A structured ultrasound report generation method based on a large language model

Abstract

A structured ultrasound report generation method based on a large language model, belonging to the field of medical image processing, firstly utilizes a multi-agent data construction framework to automatically synthesize large-scale, high-quality "verbal diagnostic clues-structured report" data pairs from unpaired ultrasound report text. Next, a template-enhanced supervised fine-tuning technique is employed to align verbal clues with institutional standard templates, using a pre-set template library to standardize the report's structure and terminology, bridging the domain knowledge gap of the model. Finally, a defect-oriented preference optimization strategy allows the model to learn from its own generated defective reports, continuously improving its diagnostic accuracy and reliability. This invention can convert fragmented diagnostic information verbally uttered by doctors during examinations into a complete, standardized, and high-quality structured clinical report in real time and automatically, significantly improving the efficiency and standardization of report generation.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and relates to a structured ultrasound report generation method based on a large language model. Background Technology

[0002] The explosive growth in ultrasound diagnostic examinations has led to a heavy clinical workload. In the current clinical workflow, physicians must manually write complete diagnostic reports while performing scanning procedures. This process is not only repetitive and time-consuming, but also prone to errors due to fatigue, and there are significant differences between operators, affecting the standardization of reports and the value of data.

[0003] Previous studies on automated report generation have primarily followed two models, but neither has been well integrated with actual clinical workflows. The first model relies on full speech recognition transcription of complete paragraphs dictated by physicians, a method that remains time-consuming and requires post-editing. The second model attempts to generate reports directly from images, but current technology cannot achieve "full-specialty" diagnostic capabilities and carries the risk of producing hallucinatory findings when dealing with complex or diverse ultrasound pathologies, thus its reliability is insufficient.

[0004] To better reflect clinical presentation, the latest approach involves converting clinicians' verbal diagnostic cues into structured, high-quality ultrasound reports. However, this approach faces two challenges: first, data scarcity, as there is currently a lack of publicly available datasets pairing verbal cues with structured reports; and second, Large Language Models (LLMs), lacking domain knowledge, may produce fabricated content or omit key descriptions when processing brief cues. Therefore, developing a method to understand concise diagnostic cues and generate high-quality, structured reports that conform to clinical standards is crucial to resolving the current challenges in ultrasound reporting workflows. Summary of the Invention

[0005] To address the issues of existing ultrasound report generation methods being disconnected from clinical practice and lacking accuracy, this invention provides a structured ultrasound report generation method based on a large language model. The process, driven by the large language model, transforms diagnostic clues dictated by physicians during examinations into standardized, high-quality ultrasound reports. This method not only solves the problems of traditional manual report writing being time-consuming, labor-intensive, error-prone, and susceptible to operator variations, but also effectively improves the quality and clinical applicability of generated ultrasound diagnostic reports.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] A structured ultrasound report generation method based on a large language model includes the following steps:

[0008] Step S1: Through a data construction framework involving multiple intelligent agents, the corpus containing only unpaired, raw ultrasound report text is synthesized into a high-quality training dataset containing paired oral diagnostic cues and corresponding structured reports.

[0009] Step S2: Based on the training dataset and the ultrasound report template obtained by hybrid retrieval, a basic large language model is subjected to supervised fine-tuning with template enhancement to obtain a primary report generation model.

[0010] Step S3: Use the primary report generation model to reason about the training data, identify and collect the defective report samples generated by it, form a preference pair with the original high-quality report samples in the dataset, and perform defect-guided preference optimization on the primary report generation model to obtain the ultrasound report generation model.

[0011] Step S4: Input the user's verbal diagnostic clues into the ultrasound report generation model to obtain a structured ultrasound report.

[0012] Furthermore, the data construction framework for the collaboration of multiple intelligent agents includes:

[0013] Clue Simulation Agent It was obtained by fine-tuning a seed dataset annotated by human experts. Its function is to receive a raw, complete ultrasound report text and extract and generate a concise diagnostic clue that mimics the speaking habits of ultrasound doctors.

[0014] Intelligent report generation By training on temporary paired data consisting of clues generated by the clue-simulating agent and the original report, its function is to expand a concise diagnostic clue into a complete, structured report containing examination findings and diagnostic impressions.

