A renal artery ultrasound report generation method and system, electronic equipment and storage medium

Abstract

This application discloses a method, system, electronic device, and storage medium for generating renal artery ultrasound reports. The method includes: keyword matching of the clinical examination text of the ultrasound case to obtain renal artery diagnostic indicators; generating search prompts based on the renal artery diagnostic indicators; acquiring general medical knowledge and clinical expert specialty knowledge from a large model and obtaining global dual-source knowledge embedding; encoding the clinical examination text of the ultrasound case to generate basic embedding; obtaining knowledge fusion residual features based on the basic embedding and global dual-source knowledge embedding; extracting features from multiple frames of ultrasound images of the renal artery to be examined to obtain visual global features; performing element-wise additive fusion of the visual global features, global dual-source knowledge embedding, and knowledge fusion residual features to obtain global unified features; and generating a renal artery ultrasound diagnostic report based on the global unified features. Through the above settings, the diagnostic accuracy and terminology standardization of the report can be improved.

Description

Technical Field

[0001] This application relates to the field of medical technology, and in particular to a method, system, electronic device and storage medium for generating renal artery ultrasound reports. Background Technology

[0002] With the widespread application of deep learning technology in medical ultrasound image analysis, intelligent generation of renal artery ultrasound reports based on artificial intelligence has become an important research direction for improving the efficiency of renal artery disease diagnosis and reducing the workload of physicians in paperwork. Existing ultrasound report generation methods mostly employ an encoding and decoding architecture, using a visual encoder to extract ultrasound image features and then a sequence decoder to generate a text report.

[0003] However, existing methods rely solely on visual information from a single image for report generation, resulting in insufficient diagnostic accuracy, terminology standardization, and clinical adaptability of the reports, making it difficult to meet the actual clinical needs for accurate diagnosis and standardized report output for renal artery stenosis. Summary of the Invention

[0004] To address the shortcomings mentioned above, this application provides a method, system, electronic device, and storage medium for generating renal artery ultrasound reports.

[0005] Firstly, to achieve the above objectives, this application provides a method for generating a renal artery ultrasound report, comprising: Keyword matching was performed on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, and a diagnostic factor pool was constructed based on the renal artery diagnostic indicators. Based on the diagnostic factor pool, search suggestions are generated, and general medical knowledge and clinical expert specialty knowledge of the large model are obtained. The general medical knowledge and clinical expert specialty knowledge of the large model are encoded to obtain global dual-source knowledge embedding. The clinical examination text of ultrasound cases is encoded to generate basic embeddings. Based on the basic embeddings and global dual-source knowledge embeddings, knowledge-aware features are obtained. The knowledge-aware features are fused with the basic embeddings to obtain knowledge fusion residual features. Feature extraction is performed on multiple frames of ultrasound images of the renal artery to be tested to obtain global visual features. The global visual features, global dual-source knowledge embedding, and knowledge fusion residual features are then additively fused element by element to obtain a unified global feature. Based on the globally consistent characteristics, a renal artery ultrasound diagnostic report is obtained.

[0006] Preferably, keyword matching is performed on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, wherein: Renal artery diagnostic indicators include peak PSV velocity, resistance index (RI), acceleration time, vessel diameter, lesion morphology, and stenosis grade.

[0007] Preferably, search suggestions are generated based on the diagnostic factor pool to obtain general medical knowledge from the large model and specialized knowledge from clinical experts. This general medical knowledge and specialized knowledge from clinical experts are then encoded to obtain a global dual-source knowledge embedding, including: Based on the diagnostic factor pool, search suggestions are generated, and searches are initiated simultaneously to the first knowledge source and the second knowledge source. The first knowledge source is a general language model, which outputs general medical knowledge including the anatomical principles of the renal artery, the physiological and pathological mechanisms of blood flow indicators, and general common sense in basic medicine. The second knowledge source is the Renal Artery Specialty Diagnosis and Treatment Knowledge Base, which stores clinical expert knowledge including diagnostic thresholds for renal artery stenosis, blood flow parameter interpretation standards, lesion grading and treatment guidelines, and clinical report writing guidelines. After aggregating and integrating general medical knowledge from large models with specialized knowledge from clinical experts, the data is input into a knowledge encoder for unified dimensional encoding, generating a global dual-source knowledge embedding.

