] If sim(F,S)≤θ t , it is considered that there is inconsistency between the content of the image and the content of the report, which may be due to errors in the description in the report, or the image fails to accurately reflect the content of the report. The difference part ΔS in the feature vector F and the semantic vector S is marked, and an error correction instruction is issued to the automatic error correction module, and the feature vector F, the semantic vector S and the difference part ΔS are transmitted to the automatic error correction module and the user interaction module.

[0136] S4. The automatic error correction module receives the error correction instruction, the feature vector F, the semantic vector S and the difference part ΔS, starts the automatic error correction program, and implements the following operation process:

[0137] S41, the automatic error correction module analyzes the error type;

[0138] Among them, the error types include the following three categories:

[0139] Image description error: the report content does not match the actual information in the image;

[0140] Missing image information: The image information is not fully reflected in the report;

[0141] Language errors: The report contains grammatical problems or inaccurate medical terms;

[0142] S42, inputting the feature vector F, the semantic vector S and the difference part ΔS into the pre-trained automatic error correction generation model, the automatic error correction generation model repairs the input semantic vector S to obtain the corrected semantic vector S' and the text report Text';

[0143] S43, calculate the similarity between the corrected semantic vector S' and the feature vector F:

[0144]

[0145] In the formula, sim(F,S') represents the similarity between the feature vector F and the corrected semantic vector S';

[0146] S44, compare sim(F,S') with the set similarity threshold θ t To compare:

[0147] If sim(F,S')>θ t , it is considered that the content of the repaired report matches the image content, and the error correction pass information is sent to the user interaction module, and the text report Text' is transmitted to the user interaction module;

[0148] If sim(F,S)≤θ t , it is marked as repair failure, and a correction failure message is sent to the user interaction module, which displays that manual review is required.

[0149] S5. The user interaction module generates an error correction report and visualizes the error correction report through an interactive interface, allowing the user to review the error correction report and make a final confirmation. The user can choose to accept or reject the modification suggestions. Based on this, the user interaction module generates and saves the medical imaging report error correction confirmation report.

[0150] Embodiment 3, the present invention proposes a method for automatic quality control and error correction of medical image reports, further comprising an automatic error correction generation model, the components of which are as follows:

[0151] Encoder: It is used to process the difference part ΔS of the input medical report semantic vector S. The encoder extracts the context information of the text through the self-attention mechanism to capture the semantic connection between the various parts of the report.

[0152] Decoder: Generates the corrected semantic vector S' and text report Text' based on the image feature vector F and the context information output by the encoder. The decoder also uses the self-attention mechanism and uses cross-attention to associate the semantic information between the image and the report.

[0153] Self-Attention Mechanism: This allows the automatic error correction generation model to pay attention to the information of all other words in the sequence when processing each word, thereby improving the ability to understand the context;

[0154] Cross-Attention Mechanism: By combining the image features F with the reported semantic vector S, a text report Text' is generated that is highly consistent with the image information.

[0155] The training process of the automatic error correction generation model is as follows:

[0156] S1. Collect training data, where the training data includes a medical image, an image feature vector F, a corresponding medical report, and a semantic vector of the corresponding medical report;

[0157] S2, by comparing the image feature vector F with the semantic vector of the corresponding medical report, the training model learns how to generate a text report that matches the image content;

[0158] S3. Define the loss function:

[0159] Where Loss represents the total loss value of the automatic error correction generation model, that is, the difference between the generated text report Text' and the target report. The smaller the value, the more consistent the semantics and content of the generated text report Text' with the target report; T represents the total number of words in the report, that is, the number of words in the generated text report Text'; P θ (y t |x1,x2,...,x t-1 ) represents the conditional probability of the t-th word predicted by the generated model under the model parameter θ, and the probability distribution of the t-th word is predicted based on the previous t-1 words; x1, x2, ..., x n Indicates the first t-1 words that the model has generated when generating the current report, that is, the context information output by the encoder;

[0160] S4. Use the gradient descent optimization algorithm Adam optimizer to minimize the loss function and update the parameters of the model so that the quality of the generated text report is gradually improved;

[0161] S5. Train the model until the maximum number of iterations is reached or the model performance meets the requirements, obtain the automatic error correction generation model, and deploy the automatic error correction generation model to the automatic error correction module.