[0015] Quality assessment agent Based on a set of predefined clinical quality dimensions, the consistency and quality among the report generated by the report generation agent, the input diagnostic clues, and the original report are scored in multiple dimensions. Data pairs are then selected according to preset scoring thresholds to form the training dataset.

[0016] Preferably, the clue simulates an intelligent agent. In this study, a small collection of "diagnostic clues - structured reports" annotated by ultrasound specialists was compiled. ,in The seed dataset represents the number of "diagnostic clues - structured reports". ;

[0017] Based on seed dataset Fine-tuning the basic large language model yields a cue-simulated intelligent agent. Diagnostic clues As input, structured reports As a reference output, supervised fine-tuning is performed, with the following loss function:

[0018] ;

[0019] in, Indicates a clue-simulating intelligent agent The loss function is denoted by log, where log represents the logarithmic function. Indicated by As a clue-simulating intelligent agent The input and output are The probability, These are the parameters of the basic large language model. This indicates a summation operation.

[0020] More preferably, the quality assessment intelligent agent In China, the clinical quality dimensions include:

[0021] Accuracy ensures that the report is free of factual errors;

[0022] Completeness, ensuring coverage of all clinically relevant findings;

[0023] Relevance ensures the content has diagnostic value;

[0024] The report should be well-structured to ensure a clear and easy-to-read format.

[0025] Clarity ensures that the language is unambiguous;

[0026] Simplicity ensures that the content is concise and free of redundancy.

[0027] More preferably, in step S1, the synthesis process of the training dataset is as follows:

[0028] Simulate an intelligent agent using clues. From the original structured ultrasound report text set A set of diagnostic clues was obtained. ,in This indicates the number of ultrasound reports in the original structured ultrasound report text set:

[0029] ;

[0030] in, Represents the independent variable that makes the function reach its maximum value. Indicated by As a clue-simulating intelligent agent The input and output are c The probability,c For output;

[0031] A basic large language model is used in the "diagnostic clues - original structured report" set. Fine-tuning was performed to obtain a report that generated the intelligent agent. Next, the diagnostic clues will be collected. As a report-generating agent Input, get report set :

[0032] ;

[0033] in, Represents the independent variable that makes the function reach its maximum value. Indicated by As a report-generating agent The input and output are r The probability, r For output; diagnostic clues Original report and the generated report As a quality assessment intelligent agent The input yields the score vector. :

[0034] ;

[0035] in, Represents a quality assessment agent;

[0036] When the fraction vector There exists any component that is less than the threshold. Then remove the corresponding value of the fraction vector. The remaining Composing a high-quality training dataset ;

[0037] The report generating agent The process of obtaining diagnostic clues is as follows: As input, the original report As a reference output, supervised fine-tuning is performed, with the following loss function:

[0038] ;

[0039] in, Indicates that the report generates an intelligent agent. loss function, Indicated by As a report-generating agent The input and output are The probability, These are the parameters of the basic large language model.

[0040] Furthermore, the process of step S2 is as follows:

[0041] Training dataset based on ICD-11 encoding and semantic similarity calculation Diagnostic clues With ultrasound report template The mixed similarity is calculated using the following formula:

[0042] ;

[0043] in, It is an encoding function that maps all clinical entities in the input parameters to ICD-11 codes, resulting in an ICD-11 encoding set; It is an indicator function, when the template Diagnostic clues The Jaccard similarity of the ICD-11 encoding set is greater than the threshold. If the condition is met, return 1; otherwise, return 0. and Representing templates respectively Diagnostic clues The embedding vector; express and Cosine similarity; This represents the weight hyperparameter;

[0044] diagnostic clues With template set Calculate the mixed similarity of all templates and extract those with similarity greater than a threshold. The templates form a subset of templates. ;

[0045] Subset of templates and diagnostic clues As input, report As the standard output, the loss function is defined as follows:

[0046] ;

[0047] in, This represents a basic report generation model. loss function, Indicated by and As a report generation model The input and output are The probability, These are the parameters of the basic large language model;

[0048] The basic large language model was then subjected to supervised fine-tuning to obtain a preliminary report generation model. .