[0008] Preferably, the clinical examination text of ultrasound cases is encoded to generate a basic embedding. Based on the basic embedding and the global dual-source knowledge embedding, knowledge-aware features are obtained. The knowledge-aware features are then fused with the basic embedding to obtain knowledge fusion residual features, including: The clinical examination text of the ultrasound case is input into a text encoder to generate a basic embedding; Query vectors are generated from basic embeddings, and cross-attention operations are performed on global dual-source knowledge embeddings to obtain knowledge-aware features. Generate gated weights by applying the basic embedded linear transformation and Sigmoid activation; Based on the gating weights, the knowledge-aware features are fused with the basic embeddings to obtain the knowledge fusion residual features.

[0009] Preferably, the knowledge-aware features are fused with the basic embeddings according to the gating weights to obtain the knowledge fusion residual feature R, as follows: R=g⊙ +(1-g)⊙ ; In the formula: Based on embedding; denoted as knowledge perception feature; g represents the gating weight.

[0010] Preferably, the visual global features, global dual-source knowledge embedding, and knowledge fusion residual features are additively fused element-wise to obtain a globally unified feature, as shown in the following formula: = ⊕ ⊕R; In the formula: R represents the residual feature of knowledge fusion; Embedded with global dual-source knowledge; It represents the global visual features.

[0011] Preferably, a renal artery ultrasound diagnostic report is obtained based on globally consistent features, including: Input the globally unified features into the decoder; Contextual features are obtained through multi-head self-attention masking in the decoder; Normalize the context features to obtain normalized context features; Align the normalized context features and the globally unified features to obtain aligned features; The alignment features are subjected to a nonlinear feature transformation to obtain the transformed features; After layer normalization of the transformed features, they are mapped to the vocabulary space through the fully connected layer of the decoder to obtain the vocabulary probability distribution; The renal artery ultrasound diagnostic report is generated by iterative autoregression based on the probability distribution of words.

[0012] Secondly, this application also provides a renal artery ultrasound report generation system, comprising: The acquisition module performs keyword matching on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, and constructs a diagnostic factor pool based on the renal artery diagnostic indicators. It generates search prompts based on the diagnostic factor pool, acquires general medical knowledge from the large-scale model and specialized knowledge from clinical experts, encodes the general medical knowledge from the large-scale model and the specialized knowledge from clinical experts to obtain a global dual-source knowledge embedding. It encodes the clinical examination text of ultrasound cases to generate a basic embedding, obtains knowledge-aware features based on the basic embedding and the global dual-source knowledge embedding, and fuses the knowledge-aware features with the basic embedding to obtain knowledge fusion residual features. The processing module is used to extract features from the multi-frame ultrasound images of the renal artery to be detected to obtain global visual features. The global visual features, global dual-source knowledge embedding, and knowledge fusion residual features are then additively fused element by element to obtain global unified features. The execution module is used to generate a renal artery ultrasound diagnostic report based on globally unified features.

[0013] Thirdly, this application also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described method.

[0014] Fourthly, this application also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: This application generates search prompts based on a diagnostic factor pool and encodes dual-source medical knowledge, enabling the simultaneous introduction of general medical knowledge and specialized diagnostic and treatment knowledge to construct a comprehensive and highly professional external knowledge support system. This compensates for the deficiencies in diagnostic evidence and terminology that result from relying solely on single visual information. Furthermore, by generating renal artery ultrasound diagnostic reports based on globally unified features, it significantly improves the diagnostic accuracy, terminology standardization, and clinical adaptability of the reports, meeting the clinical needs for accurate diagnosis and standardized report output for renal artery stenosis. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method for generating renal artery ultrasound reports in this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] This application provides a method for generating a renal artery ultrasound report, including: Keyword matching was performed on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, and a diagnostic factor pool was constructed based on the renal artery diagnostic indicators. Specifically, keyword matching is performed on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, including peak PSV velocity, resistance index (RI), acceleration time, vessel diameter, lesion morphology, and stenosis grade. Through this setup, structured parsing of case texts can be achieved, constructing a diagnostic factor pool tailored to the renal artery diagnosis and treatment scenario. This provides accurate and dedicated retrieval criteria for subsequent knowledge searches, avoiding mismatches or discrepancies between knowledge and case information.

