Medical image question interaction method and system based on pre-labeled standard answers

Abstract

The application discloses a medical image question interaction method and system based on pre-labeled standard answers, and the method comprises the following steps: receiving a user's reading answer text for a medical image recognition question and performing text preprocessing; according to a key knowledge point list of the pre-labeled answer, performing keyword matching on the reading answer text to obtain a keyword score; inputting the reading answer text and the pre-labeled answer into a medical text embedding model to calculate a semantic similarity score; performing keyword matching on the reading answer text and a common error mode list to obtain a common error deduction score; obtaining a comprehensive score of the answer according to the keyword score, the semantic similarity score and the common error deduction score; judging whether the reading answer text is correct according to the comprehensive score, if not, then returning a hierarchical prompt text to the user in a progressive manner, and receiving a re-input reading answer text. The application does not completely rely on AI real-time image recognition, and effectively reduces the evaluation distortion risk of AI reading and clinical standards.

Description

Technical Field

[0001] This invention relates to the field of information interaction technology, and in particular to a method and system for interactive medical imaging questions based on pre-annotated standard answers. Background Technology

[0002] Existing interactive solutions for medical image recognition questions mainly follow two technical approaches:

[0003] The first approach is based on AI image recognition. This approach uses a deep learning model to automatically analyze and identify features in medical images (such as ultrasound images). The AI ​​model generates interpretation results, which are then used as the standard answer for users to compare. The shortcomings of this approach are: First, training the AI ​​image recognition model requires tens of thousands of professionally labeled image data and substantial GPU computing resources, resulting in high construction and deployment costs. Second, the AI ​​model's interpretation standards and logic may deviate from the standards actually used by clinical experts, leading to a risk of inconsistencies between AI interpretation results and clinical standards. This inconsistency can directly distort assessment conclusions in teaching and evaluation scenarios. Third, when new image types or anatomical structure recognition requirements are needed, the model must be retrained or fine-tuned, with deployment cycles typically lasting several weeks, making it difficult to quickly respond to changes in the question bank content. Fourth, the deep learning model exhibits differences in recognition performance for images acquired by different brands of ultrasound equipment and different operator techniques, affecting the consistency and fairness of the assessment.

[0004] The second approach is the traditional static question bank solution. This solution packages medical images with pre-set multiple-choice options into fixed questions. Users select answers from pre-set options (A, B, C, D, etc.), and the system judges whether the selected option matches the pre-set correct option. The shortcomings of this solution are: First, users can use elimination, guessing, or semantic analysis of the option text to find the correct answer. The system can only determine whether the user selected the correct option, but cannot verify whether the user truly possesses the ability to independently observe images and form interpretations. Second, the standard answer is clearly exposed to the user in the form of options, allowing the user to deduce the interpretation logic from the options rather than making judgments based on their own understanding of the images. Third, it lacks a tiered prompting mechanism; after a user answers incorrectly, the system directly provides the complete answer without a gradual guidance or progressively decreasing difficulty. Fourth, statistical analysis only records correct and incorrect answers, lacking the ability to summarize and analyze user answer data by classification dimensions (such as by disease type, anatomical structure, or difficulty level), and cannot identify the user's weaknesses in specific knowledge domains.

[0005] Therefore, there is an urgent need to invent a new interactive method for medical image questions based on pre-annotated standard answers, in order to solve the problems of existing technologies such as reliance on AI image recognition, limited interactive question bank formats, and inability to locate users' weak points according to knowledge domain dimensions. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method and system for interactive medical imaging questions based on pre-annotated standard answers, which at least partially solves the problems existing in the prior art.

[0007] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0009] According to a first aspect of the present invention, a method for interactive medical imaging questions based on pre-annotated standard answers is provided, the method comprising:

[0010] Receive the user's interpretation and answer text for medical image recognition questions, along with the question's unique identifier;

[0011] The judgment response text is preprocessed to obtain the preprocessed judgment response text;

[0012] Based on the list of key knowledge points of the pre-annotated answer corresponding to the unique identifier of the question, keyword matching is performed on the preprocessed interpretation and answer text to obtain a keyword matching score.

[0013] The interpreted response text and the pre-labeled answer are input into a pre-trained medical text embedding model to calculate a semantic similarity score;

[0014] The preprocessed interpretation and answer text is matched with the list of common error patterns in the pre-labeled answers using keywords to obtain the common error deductions;

[0015] The overall answer score is obtained based on the keyword matching score, the semantic similarity score, and the common error deductions.

[0016] The system determines whether the interpreted answer text is correct based on the overall score of the answer. If the interpreted answer text is incorrect, it progressively returns hierarchical prompt text to the user and receives the interpreted answer text input by the user based on the hierarchical prompt text.

[0017] Furthermore, it receives the user's interpretation and answer text for medical image recognition questions, along with the question's unique identifier, including:

[0018] After receiving a question request from a user on the client side, the server sends the medical image recognition question data to the interception layer.

[0019] In the interception layer, the medical image recognition question data is compared with a preset whitelist, preset fields are removed, and the medical image recognition question data that has passed the interception layer is sent to the client. The preset fields include standard answer text, key knowledge point list, common error pattern list and prompt text.

[0020] It receives the user's interpretation of medical image recognition questions and the question's unique identifier from the client.