[0049] Furthermore, the process of step S3 is as follows:

[0050] training dataset As a primary report generation model The input is used to perform reasoning to obtain a new generated report set. The calculation formula is as follows:

[0051] ;

[0052] in, This indicates a basic report generation model. A subset of the template;

[0053] Will , and As input, use the quality assessment agent judge The production quality is calculated using the following formula:

[0054] ;

[0055] score vector The accuracy and integrity components are used as evaluation dimensions. When either the accuracy or integrity component falls below a threshold... The corresponding generated report Diagnostic clues and the original report Constructing a preference triplet All preference triples Constructing a preference dataset ;

[0056] The loss function is defined as follows:

[0057] ;

[0058] in, This represents the loss function of the final ultrasound report generation model. The model to be optimized As a reference model for freezing parameters, both are initialized to... , This indicates a direct preference for optimizing the temperature hyperparameter. This represents the sigmoid activation function. Represents the mathematical expectation. Represents a triplet From dataset , For diagnostic clues, For the original report, To generate a report, This indicates that when the input is At that time, the model to be optimized The output is The probability, This indicates that when the input is At that time, reference model The output is The probability, This indicates that when the input is At that time, the model to be optimized The output is The probability, This indicates that when the input is At that time, reference model The output is The probability of;

[0059] and the initial report generation model In the preference dataset The model was then further fine-tuned using direct preference optimization to obtain the final ultrasound report generation model.

[0060] The beneficial effects of this invention are mainly reflected in the following aspects: Firstly, by introducing a data construction framework involving multiple intelligent agents, a high-quality "verbal cues-structured report" corpus is autonomously synthesized from a large number of unpaired ultrasound reports, thus solving the long-standing bottleneck problem of scarce training data in this field. Secondly, by designing a two-stage learning strategy—template-enhanced supervised fine-tuning and preference optimization based on self-defects—the quality and reliability of report generation are significantly improved. Furthermore, template-enhanced supervised fine-tuning, by introducing ultrasound report templates as contextual references, ensures the standardization of generated reports and effectively compensates for the model's lack of clinical knowledge. Simultaneously, the preference optimization stage utilizes the model's own generated defective reports as negative preferences, forcing the model to learn from errors, thereby suppressing content illusions and enhancing diagnostic accuracy. This application effectively solves the problems of low efficiency, error susceptibility, and poor consistency in traditional manual reporting processes, providing the ultrasound diagnostic field with an efficient, accurate, and highly standardized report generation solution that can seamlessly integrate into clinical workflows. Attached Figure Description

[0061] Figure 1 This is a flowchart of the structured ultrasound report generation method based on a large language model, as described in this application.

[0062] Figure 2A schematic diagram of the framework structure for constructing multi-agent data in this application.

[0063] Figure 3 This is a schematic diagram of the oversight and fine-tuning process for enhancing the template of this application.

[0064] Figure 4 This is a schematic diagram of the preference optimization process guided by the defects of this application. Detailed Implementation

[0065] The present invention will now be further described with reference to the accompanying drawings.

[0066] Reference Figures 1-4 A structured ultrasound report generation method based on a large language model includes:

[0067] Step S1: Through a data construction framework involving multiple agents, the corpus containing only unpaired, raw ultrasound report text is synthesized into a high-quality training dataset containing paired oral diagnostic cues and corresponding structured reports.

[0068] In a specific embodiment, such as Figure 2 As shown, the data construction framework for the collaboration of multiple intelligent agents includes:

[0069] Clue Simulation Agent The agent was fine-tuned on a seed dataset annotated by ultrasound experts. Its function is to receive a raw, complete ultrasound report text and extract and generate a concise diagnostic clue that mimics the speaking habits of ultrasound doctors.

[0070] Report generating intelligent agent The agent is trained on temporary paired data consisting of clues generated by the clue simulation agent and the original report. Its function is to expand a concise diagnostic clue into a complete, structured report containing examination findings and diagnostic impressions.

[0071] Quality assessment agent Its function is to perform multi-dimensional scoring on the consistency and quality among the report generated by the report generation agent, the input diagnostic clues, and the original report based on a set of predefined clinical quality dimensions, and to filter data pairs according to preset scoring thresholds to finally form the training dataset.

[0072] In one specific embodiment, the clue-simulating agent The agent was fine-tuned on a seed dataset annotated by ultrasound experts, including:

[0073] Collect a small collection of "diagnostic clues - structured reports" annotated by ultrasound specialists. ,in The number of "diagnostic clues - structured reports" represents the number of entries in the seed dataset. .