[0019] Based on the diagnostic factor pool, search suggestions are generated, and general medical knowledge and clinical expert specialty knowledge of the large model are obtained. The general medical knowledge and clinical expert specialty knowledge of the large model are encoded to obtain global dual-source knowledge embedding. Specifically, search suggestions are generated based on the diagnostic factor pool, and searches are initiated simultaneously into a first knowledge source and a second knowledge source. The first knowledge source is a general-purpose large language model, outputting general medical knowledge including renal artery anatomy, the physiological and pathological mechanisms of blood flow parameters, and basic medical common sense. The second knowledge source is a renal artery specialist diagnosis and treatment knowledge base, which stores clinical expert specialist knowledge including renal artery stenosis diagnostic thresholds, blood flow parameter interpretation standards, lesion grading and treatment guidelines, and clinical report writing guidelines. The general-purpose medical knowledge from the large model and the clinical expert specialist knowledge are aggregated and integrated, then input into a knowledge encoder for unified dimensional encoding, generating a global dual-source knowledge embedding. Through this setup, both general medical common sense and specialist diagnosis and treatment guidelines can be introduced simultaneously, constructing a comprehensive and highly professional external knowledge system, compensating for the deficiencies of relying solely on single visual information, such as insufficient diagnostic evidence, non-standard terminology, and deviation from clinical guidelines.

[0020] The clinical examination text of ultrasound cases is encoded to generate basic embeddings. Based on the basic embeddings and global dual-source knowledge embeddings, knowledge-aware features are obtained. The knowledge-aware features are fused with the basic embeddings to obtain knowledge fusion residual features. Specifically, the clinical examination text of the ultrasound case is input into a text encoder to generate a basic embedding; Query vectors are generated from basic embeddings, and cross-attention operations are performed on global dual-source knowledge embeddings to obtain knowledge-aware features. Generate gated weights by applying the basic embedded linear transformation and Sigmoid activation; Based on the gating weights, the knowledge-aware features are fused with the basic embeddings to obtain the knowledge fusion residual features. The formula for the fused residual feature R is as follows: R=g⊙ +(1-g)⊙ ; In the formula: Based on embedding; denoted as knowledge perception feature; g represents the gating weight.

[0021] Through the above settings, cross-attention achieves accurate matching of case information and dual-source knowledge, and adopts dynamic gating weight adaptive fusion, which can avoid the knowledge redundancy, conflict and overload problems caused by fixed hard splicing. While retaining the original examination information, it strengthens effective pathological knowledge and improves the pertinence and reliability of features.

[0022] Feature extraction is performed on multiple frames of ultrasound images of the renal artery to be tested to obtain global visual features. The global visual features, global dual-source knowledge embedding, and knowledge fusion residual features are then additively fused element by element to obtain a unified global feature. Specifically, the fusion formula is as follows: = ⊕ ⊕R; In the formula: R represents the residual feature of knowledge fusion; Embedded with global dual-source knowledge; This provides a global visual feature. Through the above settings, it is possible to fully integrate the visual features of ultrasound images, dual-source external knowledge features, and original case examination features, forming a globally unified feature with complete representation and strong semantic constraints, thus solving the problems of insufficient single-modal information and inadequate utilization of multi-source information.

[0023] Based on the globally consistent characteristics, a renal artery ultrasound diagnostic report is obtained.