[0021] Further, the interpretation response text is preprocessed to obtain a preprocessed interpretation response text, including:

[0022] The interpretation and response text is cleaned by removing punctuation marks and common stop words to obtain pure content text;

[0023] The plain text is segmented by calling a word segmenter loaded with a medical dictionary to obtain segmented text;

[0024] The system identifies negative expressions and paraphrasing prompts in the segmented text, marks text segments containing negative expressions with negative context, and marks text segments containing paraphrasing prompts with paraphrasing prompts.

[0025] Furthermore, based on the list of key knowledge points in the pre-annotated answer corresponding to the unique identifier of the question, keyword matching is performed on the preprocessed interpretation and answer text to obtain a keyword matching score, including:

[0026] For each key knowledge point object in the list of key knowledge points in the pre-annotated answer corresponding to the unique identifier of the question, based on the preset hierarchical matching rules, the key knowledge point objects are used to perform hierarchical matching in the pre-processed interpretation and answer text.

[0027] If all key keywords in the key knowledge point object match the preprocessed interpretation and answer text, then the corresponding key keywords are defined as a complete match, and the matching weight of the current key knowledge point object is assigned the first preset matching weight. The key keywords include core key keywords and bonus key keywords. The core key keywords correspond to the core type weight, and the bonus key keywords correspond to the bonus type weight.

[0028] If at least one key keyword in the key knowledge point object matches the preprocessed interpretation and response text, then the corresponding key keyword is defined as a partial match, and the matching weight of the current key knowledge point object is assigned the second preset matching weight.

[0029] If the key keywords in the key knowledge point object do not appear in the preprocessed judgment and answer text, the corresponding key keywords are defined as unmatched, and the matching weight of the current key knowledge point object is assigned to the third preset matching weight, wherein the first preset matching weight is greater than the second preset matching weight, and the second preset matching weight is greater than the third preset matching weight.

[0030] Based on the matching weight and type weight of the key keywords in each key knowledge point object, the comprehensive keyword matching score of the interpretation and response text is obtained.

[0031] Furthermore, the method also includes:

[0032] Based on the paraphrase guidance mark, determine whether the key keywords that are successfully matched in the key knowledge point object only appear in the paraphrase or retelling text in the interpretation and answer text, and the key keywords that are successfully matched are key keywords that are completely matched or partially matched;

[0033] If the key knowledge point object that is successfully matched only appears in the paraphrased or retelled text in the interpretation and answer text, then the key knowledge point keyword is downgraded to unmatched;

[0034] If the key knowledge point object that is successfully matched appears not only in the paraphrased or retelled text in the interpretation and answer text, then the key knowledge point keyword will not be downgraded to unmatched.

[0035] Based on the negative context marker, it is determined whether the key keywords that are successfully matched in the key knowledge point object appear within the scope of the negative context marker. The scope of the negative context marker is from the beginning of the negative expression word in the clause to the end of the clause.

[0036] If a key keyword that is successfully matched in the key knowledge point object appears within the scope of the negative context marker, then the key keyword is downgraded to unmatched.

[0037] If the key knowledge point object that is successfully matched does not appear in the scope of the negative context marker, then the key knowledge point keyword will not be downgraded to unmatched;

[0038] Determine whether the character length of high-density sentences in the response text is less than a preset length threshold. The high-density sentences are those in which the number of successfully matched key words is greater than or equal to the preset key word number threshold.

[0039] If the character length of a high-density sentence in the response text is less than a preset length threshold, then the exact matching keywords in the high-density sentence will be downgraded to partial matching, while the partial matching keywords in the high-density sentence will remain unchanged.

[0040] If the character length of a high-density sentence in the response text is greater than or equal to a preset length threshold, then the key keywords in that high-density sentence will not be downgraded.

[0041] Further, the interpreted response text and the pre-labeled answer are input into a pre-trained medical text embedding model to calculate a semantic similarity score, including:

[0042] The interpretation response text and the pre-labeled answer are input into the same pre-trained medical text embedding model. The cosine similarity between the semantic vectors corresponding to the interpretation response text and the pre-labeled answer input is calculated. The medical text embedding model is a medical-specific text embedding model obtained by domain fine-tuning with a large number of Chinese medical literature and clinical texts on the basis of a general Chinese pre-trained language model.

[0043] If the keyword matching score is within the preset passing range, the semantic weight will be assigned a preset high semantic weight value.

[0044] If the keyword matching score is not within the preset passing threshold range, the semantic weight will be assigned a preset low semantic weight value.

[0045] The semantic similarity score is obtained based on the semantic weight and the cosine similarity.

[0046] Furthermore, the preprocessed interpretation response text is matched with the list of common error patterns in the pre-annotated answers using keywords to obtain common error deductions, including:

[0047] For each common error pattern object in the list of common error patterns of the pre-annotated answer corresponding to the unique identifier of the question, the common error keywords in the common error pattern object are matched in the preprocessed interpretation and answer text, and the number of matched error keywords is counted.

[0048] Calculate the hit rate of incorrect keywords based on the number of incorrect keywords hit;

[0049] Determine whether the hit rate of the erroneous keywords is greater than or equal to a preset hit rate threshold;

[0050] If the hit rate of the erroneous keyword is greater than or equal to the preset hit rate threshold, then the common error pattern object corresponding to the hit rate of the erroneous keyword is hit;

[0051] If the hit rate of the erroneous keyword is less than the preset hit rate threshold, then the common error pattern object corresponding to the hit rate of the erroneous keyword is not hit;

[0052] The common error deduction is obtained based on the number of common error pattern objects hit and the preset deduction coefficient.