[0074] Based on seed dataset Fine-tuning the basic large language model yields a cue-simulated intelligent agent. Diagnostic clues As input, structured reports As a reference output, supervised fine-tuning is performed, with the following loss function:

[0075] ;

[0076] in, Indicates a clue-simulating intelligent agent The loss function is denoted by log, where log represents the logarithmic function. Indicated by As a clue-simulating intelligent agent The input and output are The probability, These are the parameters of the basic large language model. This indicates a summation operation.

[0077] In one specific embodiment, the predefined clinical quality dimensions include:

[0078] Accuracy ensures that the report is free of factual errors;

[0079] Completeness, ensuring coverage of all clinically relevant findings;

[0080] Relevance ensures the content has diagnostic value;

[0081] The report should be well-structured to ensure a clear and easy-to-read format.

[0082] Clarity ensures that the language is unambiguous;

[0083] Simplicity ensures that the content is concise and free of redundancy.

[0084] In one specific embodiment, the process of synthesizing a high-quality training dataset containing paired verbal diagnostic cues and corresponding structured reports from corpus consisting only of unpaired, raw structured ultrasound report text includes:

[0085] Step 1.1: Simulate an intelligent agent using cues From the original structured ultrasound report text set A set of diagnostic clues was obtained. ,in This indicates the number of ultrasound reports in the original structured ultrasound report text set:

[0086] ;

[0087] in, Represents the independent variable that makes the function reach its maximum value. Indicated by As a clue-simulating intelligent agent The input and output are c The probability, c For output;

[0088] Step 1.2: Apply a basic large language model to the "diagnostic clues - original structured report" set. Fine-tuning was performed to obtain a report that generated the intelligent agent. Next, the diagnostic clues will be collected. As a report-generating agent Input, get report set :

[0089] ;

[0090] in, Represents the independent variable that makes the function reach its maximum value. Indicated by As a report-generating agent The input and output are r The probability, r For output;

[0091] Step 1.3: Diagnostic clues Original report and the generated report As a quality assessment intelligent agent The input yields the score vector. :

[0092] ;

[0093] in, Represents a quality assessment agent;

[0094] Step 1.4, when the fraction vector There exists any component that is less than the threshold. Then remove the corresponding value of the fraction vector. The remaining Composing a high-quality training dataset .

[0095] In one specific embodiment, a basic large language model is used in the "diagnostic clues-original report" set. Fine-tuning was performed to obtain a report that generated the intelligent agent. ,include:

[0096] diagnostic clues As input, the original report As a reference output, supervised fine-tuning is performed, with the following loss function:

[0097] ;

[0098] in, Indicates that the report generates an intelligent agent. loss function, Indicated by As a report-generating agent The input and output are The probability, These are the parameters of the basic large language model;

[0099] Step S2: Based on the training dataset and the ultrasound report templates obtained from hybrid retrieval, a basic large language model is subjected to supervised fine-tuning with template enhancement to obtain a primary report generation model.

[0100] In one specific embodiment, such as Figure 3 As shown, based on the training dataset and the retrieved ultrasound report templates, a basic large language model is subjected to supervised fine-tuning with template enhancement to obtain a preliminary report generation model, including:

[0101] Step 2.1: Calculate the training dataset based on ICD-11 encoding and semantic similarity. Diagnostic clues With ultrasound report template The mixed similarity is calculated using the following formula:

[0102] ;

[0103] in It is an encoding function that maps all clinical entities in the input parameters to ICD-11 codes, resulting in an ICD-11 encoding set; It is an indicator function, when the template Diagnostic clues The Jaccard similarity of the ICD-11 encoding set is greater than the threshold. If the condition is met, return 1; otherwise, return 0. and Representing templates respectively Diagnostic clues The embedding vector; express and Cosine similarity; This represents the weight hyperparameter;

[0104] Step 2.2: Diagnostic clues With template set Calculate the mixed similarity of all templates and extract those with similarity greater than a threshold. The templates form a subset of templates. ;

[0105] Step 2.3: Subset the template and diagnostic clues As input, report As the standard output, the loss function is defined as follows:

[0106] ;

[0107] in, This represents a basic report generation model. loss function, Indicated by and As a report generation model The input and output are The probability, These are the parameters of the basic large language model;

[0108] The basic large language model was then subjected to supervised fine-tuning to obtain a preliminary report generation model. .