[0162] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for automated quality control and error correction of medical image reports, characterized in that: The specific implementation steps include the following: S1. Import medical image data, analyze the medical image data, standardize the image, use contrast-limited adaptive histogram equalization method to enhance the local contrast of the image, then extract the target area, automatically perform pixel-level segmentation based on the context information of the target area, extract the key area, and obtain the feature vector based on it, and then construct a coupled information group <patient identification number, patient detection time, feature vector>; S2, import the report text data corresponding to the medical imaging data, split the report text data into word sequences through word segmentation technology, generate a standardized word set, extract keywords, convert each keyword into an embedded vector representation, output a keyword vector set, and use the text model to generate a semantic vector, and construct a coupled information group <patient identification number, patient detection time, semantic vector> based on this; S3, extracting the feature vector of the medical image data and the semantic vector of the corresponding report text data, calculating the similarity between the feature vector and the semantic vector, and comparing the similarity with the set similarity threshold: if the similarity exceeds the set similarity threshold, it is considered that the content of the report matches the content of the image; otherwise, it is considered that there is an inconsistency between the content of the image and the content of the report, identifying the difference, and executing S4; S4, start the automatic error correction program, input the feature vector, the semantic vector and the difference part into the pre-trained automatic error correction generation model, the automatic error correction generation model repairs the input semantic vector, obtains the corrected semantic vector and the corrected text report, and compares the similarity between the corrected semantic vector and the feature vector with the set similarity threshold, and outputs the comparison result; S5. Generate an error correction report and visualize the error correction report through an interactive interface, allowing the user to review the error correction report and make a final confirmation. The user can choose to accept or reject the modification suggestions on their own, and a medical imaging report error correction confirmation report is generated and saved accordingly.

2. The method for automatic quality control and error correction of medical image reports according to claim 1, characterized in that: The key area extraction process is as follows: S21. Analyze medical imaging data and extract the imaging matrix I(x,y) through open source tools. Among them, the image matrix I(x,y) represents the distribution of two-dimensional pixel grayscale values, which preliminarily describes the quality and structural details of medical image data; S22. Standardize the image to eliminate differences between different scanning devices or imaging environments, and linearly convert the grayscale value range from device-specific values ​​to a unified standard range: In the formula, I norm (x, y) represents the normalized grayscale value; (x, y) represents the position coordinates of a pixel in the medical imaging data; I min and I max Respectively represent the minimum and maximum grayscale values ​​of the image matrix I(x,y); S23. A contrast-limited adaptive histogram equalization method is used to enhance the local contrast of the image. The grayscale distribution of the local area in the image is adjusted to highlight important details while avoiding over-processing caused by global contrast adjustment. The calculation formula of the contrast-limited adaptive histogram equalization method is: In the formula, g new Represents the gray value of the processed image; clip() means to ensure that each gray value g new In the limited range [G min , G max ]Inside; G min and G max Respectively represent the minimum and maximum grayscale range of the image; CDF(g) represents the cumulative distribution function of the image grayscale value g; CDF min Represents the minimum cumulative distribution value; N clip Indicates the threshold set for limiting contrast; S24, extract edge information from the image through the Canny edge detection algorithm, identify the contour of the lesion or organ, and based on the set high threshold T high and low threshold T low , determine the edge pixels and obtain the target area; In the formula, G(x,y) represents the pixel gradient value, that is, the edge strength; G x and G y Represents the gradient of the image in the x direction and the gradient in the y direction If the gradient strength G(x,y) of a pixel is greater than the high threshold T high , then the pixel is considered to be an edge pixel; if the gradient strength G(x,y) of a pixel is lower than the high threshold T high , but connected to high gradient pixels belonging to the edge and whose gradient intensity is higher than the low threshold T low , and the pixel is also marked as an edge pixel; other pixels are marked as non-edge pixels, and the target area is obtained accordingly; S25. Use the pre-trained UNet model to automatically segment the target area, automatically perform pixel-level segmentation based on the context information of the target area, and extract the key area in the target area.

3. The method for automatic quality control and error correction of medical image reports according to claim 1, characterized in that: The feature vector extraction process is as follows: using the image model to extract the features in the key areas of the medical image data, and obtaining the feature vector F = {F1, F2, ..., F i , ..., F m }.

4. The method for automatic quality control and error correction of medical image reports according to claim 1, characterized in that: The keyword extraction process is as follows: S41. Use TF-IDF weight method to extract keywords and set TF-IDF threshold TF-IDF min , only TF-IDF(w i ')≥TF-IDF min Keywords, get the initial keyword set K TI ={k1, k2, ..., k i , ..., k s }, k i represents the i-th keyword, s represents the number of initial keywords; The calculation formula based on TF-IDF weight is as follows: TF-IDF(w i ')=TF(w i ')×IDF(w i '); In the formula, TF-IDF(w i ') indicates word w i 'TF-IDF weight; w i ' represents the standardized word segmentation result W' = {w1', w'2, ..., w i ',...,w' n } the i-th normalized word in TF(w i ') indicates word w i 'Word frequency in the report text data; n i Word w i 'Number of times in the report text data; Indicates the total number of words in the report text data; IDF(w i ') indicates word w i ''s inverse document frequency; N represents the total number of all medical reports; Word w i 'Number of reports; S42. Use the BERT-based sentence embedding model to further evaluate the importance of words and generate an attention score A(k i ), set the attention score threshold A min , if A(k i )<A min , then the semantic importance of the word is considered insufficient, and the word is removed to obtain the optimized keyword set K DL ={k1, k2, ..., k i , ..., k t }, t represents the number of optimized keywords; The attention score formula is: A(k i )=Softmax(W q ·Embed(k i )·(W k ) T ); In the formula, A(k i ) represents the initial keyword k i The attention score; Softmax() represents the Softmax function, that is, the normalization function, which maps the score to [0,1]; W q and W k represents the query and key weight matrix in the attention mechanism; Embed(k i ) represents the initial keyword k i The embedding vector of k ) T Represents the matrix W k The transposed matrix of S43, supplement and optimize the keyword set by combining the medical field dictionary and regular expression rules to generate a final keyword set K = {k1, k2, ..., k i , ..., k h }; Among them, h represents the final number of keywords.