[0053] Furthermore, judging whether the interpreted response text is correct based on the overall score of the response includes:

[0054] If the overall score of the answer is greater than or equal to the preset correctness threshold, and all the current core keywords are completely matched, and no common error pattern objects are hit, then the user's answer text for the medical image recognition question will be judged as correct.

[0055] If the overall score of the answer is less than the preset correctness threshold, or if the current core keyword is not a complete match, or if a common error pattern is hit, then the user's answer to the medical image recognition question will be judged as incorrect.

[0056] Furthermore, if the interpreted response text is incorrect, a progressively layered prompt text is returned to the user, and the interpreted response text input by the user based on the layered prompt text is received, including:

[0057] For interpretations of the response text that are deemed incorrect, receive prompt requests sent by the user;

[0058] Determine whether the level parameter of the current prompt request is within the preset valid level range;

[0059] If the level parameter of the current prompt request is not within the preset valid level range, the prompt request will be rejected.

[0060] If the level parameter of the current prompt request is within the preset valid level range, then determine whether all the preceding requests of the current prompt request have been completed;

[0061] If not all the prerequisite requests for the current prompt request have been completed, then the prompt request is rejected;

[0062] If all prerequisite requests for the current prompt request have been completed, then determine whether there is a prompt text corresponding to the current prompt request level in the prompt text array that uniquely identifies the question.

[0063] If the prompt text array that uniquely identifies the question does not contain the prompt text corresponding to the current prompt request level, then the prompt request is rejected;

[0064] If the prompt text array that uniquely identifies the question contains a prompt text corresponding to the current prompt request level, then the prompt text corresponding to the current prompt request level will be output to the client.

[0065] According to a second aspect of the present invention, a medical imaging question interaction system based on pre-annotated standard answers is provided, the system comprising:

[0066] The question bank storage module is used to store medical image recognition questions and pre-annotated answers;

[0067] The data receiving module is used to receive the user's interpretation and answer text for medical image recognition questions and the unique identifier of the question;

[0068] The semantic comparison engine module includes a text preprocessing unit, a keyword matching unit, a semantic similarity calculation unit, a common error detection unit, and a comprehensive scoring unit;

[0069] The text preprocessing unit is used to preprocess the judgment response text to obtain the preprocessed judgment response text.

[0070] The keyword matching unit is used to perform keyword matching on the preprocessed interpretation and answer text based on the list of key knowledge points of the pre-annotated answer corresponding to the unique identifier of the question, and to obtain a keyword matching score.

[0071] The semantic similarity calculation unit is used to input the interpreted answer text and the pre-labeled answer into a pre-trained medical text embedding model to calculate the semantic similarity score;

[0072] The common error detection unit is used to perform keyword matching between the preprocessed interpretation and answer text and the list of common error patterns of the pre-labeled answer to obtain common error deductions;

[0073] The comprehensive scoring unit is used to obtain a comprehensive answer score based on the keyword matching score, the semantic similarity score, and the common error deductions, and to determine whether the interpreted answer text is correct based on the comprehensive answer score.

[0074] The layered prompting module is used to receive prompt requests from users for incorrect interpretation of the response text, progressively return layered prompt text to the user, and receive the interpretation response text entered by the user based on the layered prompt text.

[0075] This invention provides a method and system for interactive medical image questions based on pre-annotated standard answers. The method includes: receiving and preprocessing a user's interpretation of a medical image recognition question; performing keyword matching on the interpretation of the interpretation of the question based on a list of key knowledge points in the pre-annotated answer to obtain a keyword matching score; inputting the interpretation of the question and the pre-annotated answer into a medical text embedding model to calculate a semantic similarity score; performing keyword matching on the interpretation of the question and a list of common error patterns to obtain common error deductions; obtaining a comprehensive answer score based on the keyword matching score, semantic similarity score, and common error deductions; determining whether the interpretation of the question is correct based on the comprehensive score; if incorrect, progressively returning hierarchical prompts to the user and receiving a re-entered interpretation of the question. This invention does not rely entirely on real-time AI image recognition, effectively reducing the risk of evaluation distortion due to inconsistencies between AI interpretation and clinical standards. Attached Figure Description

[0076] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0077] Figure 1 A flowchart illustrating the interactive method for medical imaging questions based on pre-annotated standard answers provided in an embodiment of the present invention;

[0078] Figure 2 This is a schematic diagram illustrating the principle of the layered prompting process provided in an embodiment of the present invention;

[0079] Figure 3 A schematic diagram illustrating the two-dimensional difficulty control principle provided in this embodiment of the invention;

[0080] Figure 4 This is a schematic diagram of the architecture of the client, various interfaces, and internal components of the server provided in an embodiment of the present invention. Detailed Implementation

[0081] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0082] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0083] Figure 1 A flowchart of a medical imaging question interaction method based on pre-annotated standard answers according to an embodiment of the present invention is shown.

[0084] like Figure 1 As shown, the medical image question interaction method based on pre-annotated standard answers according to an embodiment of the present invention may include steps S100, S200, S300, S400, S500, S600 and S700.