[0109] Step S3: Use the primary report generation model to infer the training data, identify and collect the defective report samples generated by it, form a preference pair with the original high-quality report samples in the dataset, and perform defect-guided preference optimization on the primary report generation model to obtain the final ultrasound report generation model.

[0110] In one specific embodiment, such as Figure 4 As shown, the primary report generation model is used to infer from the training data, identify and collect defective report samples generated by it, and form preference pairs between the defective samples and the original high-quality report samples in the dataset. The primary report generation model is then optimized using defect-guided preferences to obtain the final ultrasound report generation model, including:

[0111] Step 3.1: Transfer the training dataset As a basic report generation model The input is used to perform reasoning to obtain a new generated report set. The calculation formula is as follows:

[0112] ;

[0113] in, This indicates a basic report generation model. A subset of the template;

[0114] Step 3.2, , and As input, use the quality assessment agent judge The production quality is calculated using the following formula:

[0115] ;

[0116] Step 3.3: Convert the scoring vector The accuracy and integrity components are used as evaluation dimensions. When either the accuracy or integrity component falls below a threshold... The corresponding generated report Diagnostic clues and the original report Constructing a preference triplet All preference triples Constructing a preference dataset ;

[0117] Step 3.4: Define the loss function as follows:

[0118] ;

[0119] in, This represents the loss function of the final ultrasound report generation model. The model to be optimized As a reference model for freezing parameters, both are initialized to... , This indicates a direct preference for optimizing the temperature hyperparameter. This represents the sigmoid activation function. Represents the mathematical expectation. Represents a triplet From dataset , For diagnostic clues, For the original report, To generate a report, This indicates that when the input is At that time, the model to be optimized The output is The probability, This indicates that when the input is At that time, reference model The output is The probability, This indicates that when the input is At that time, the model to be optimized The output is The probability, This indicates that when the input is At that time, reference model The output is The probability of;

[0120] and the initial report generation model In the preference dataset The model was further fine-tuned using direct preference optimization to obtain the final ultrasound report generation model.

[0121] Step S4: Input the user's verbal diagnostic clues into the final report generation model to obtain a structured ultrasound report.

[0122] Specifically, the ultrasound report generation model obtained after the above training can generate corresponding structured ultrasound report text based on the input verbal diagnostic clues, and ensure the standardization and high accuracy of the generated ultrasound report.

[0123] This embodiment uses the large language model Qwen3-8B as the basic large language model, and performs fine-tuning and optimization on the basis of the basic large language model to obtain the ultrasound report generation model of the present invention.

[0124] The structured ultrasound report in this embodiment is a standardized and itemized summary of the ultrasound examination results of the examined site. The report includes at least two parts: "Ultrasound Findings" and "Diagnostic Opinion." "Ultrasound Findings" records objective imaging information obtained by the physician based on ultrasound images, such as the location, size, shape, boundary echo characteristics, and blood flow signals of lesions. "Diagnostic Opinion" is a conclusive opinion given by the physician based on the "Ultrasound Findings," combined with clinical data and relevant treatment guidelines, after a comprehensive analysis of the possible nature, severity, and differential diagnosis of the disease. It may also provide further suggestions for follow-up or additional examinations. This structured design makes the expression of ultrasound examination results in the report more standardized and consistent, facilitating subsequent retrieval, statistical analysis, and processing.

[0125] This invention addresses the long-standing bottleneck of scarce training data in the field by autonomously synthesizing a high-quality "verbal cue-structured report" corpus from a large number of unpaired ultrasound reports using a multi-agent data construction framework. Template-enhanced supervised fine-tuning, by introducing ultrasound report templates as contextual references, ensures the standardization of generated reports and effectively compensates for the model's lack of clinical knowledge. Furthermore, preference optimization utilizes the model's own generated flawed reports as negative preferences, forcing the model to learn from errors, thereby suppressing content illusion and enhancing diagnostic accuracy. Through the comprehensive application of these techniques, this application significantly improves the efficiency of ultrasound examinations, reduces the burden on physicians, and effectively solves the problems of inefficiency, error-proneness, and poor consistency in traditional manual reporting processes.