5. The method for automatic quality control and error correction of medical image reports according to claim 1, characterized in that: The text model and the image model share a target embedding dimension m. During the training process, the dimension and semantic consistency of the feature vector F and the semantic vector S are constrained through contrastive learning or cross-modal loss.

6. The method for automatic quality control and error correction of medical image reports according to claim 1, characterized in that: The calculation process of the similarity between the feature vector and the semantic vector is: S61. Extract the coupling information group according to the patient identification number ID <ID,T i , F> and coupling information group <ID,T i , S>, match the feature vector F with the semantic vector S, and get the coupling information group<F,S> ; Where ID represents the patient identification number; T i It represents the time when the patient's medical imaging data is obtained, that is, the patient's detection time; F represents the feature vector; S represents the semantic vector; S62, according to the coupling information group<F,S> , calculate the similarity between the feature vector F and the semantic vector S, the calculation formula is as follows: In the formula, sim(F,S) represents the similarity between the feature vector F and the semantic vector S; F·S represents the dot product of the feature vector F and the semantic vector S; F i and S i They respectively represent the component of the i-th dimension of the feature vector F and the component of the i-th dimension of the semantic vector S; ||F|| and ||S|| represent the modulus of the feature vector F and the modulus of the semantic vector S, respectively.

7. The method for automatic quality control and error correction of medical image reports according to claim 1, characterized in that: The components of the automatic error correction generation model include: The encoder is used to process the difference part ΔS of the input medical report semantic vector S. The encoder extracts the context information of the text through the self-attention mechanism to capture the semantic connection between the various parts of the report; The decoder is used to generate the corrected semantic vector S' and the text report Text' based on the image feature vector F and the context information output by the encoder. The decoder also uses the self-attention mechanism and uses cross attention to associate the semantic information between the image and the report; The self-attention mechanism is used to enable the automatic error correction generation model to pay attention to the information of all other words in the sequence when processing each word, thereby improving the ability to understand the context; The cross-attention mechanism is used to generate a text report Text' with high consistency to the image information by combining the image features F with the reported semantic vector S.

8. The method for automatic quality control and error correction of medical image reports according to claim 7, characterized in that: The training process of the automatic error correction generation model is as follows: S81, collecting training data, where the training data includes a medical image, an image feature vector F, a corresponding medical report, and a semantic vector of the corresponding medical report; S82, by comparing the image feature vector F with the semantic vector of the corresponding medical report, the training model learns how to generate a text report matching the image content; S83. Define the loss function: Where Loss represents the total loss value of the automatic error correction generation model, that is, the difference between the generated text report Text' and the target report. The smaller the value, the more consistent the semantics and content of the generated text report Text' with the target report; T represents the total number of words in the report, that is, the number of words in the generated text report Text'; P θ (y t |x1,x2,...,x t-1 ) represents the conditional probability of the t-th word predicted by the generated model under the model parameter θ, and the probability distribution of the t-th word is predicted based on the previous t-1 words; x1, x2, ..., x n Indicates the first t-1 words that the model has generated when generating the current report, that is, the context information output by the encoder; S84, using the gradient descent optimization algorithm Adam optimizer to minimize the loss function and update the parameters of the model so that the quality of the generated text report is gradually improved; S85. Train the model until the maximum number of iterations is reached or the model performance meets the requirements, obtain the automatic error correction generation model, and deploy the automatic error correction generation model to the automatic error correction module.

9. The method for automatic quality control and error correction of medical image reports according to claim 1, characterized in that: The image model is a convolutional neural network and the text model is a Transformer network.

10. A medical image report automated quality control and error correction system, which is used to execute the medical image report automated quality control and error correction method according to claims 1 to 9, characterized in that: include: A medical image data processing module, used to extract medical image data from imaging equipment and extract feature vectors; The report parsing module is used to perform word segmentation, grammatical analysis and semantic parsing on the generated medical report text and generate a semantic vector; Image-report semantic matching module, used to check the semantic consistency between medical image features and report content; Automatic error correction module, which combines the predefined medical terminology library and AI model to identify spelling errors, grammatical problems and semantic mismatches in reports and provide correction suggestions; The user interaction module is used to provide an interactive interface for users, allowing them to review the error correction suggestions and make final confirmations, and generate quality control reports; The system database is used to store data generated during the automated quality control and error correction process of medical image reports.

Images