[0085] In step S100, the system receives the user's interpretation of a medical image recognition question and the question's unique identifier.

[0086] Specifically, the above steps include:

[0087] Upon receiving a question request from a user on the client side, the server sends the medical image recognition question data to the interception layer. At the interception layer, the medical image recognition question data is compared with a preset whitelist (containing fields such as id, category, subcategory, difficulty, image_url, image_type, and question_text). Preset fields (standard_answer, key_points, common_mistakes, and hints) are removed. The medical image recognition question data that passes through the interception layer is then sent to the client. The aforementioned preset fields include the standard answer text, a list of key knowledge points, a list of common error patterns, and hint text.

[0088] This invention ensures that sensitive fields do not appear in the response data of the question-generating interface at the transmission layer by removing non-whitelisted fields before serialization. The excluded fields can only be returned on demand and in a controlled manner through the answer comparison interface and the hierarchical prompt interface. That is, users must first submit their own interpretations before they can indirectly understand the content coverage of the standard answer through the comparison results.

[0089] It receives the user's interpretation of medical image recognition questions and the question's unique identifier from the client.

[0090] The aforementioned server runs as an independent service process, exposing five tool interfaces through a standardized tool call protocol. Client systems—including but not limited to AI dialogue agents, web front-end applications, and mobile applications—use the question bank service by calling these tool interfaces. Clients do not need to directly access the question bank metadata database, nor do they have direct read permissions for the question bank metadata. All five tool interfaces (category query interface, question retrieval interface, answer comparison interface, hierarchical hint interface, and batch statistics interface) adopt a request-response communication model, and the data exchange format is a structured key-value pair document format (such as JSON). Each interface follows a stateless design principle—the server does not maintain client session state; each request carries all the parameter information required to complete the request, and requests are independent of each other. This stateless design allows the server to be horizontally scaled, with multiple server instances serving different client requests simultaneously without needing to synchronize session state between instances.

[0091] Figure 4 The diagram illustrates the architecture of the client, interfaces, and internal server components provided in this embodiment of the invention.

[0092] The following is the complete metadata (presented in JSON document format) for a question belonging to the "BLUE Protocol" category, the A-line recognition subcategory, and the basic difficulty level. The question's unique identifier is "BLUE_001," indicating that it is the first question under the BLUE Protocol category. The primary category is "BLUE Protocol," and the secondary subcategory is "A-line recognition." The difficulty level is "basic." The image file path is "images / BLUE / A_line_normal.png," pointing to a static ultrasound image on the storage medium showing a normal A-line appearance at the right upper lung point. The image type is "static." The question text is: "This is an ultrasound image of the right upper lung point. Please describe what you see and provide your interpretation."

[0093] The standard answer text is written by a domain expert, with the content as follows: Multiple horizontal hyperechoic lines parallel to the pleural line and arranged at equal intervals are visible below the pleural line, namely the A-lines. The A-lines are reverberation artifacts rather than anatomical structures. Their clinical significance is that the subpleural region is gas-containing lung tissue, indicating that there is no interstitial syndrome in this region. In the BLUE protocol, bilateral A-lines plus the presence of lung sliding rule out pulmonary edema, and the next step is to examine the deep veins to rule out pulmonary embolism.

[0094] Next, in step S200, text preprocessing is performed on the interpretation response text to obtain a preprocessed interpretation response text.

[0095] Specifically, the above steps include:

[0096] Basic cleaning processing is performed on the interpretation response text to remove punctuation marks and general stop words (such as "de", "le", "zai", etc.), retaining only words with actual semantic content to obtain a pure content text.

[0097] A word segmenter loaded with a medical professional dictionary is called to perform word segmentation processing on the pure content text to obtain a segmented text. Based on the general Chinese word segmentation dictionary, the aforementioned word segmenter is additionally loaded with a medical term dictionary, an ultrasound imaging term dictionary and an anatomical term dictionary, which can correctly identify and segment medical compound terms such as "pleural line", "inferior vena cava", "interstitial syndrome", "tricuspid annular plane systolic excursion", etc., and avoid incorrectly segmenting professional terms into fragmented words.

[0098] Negative expressions (such as "not", "no", "not at all", "not found", "does not indicate", "can rule out", etc.) and paraphrase guiding phrases in the segmented text are identified, negative context marking is performed on text fragments containing negative expressions, and paraphrase guiding marking is performed on text fragments containing paraphrase guiding phrases, so as to provide negative context position information for the subsequent false positive exclusion step.

[0099] In step S300, keyword matching is performed on the preprocessed interpretation response text according to the key knowledge point list of the pre-annotated answer corresponding to the unique question identifier, so as to obtain a keyword matching score.

[0100] Specifically, the above steps include:

[0101] For each key knowledge point object in the pre-annotated answer key list (key_points) corresponding to the unique identifier of the question, a hierarchical matching is performed on the pre-processed interpretation answer text based on preset hierarchical matching rules. For example, the key knowledge point list (key_points) contains four key point objects: Key point 1: Key point text is "Identify the appearance of A-line, parallel to the pleural line, isocentric horizontal hyperechoic line", weight type is required, matching keywords include "A-line", "parallel", "pleural line", "isocentric", "horizontal", "hyperechoic", "reverberation artifact"; Key point 2: Key point text is "Clarify that A-line is a reverberation artifact rather than an anatomical structure", weight type is required, matching keywords include "artifact". Examples include "reverberation," "not anatomy," and "not a structure"; Point 3: The key text is "clinical significance, subpleural aerated lung tissue," with a weight type of required, and matching keywords including "aerated lung," "normal lung," "exclude pulmonary edema," and "exclude interstitial lung"; Point 4: The key text is "BLUE decision, bilateral A-line + lung sliding → exclude pulmonary edema, check deep veins," with a weight type of bonus, and matching keywords including "BLUE," "deep veins," "pulmonary embolism," "exclude pulmonary edema," and "next step."