[0126] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms stated in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A structured ultrasound report generation method based on a large language model, characterized in that, The method includes the following steps: Step S1: Using a data construction framework involving multiple agents, the corpus containing only unpaired, raw ultrasound report text is synthesized into a training dataset containing paired verbal diagnostic cues and corresponding structured reports; the data construction framework involving multiple agents includes: Clue Simulation Agent It was obtained by fine-tuning on a labeled seed dataset. Its function is to receive a raw ultrasound report text and extract and generate a diagnostic clue from it. Report generating intelligent agent By training on temporary paired data consisting of clues generated by the clue simulation agent and the original report, its function is to expand a diagnostic clue into a structured report. Quality assessment agent Based on the clinical quality dimension, the consistency and quality among the report generated by the report generation agent, the input diagnostic clues, and the original report are scored in multiple dimensions, and data pairs are selected according to the preset scoring threshold to form the training dataset. Step S2: Based on the training dataset and the ultrasound report template obtained by hybrid retrieval, a basic large language model is subjected to supervised fine-tuning with template enhancement to obtain a primary report generation model; Step S3: Use the primary report generation model to reason about the training data, identify and collect the defective report samples generated by it, form a preference pair with the original high-quality report samples in the dataset, and perform defect-guided preference optimization on the primary report generation model to obtain the ultrasound report generation model. Step S4: Input the user's verbal diagnostic clues into the ultrasound report generation model to obtain a structured ultrasound report.

2. The structured ultrasound report generation method based on a large language model as described in claim 1, characterized in that, The clue simulates an intelligent agent In the middle, a small number of labeled "diagnostic clues-structured reports" sets were collected. ,in The seed dataset represents the number of "diagnostic clues - structured reports". ; Based on seed dataset Fine-tuning the basic large language model yields a cue-simulated intelligent agent. Diagnostic clues As input, structured reports As a reference output, supervised fine-tuning is performed and a loss function is defined.

3. The structured ultrasound report generation method based on a large language model as described in claim 1, characterized in that, The quality assessment intelligent agent In this context, clinical quality dimensions include accuracy, completeness, relevance, structure, clarity, and conciseness.

4. The structured ultrasound report generation method based on a large language model as described in claim 1, characterized in that, In step S1, the process of synthesizing the training dataset is as follows: Simulate an intelligent agent using clues. From the original structured ultrasound report text set A set of diagnostic clues was obtained. ,in This indicates the number of ultrasound reports in the original structured ultrasound report text set; A basic large language model is used in the "diagnostic clues - original structured report" set. Fine-tuning was performed to obtain a report that generated the intelligent agent. Next, the diagnostic clues will be collected. As a report-generating agent Input, get report set ; Diagnostic clues Original report and the generated report As a quality assessment intelligent agent The input yields the score vector. ; When the fraction vector There exists any component that is less than the threshold. Then remove the corresponding value of the fraction vector. The remaining Composing a high-quality training dataset .

5. The structured ultrasound report generation method based on a large language model as described in claim 1, characterized in that, The report generating agent The process of obtaining diagnostic clues is as follows: As input, the original report As a reference output, supervised fine-tuning is performed and a loss function is defined.

6. A structured ultrasound report generation method based on a large language model as described in any one of claims 1 to 5, characterized in that, The process of step S2 is as follows: Training dataset based on ICD-11 encoding and semantic similarity calculation Diagnostic clues With ultrasound report template Mixed similarity; Diagnostic clues With template set Calculate the mixed similarity of all templates and extract those with a similarity greater than a threshold. The templates form a subset of templates. ; Subset of templates and diagnostic clues As input, report As the standard output, a loss function is defined, and the basic large language model is fine-tuned under supervision to obtain a preliminary report generation model. .

7. A structured ultrasound report generation method based on a large language model as described in any one of claims 1 to 5, characterized in that, The process of step S3 is as follows: training dataset As a primary report generation model The input is used to perform reasoning to obtain a new generated report set. ; Will , and As input, use the quality assessment agent judge The quality of the generated product; score vector The accuracy and integrity components are used as evaluation dimensions. When either the accuracy or integrity component falls below a threshold... The corresponding generated report Diagnostic clues and the original report Constructing a preference triplet All preference triples Constructing a preference dataset ; Define the loss function and implement the initial report generation model. In the preference dataset The model was then further fine-tuned using direct preference optimization to obtain the final ultrasound report generation model.

Images