[0024] Specifically, globally unified features are input into the decoder; Contextual features are obtained through multi-head self-attention masking in the decoder; Normalize the context features to obtain normalized context features; Align the normalized context features and the globally unified features to obtain aligned features; The alignment features are subjected to a nonlinear feature transformation to obtain the transformed features; After layer normalization of the transformed features, they are mapped to the vocabulary space through the fully connected layer of the decoder to obtain the vocabulary probability distribution; The renal artery ultrasound diagnostic report is generated by iterative autoregression based on the probability distribution of words.

[0025] With the above settings, autoregressive decoding based on globally unified features can generate renal artery ultrasound reports with standardized structure, accurate terminology, and rigorous diagnostic logic under the joint constraints of visual information, original examination information, and dual-source clinical knowledge. This significantly improves the accuracy, standardization, and clinical adaptability of the reports, meeting the needs for accurate diagnosis and standardized output of renal artery stenosis.

[0026] Secondly, this application also provides a renal artery ultrasound report generation system, comprising: The acquisition module performs keyword matching on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, and constructs a diagnostic factor pool based on the renal artery diagnostic indicators. It generates search prompts based on the diagnostic factor pool, acquires general medical knowledge from the large-scale model and specialized knowledge from clinical experts, encodes the general medical knowledge from the large-scale model and the specialized knowledge from clinical experts to obtain a global dual-source knowledge embedding. It encodes the clinical examination text of ultrasound cases to generate a basic embedding, obtains knowledge-aware features based on the basic embedding and the global dual-source knowledge embedding, and fuses the knowledge-aware features with the basic embedding to obtain knowledge fusion residual features. The processing module is used to extract features from the multi-frame ultrasound images of the renal artery to be detected to obtain global visual features. The global visual features, global dual-source knowledge embedding, and knowledge fusion residual features are then additively fused element by element to obtain global unified features. The execution module is used to generate a renal artery ultrasound diagnostic report based on globally unified features.

[0027] This application also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described method.

[0028] This application also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0029] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating a renal artery ultrasound report, characterized in that, include: Keyword matching was performed on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, and a diagnostic factor pool was constructed based on the renal artery diagnostic indicators. Based on the diagnostic factor pool, search prompts are generated, and general medical knowledge and clinical expert specialty knowledge of the large model are obtained. The general medical knowledge and clinical expert specialty knowledge of the large model are encoded to obtain global dual-source knowledge embedding. The clinical examination text of the ultrasound case is encoded to generate a basic embedding. Based on the basic embedding and the global dual-source knowledge embedding, knowledge-aware features are obtained. The knowledge-aware features are fused with the basic embedding to obtain knowledge fusion residual features. Visual global features are obtained by extracting features from multi-frame ultrasound images of the renal artery to be detected. The visual global features, the global dual-source knowledge embedding, and the knowledge fusion residual features are then additively fused element-by-element to obtain global unified features. Based on the aforementioned global unified features, a renal artery ultrasound diagnostic report is obtained.

2. The method for generating a renal artery ultrasound report according to claim 1, characterized in that, The clinical examination text of the ultrasound case was subjected to keyword matching to obtain renal artery diagnostic indicators, among which: The renal artery diagnostic indicators include peak PSV velocity, resistance index (RI), acceleration time, vessel diameter, lesion morphology, and stenosis grade.

3. The method for generating a renal artery ultrasound report according to claim 1, characterized in that, The process involves generating search suggestions based on the diagnostic factor pool, acquiring general medical knowledge from the large model and specialized knowledge from clinical experts, encoding the general medical knowledge from the large model and specialized knowledge from clinical experts to obtain a global dual-source knowledge embedding, including: Based on the diagnostic factor pool, the search suggestion words are generated, and searches are initiated to the first knowledge source and the second knowledge source simultaneously. The first knowledge source is a general large language model, which outputs general medical knowledge of the large model, including the anatomical principles of the renal artery, the physiological and pathological mechanisms of blood flow indicators, and basic medical common sense. The second knowledge source is a renal artery specialist diagnosis and treatment knowledge base, which stores clinical expert specialist knowledge including renal artery stenosis diagnostic thresholds, blood flow parameter interpretation standards, lesion grading and treatment guidelines, and clinical report writing guidelines. The general medical knowledge of the large model and the specialized knowledge of the clinical experts are summarized and integrated, and then input into the knowledge encoder for unified dimensional encoding to generate the global dual-source knowledge embedding.