[0102] If all key keywords in the key knowledge point object match the preprocessed interpretation and response text (keywords are not required to appear in the exact same order or consecutively, only that they all exist), then the corresponding key keyword is defined as a complete match, and the matching weight of the current key knowledge point object is assigned the first preset matching weight (preferably 1.0), indicating that the user's expression of the knowledge point is complete and accurate.

[0103] The above key keywords include core key keywords and bonus key keywords. Core key keywords correspond to core type weights, and bonus key keywords correspond to bonus type weights.

[0104] If at least one key keyword in the key knowledge point object matches the preprocessed interpretation and response text, then the corresponding key keyword is defined as a partial match, and the matching weight of the current key knowledge point object is assigned to the second preset matching weight (preferably 0.5), indicating that the user has touched on the knowledge point but the expression is not complete or accurate enough.

[0105] If the key keywords in the key knowledge point object do not appear in the preprocessed interpretation and response text, the corresponding key keywords are defined as unmatched, indicating that the user did not mention the knowledge point at all. The matching weight of the current key knowledge point object is assigned to the third preset matching weight (preferably 0), where the first preset matching weight is greater than the second preset matching weight, and the second preset matching weight is greater than the third preset matching weight.

[0106] Based on the matching weight and type weight of the key keywords in each key knowledge point object, the comprehensive keyword matching score of the interpretation and response text is obtained.

[0107] For key knowledge points marked as complete or partial matches, perform secondary filtering using the following three exclusion rules to correct false match judgments to the correct match level.

[0108] Rule A: Based on the paraphrase guide marker, determine whether the successfully matched key keywords in the key knowledge point object only appear in the paraphrased or relayed text of the interpretation answer; if the successfully matched key keywords in the key knowledge point object only appear in the paraphrased or relayed text of the interpretation answer, then the key keyword is downgraded to unmatched; if the successfully matched key keywords in the key knowledge point object not only appear in the paraphrased or relayed text of the interpretation answer, then the key keyword is not downgraded to unmatched.

[0109] For example, if the question text is "Please describe what you see in this ultrasound image", and the user answers "There are several parallel horizontal lines in the image, which should be line A", where keywords such as "parallel" and "horizontal lines" are located after the restatement prompt "The image shows...", and the user does not further demonstrate their independent interpretation and reasoning process, then this rule is triggered, and the matching level of the corresponding key point is downgraded to unmatched.

[0110] Rule B: Based on the negative context marker, determine whether the successfully matched key keywords in the key knowledge point object appear within the scope of the negative context marker. The scope of the negative context marker is from the beginning of the negative expression in the clause to the end of the clause (with commas, periods, semicolons, or newlines as clause boundary markers). If the successfully matched key keywords in the key knowledge point object appear within the scope of the negative context marker, then the key keyword is downgraded to unmatched. If the successfully matched key keywords in the key knowledge point object do not appear within the scope of the negative context marker, then the key keyword is not downgraded to unmatched.

[0111] For example, if a user's response includes the clause "I didn't see a clear B-line," the keyword "B-line" exists in the text, but because it appears within the negative context of "didn't see," the user's intention is to express "the B-line doesn't exist" rather than "the B-line exists." In this case, the match of the key knowledge point for identifying the B-line should not be counted as a valid match. Once this rule is triggered, the matching level of the corresponding point will be downgraded to unmatched.

[0112] Rule C: Determine if the character length (number of Chinese characters after removing punctuation and spaces) of the high-density sentences in the response text is less than a preset length threshold (default setting is 15 Chinese characters). The high-density sentences are those where the number of successfully matched key keywords is greater than or equal to the preset key keyword number threshold (preferably three) (bounded by periods, semicolons, or newlines). If the character length of the high-density sentences in the response text is less than the preset length threshold, it is determined that the user has "mentioned but not elaborated" in the sentence, that is, the user has listed multiple terms but has not explained the connotation, mechanism, or clinical significance of any of them. Its essence is a mechanical piling up of keywords rather than a true understanding. The fully matched key keywords in the high-density sentences are downgraded to partially matched, while the partially matched key keywords in the high-density sentences remain unchanged. If the character length of the high-density sentences in the response text is greater than or equal to the preset length threshold, the key keywords in the high-density sentences are not downgraded.

[0113] In step S400, the interpreted response text and the pre-labeled answer are input into a pre-trained medical text embedding model to calculate the semantic similarity score.

[0114] Specifically, the above steps include:

[0115] The interpreted response text and the pre-labeled answer are input into the same pre-trained medical text embedding model and encoded as fixed-dimensional semantic vector representations. The cosine similarity between the semantic vectors corresponding to the interpreted response text and the pre-labeled answer input is calculated. The result ranges from [0.00, 1.00]. The higher the value, the more consistent the two texts are in the semantic space.

[0116] The aforementioned medical text embedding model is a domain-specific text embedding model obtained by fine-tuning a large number of Chinese medical documents and clinical texts based on a general Chinese pre-trained language model.

[0117] In the comprehensive scoring system of this invention, semantic similarity score serves as a supplementary dimension and correction factor to keyword matching score. If the keyword matching score is within a preset passing grade range (default setting is [0.45, 0.65]), the semantic weight is assigned a preset high semantic weight value; if the keyword matching score is not within the preset passing grade range (higher or lower than the preset passing grade range), the semantic weight is assigned a preset low semantic weight value. Semantic similarity only plays a minor adjustment role in the comprehensive score. This design ensures that when keyword matching cannot clearly distinguish user skill levels, semantic similarity can supplement the judgment of whether the user truly understands the core meaning of the content being interpreted, rather than simply using certain keywords by chance.

[0118] The semantic similarity score is obtained based on semantic weight and cosine similarity.

[0119] In step S500, the preprocessed interpretation answer text is matched with the list of common error patterns in the pre-labeled answer by keywords to obtain the common error deductions.

[0120] Specifically, the above steps include:

[0121] For each common error pattern object in the pre-labeled answer corresponding to the unique identifier of the question, the error keywords in the common error pattern objects are matched against the pre-processed interpretation answer text to count the number of matched error keywords. For example, the common error pattern list (common_mistakes) contains two patterns: the error description of the first pattern is "mistaking line A for the pleural line itself", the matching keywords include "pleural line" and "is the pleura", and the correction is "line A is an artifact parallel to the pleural line below it; the pleural line itself is a smooth hyperechoic line"; the error description of the second pattern is "ending the interpretation upon seeing line A without continuing according to the BLUE procedure", the matching keywords include "normal", "no problem", and "exclude", and the correction is "after seeing bilateral lines A, continue to check deep veins according to the BLUE protocol to rule out pulmonary embolism".

[0122] Calculate the incorrect keyword hit rate based on the number of incorrect keywords hit (number of hit keywords ÷ length of the keywords array).

[0123] Determine if the percentage of incorrect keyword hits is greater than or equal to a preset hit rate threshold (default setting is 60%). If the percentage of incorrect keyword hits is greater than or equal to the preset hit rate threshold, then the common error pattern object corresponding to the percentage of incorrect keyword hits is hit, that is, the typical expression of the error pattern is clearly contained in the user's answer. If the percentage of incorrect keyword hits is less than the preset hit rate threshold, then the common error pattern object corresponding to the percentage of incorrect keyword hits is not hit.

[0124] The results of the hit detection are not directly passed to the user. The information of the common error patterns detected (including error pattern descriptions and corresponding correction texts) is passed as internal data to the subsequent comprehensive scoring steps for point deduction. At the same time, it is returned to the client as part of the answer comparison interface response data, and the client decides how to present it to the user (for example, in the comparison results, it may subtly suggest "your judgment may need to be reconsidered regarding part X", rather than directly pointing out "you made error X").

[0125] The common error deduction score is obtained based on the number of common error pattern objects hit and the preset deduction coefficient.

[0126] In step S600, the comprehensive answer score is obtained based on the keyword matching score, semantic similarity score, and common error deductions.

[0127] Specifically, the above steps include:

[0128] The formula for calculating the keyword matching score is: K_score = Σ(matching weight of each key knowledge point × weight type coefficient of that point) ÷ Σ(weight type coefficient of that point), where the weight type coefficient of key knowledge points with the weight type of core point required is 1.0, and the weight type coefficient of key knowledge points with the weight type of bonus point is 0.5. For example, suppose a question contains 3 required points and 2 bonus points. The matching results for the user on this question are as follows: two required points are completely matched (weight 1.0 each), one required point is partially matched (weight 0.5), one bonus point is completely matched (weight 1.0), and one bonus point is not matched (weight 0.0). Then K_score = (1.0×1.0 + 1.0×1.0 + 0.5×1.0 + 1.0×0.5 + 0.0×0.5) ÷ (1.0×3 +0.5×2) = 3.0 ÷ 4.0 = 0.75.

[0129] Common error deduction P_score=min(number of common error patterns hit × single deduction coefficient, maximum deduction), the default single deduction coefficient is 0.15, and the maximum deduction is 0.60.

[0130] The total score is calculated as follows: Total = (K_score × 0.60 + S_score × 0.20 - P_score × 0.20) × Hint_coefficient, where Hint_coefficient is the hint conversion factor, determined based on the highest hint level the user has used on this question. If the user has not used any hints, Hint_coefficient = 1.0; if the highest level used by the user is Level 1, Hint_coefficient = 1.0 (Level 1 hints are not converted); if the highest level is Level 2, Hint_coefficient = 0.9; if the highest level is Level 3, Hint_coefficient = 0.7.

[0131] In step S700, the system determines whether the answer text is correct based on the overall answer score. If the answer text is incorrect, it progressively returns hierarchical prompt text to the user and receives the answer text input by the user based on the hierarchical prompt text.

[0132] Specifically, the above steps include:

[0133] If the overall score of the answer is greater than or equal to the preset correctness threshold (default setting is 0.70), and all current core keywords are completely matched, and no common error pattern objects are hit, then the user's answer text for the medical image recognition question will be judged as correct.