4. The method for generating a renal artery ultrasound report according to claim 2, characterized in that, The process involves encoding the clinical examination text of the ultrasound case to generate a basic embedding, obtaining knowledge-aware features based on the basic embedding and the global dual-source knowledge embedding, and fusing the knowledge-aware features with the basic embedding to obtain knowledge fusion residual features, including: The clinical examination text of the aforementioned ultrasound case is input into a text encoder to generate a basic embedding; A query vector is generated using the aforementioned basic embedding, and a cross-attention operation is performed on the global dual-source knowledge embedding to obtain the knowledge-aware features. The basic embedded linear transformation and Sigmoid activation are used to generate gated weights; The knowledge-aware features are fused with the basic embedding based on the gating weights to obtain knowledge fusion residual features.

5. The method for generating a renal artery ultrasound report according to claim 4, characterized in that, The formula for fusing the knowledge-aware features with the basic embedding based on the gating weights to obtain the knowledge fusion residual feature R is as follows: R=g⊙ +(1-g)⊙ ; In the formula: For the aforementioned basic embedding; denoted as the knowledge perception feature; g represents the gating weight.

6. The method for generating a renal artery ultrasound report according to claim 1, characterized in that, The global visual features, the global dual-source knowledge embedding, and the knowledge fusion residual features are additively fused element-wise to obtain a unified global feature, as shown in the following formula: = ⊕ ⊕R； In the formula: R represents the knowledge fusion residual feature; Embedded global dual-source knowledge; This refers to the visual global features.

7. The method for generating a renal artery ultrasound report according to claim 1, characterized in that, The process of obtaining a renal artery ultrasound diagnostic report based on the globally unified features includes: The globally unified feature is input into the decoder; Contextual features are obtained through the masked multi-head self-attention of the decoder; The context features are normalized to obtain the normalized context features. The normalized context features and the global unified features are aligned to obtain aligned features; The alignment features are subjected to nonlinear feature transformation to obtain the transformed features; After layer normalization, the transformed features are mapped to the vocabulary space through the fully connected layer of the decoder to obtain the vocabulary probability distribution. Based on the probability distribution of the vocabulary, an autoregressive process is performed to generate a renal artery ultrasound diagnostic report.

8. A renal artery ultrasound report generation system, characterized in that, include: The acquisition module is used to perform keyword matching on the clinical examination text of ultrasound cases to obtain renal artery diagnostic indicators, and to construct a diagnostic factor pool based on the renal artery diagnostic indicators. Based on the diagnostic factor pool, search prompts are generated, and general medical knowledge and clinical expert specialty knowledge of the large model are obtained. The general medical knowledge and clinical expert specialty knowledge of the large model are encoded to obtain global dual-source knowledge embedding. The clinical examination text of the ultrasound case is encoded to generate a basic embedding. Based on the basic embedding and the global dual-source knowledge embedding, knowledge-aware features are obtained. The knowledge-aware features are fused with the basic embedding to obtain knowledge fusion residual features. The processing module is used to extract features from the multi-frame ultrasound images of the renal artery to be detected to obtain visual global features, and to perform element-wise additive fusion of the visual global features, the global dual-source knowledge embedding, and the knowledge fusion residual features to obtain global unified features. The execution module is used to obtain a renal artery ultrasound diagnostic report based on the globally unified features.

9. An electronic device, characterized in that, The method includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the method described in any one of claims 1 to 7.

10. A storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the method described in any one of claims 1 to 7.

Images