[0134] If the overall score of the answer is less than the preset correctness threshold, or if the current core keyword is not a complete match, or if a common error pattern object is hit, that is, if any of the above three conditions are not met, then the user's answer text for the medical image recognition question will be judged as incorrect.

[0135] It's important to note that the answer comparison API does not directly return the original standard answer to the question when returning the comparison results to the client. The returned response data only includes: a comprehensive correctness judgment (is_correct), a comprehensive score, a list of covered key knowledge points (each point is marked with its matching level and the specific keywords it matched), a list of uncovered key knowledge points (each missing point is marked with its descriptive text and importance level, sorted in descending order of importance), a semantic similarity score, a list of common error patterns and their correction suggestions, the remaining allowed attempts, and suggested next steps. This "comparison results are visible, standard answers are not" design ensures that users must improve their answers through their own thinking and interpretation, rather than simply memorizing the standard answer.

[0136] When a user's comparison result for a question is incorrect, the standard answer is not revealed directly. Instead, progressive guidance information is provided to the user through a hierarchical prompting interface.

[0137] For incorrectly interpreted answer texts, receive prompt requests from users and the unique identifier of the question.

[0138] Determine whether the level parameter (hint_level, value 1, 2 or 3) of the current hint request is within the preset valid level range; if the level parameter of the current hint request is not within the preset valid level range (1 to 3), then reject the hint request.

[0139] If the level parameter of the current hint request is within the preset valid level range, then it is determined whether all the preceding requests of the current hint request have been completed. That is, when hint_level is 2, it must be verified that the user has successfully requested the level 1 hint of the question. When hint_level is 3, it must be verified that the user has successfully requested the level 1 hint and level 2 hint of the question. If the preceding requests of the current hint request have not been completed, the hint request is rejected and the server returns a sequence violation error.

[0140] If all prerequisite requests for the current hint request have been completed, check if the hint text array uniquely identified by the question contains a hint text corresponding to the current hint request level (hints[hint_level-1] is not empty); if the hint text array uniquely identified by the question does not contain a hint text corresponding to the current hint request level, then reject the hint request; if the hint text array uniquely identified by the question contains a hint text corresponding to the current hint request level, then output the hint text corresponding to the current hint request level to the client. For example, the three items in the three-level progressive hints array are as follows: Level 1 hints[0]: "Please note below the pleural line, do you see multiple horizontal lines parallel to the pleural line? What is the pattern of the spacing between them?"; Level 2 hints[1]: "This is one of the most common findings in normal lung ultrasound. In the BLUE protocol, if this finding is seen bilaterally and lung gliding is normal, what area should be examined next?"; Level 3 hints[2]: "This is a typical A-line finding—a reverberant artifact, indicating subpleural aerodynamic lung. Bilateral A-lines + normal lung gliding can rule out interstitial syndrome, and deep vein examination should continue to rule out pulmonary embolism."

[0141] More specifically, the content progression logic of the three-level prompt text adopted in this invention is as follows:

[0142] Level 1 hints[0] are limited to guiding users to pay attention to the location of key anatomical regions in medical images, the spatial relationship between image features, or the observation angle of morphological features, without revealing any interpretation direction or diagnostic conclusion. Level 2 hints[1] extend to indicating the interpretation direction or the thinking path for differential diagnosis, without giving a complete interpretation conclusion or specific medical terminology. Level 3 hints[2] provide a nearly complete interpretation framework, including key clinical decision points and core medical terminology, but retain some room for thought (such as not directly giving the final treatment decision).

[0143] Each level of prompt text consists of fixed natural language text pre-written by domain experts for that specific question during the question bank annotation phase, and is not generated in real-time by the AI ​​model during runtime. This design ensures the medical accuracy of the prompt content, giving domain experts ample time to review and refine the wording of the prompts during the annotation phase, thus avoiding potential guidance bias or factual errors that may occur with AI-generated prompts in real time.

[0144] The request and use of hints are subject to the following four constraints: First, requests must be made strictly in the order of Level 1 → Level 2 → Level 3, and skipping intermediate levels to directly request higher-level hints is prohibited. Second, a hint of the same level can only be requested once by the same user for the same question; repeated requests for the same level will result in an error response. Third, after each request and receipt of the hint text, the user must resubmit an interpretation (i.e., call the answer comparison interface again) before requesting the next level of hints. This constraint ensures the actual teaching effect of the hints, requiring users to think and provide a new answer after using the current hint, rather than answering again after continuously obtaining all hints. Fourth, if a user's resubmitted interpretation after using all three levels of hints is still judged as incorrect, the suggestion operation field of the answer comparison interface returns "reveal_answer," suggesting the disclosure of the standard answer. The client system then decides whether to display the standard answer text to the user based on its own interaction design.

[0145] Figure 2 The diagram illustrates the principle of the hierarchical prompting process provided in an embodiment of the present invention.

[0146] Preferably, the embodiments of the present invention further include: after a single round of answering session ends or a batch of answering sessions is completed, the client calls the server's batch statistics interface (submit_batch_result), and passes the session identifier (session_id), user identifier (user_id), and an array of all answering result records completed in this round in the request parameters.

[0147] Each answer record contains the following fields: unique identifier of the question, primary category to which the question belongs, difficulty level, user's final correctness judgment (is_correct) for the question, overall score, number of attempts actually used, list of hint levels used, answering time (in seconds), list of key knowledge points still missing in the final judgment, and text describing common error patterns hit on the question.

[0148] After receiving the above array of answer results, the server-side statistical analysis module performs aggregation calculations according to the following rules: First, using the primary category as the aggregation key, it calculates the answer accuracy rate (number of correctly judged questions ÷ total number of answers in the category), average number of attempts, average number of hints used, and average answer time for all answer records under each category; Second, it calculates the total frequency of each key knowledge point identifier being marked as "missing" across all answer records, sorts them in descending order of frequency, and identifies the knowledge points most frequently missed by the user group or the user in this round of answering; Third, it calculates the total frequency of each common error pattern description text hit across all answer records, sorts it in descending order of frequency, and identifies the error types most frequently made by the user group or the user.

[0149] Based on the above aggregation results, automatic identification of weak categories is performed. The identification rules are as follows: Rule 1: If the correct answer rate of a category is not higher than the preset weak category threshold (default setting is 60%), then the category is marked as a "weak category"; Rule 2: If the correct answer rate of a category is higher than the weak category threshold, but there is a key knowledge point in the category whose weight class is required, and the proportion of its missing frequency to the total number of answers in the category exceeds the preset proportion threshold (default setting is 50%), then the category is marked as "partially weak", and the specific weak point is noted in the marking information; Rule 3: If the cumulative number of hits of the same common error pattern in a category reaches the preset number threshold (default setting is 3 times), then the category-error combination is marked separately, indicating that there is a systematic common misunderstanding in the category.

[0150] The aggregated results and the weak category identification results are organized into structured response data and returned to the client. The response data also includes a priority improvement suggestion field automatically generated by the system, listing the categories that are recommended for priority review or reinforcement, as well as the specific knowledge points and common mistakes to focus on under each category.

[0151] Preferably, the embodiments of the present invention further include: a two-dimensional control mechanism for controlling the difficulty of the questions. This mechanism does not change the core process of the above steps, but runs independently as a question-generating strategy layer. At each decision point where the "next question" needs to be selected (i.e. when the question-generating interface is called), the mechanism selects a question from the question bank that matches the user's current ability level based on the user's historical answer performance.

[0152] The aforementioned two-dimensional control mechanism specifically includes:

[0153] The vertical dimension control module locks the range of selectable question groups based on the user-preset or system-suggested difficulty levels (D ∈ {basic, advanced, expert}). The system has pre-grouped all questions into three groups—basic, advanced, and expert—based on the difficulty field in the question bank metadata. When the vertical dimension control module receives the target difficulty level parameter, it restricts the search range of selectable questions in the question retrieval interface to the group corresponding to that level, preventing questions from being generated across levels. Each level has a different focus: the basic level focuses on the identification of anatomical structures, the judgment of basic imaging features, and preliminary decisions using standardized procedures; the advanced level focuses on the interpretation of pathophysiological mechanisms in imaging manifestations, multi-parameter integrated interpretation, and systematic differential diagnosis; the expert level focuses on the selection of imaging strategies in complex clinical scenarios, the assessment of the limitations of standardized procedures, and the handling of uncertainties.

[0154] The horizontal dimension control module is responsible for dynamically fine-tuning the difficulty of selected questions within the same difficulty level based on the user's continuous answering performance. The specific runtime control rules are as follows: The system maintains two runtime counters for each user session: a consecutive correct answer count (C) and a consecutive incorrect answer count (E). Both counters are initialized to 0 at the start of the session. After the user completes each question and obtains a comprehensive judgment result, if the judgment is correct, C is incremented by 1 and E is reset to 0; if the judgment is incorrect, E is incremented by 1 and C is reset to 0. When the question selection interface is called and the next question needs to be selected from the current level, the system reads the current C and E values ​​and executes the following question selection logic: If C ≥ the first preset threshold (default set to 3), a question is randomly selected from the high-difficulty subset of the current level, and C is reset to 0 after selection; if E ≥ the second preset threshold (default set to 2), a question is randomly selected from the low-difficulty subset of the current level, and E is reset to 0 after selection; if neither of the above two conditions is met, a question is randomly selected from the default subset of the current level, and C and E remain unchanged.

[0155] The pre-division scheme for the above-mentioned difficulty subsets is as follows: Within the same difficulty level, two indicators are calculated for each question: the total number of key knowledge points for that question and the proportion of points with the weight type "required" in that question. Questions that simultaneously meet the conditions of "key_points array length ≤ 3" and "required type point proportion ≤ 60%" are assigned to the low-difficulty subset; questions that simultaneously meet the conditions of "key_points array length ≥ 5" and "required type point proportion ≥ 80%" are assigned to the high-difficulty subset; all other questions are assigned to the default subset. This pre-division scheme is automatically calculated by the system when questions are added to the database, without requiring manual specification by annotation experts.

[0156] In addition, the vertical dimension control module includes an upgrade suggestion mechanism. When a user continuously completes a preset number of questions (default setting is 5 questions) within a certain difficulty level, and the overall score for each question is not lower than the preset upgrade standard (default setting is 0.85 points), an upgrade suggestion mark is attached to the statistical analysis results, suggesting that the user or system administrator upgrade the target difficulty level to the next level. The level change requires confirmation before it takes effect. The confirming party can be the user (in adaptive learning mode) or the system administrator (in unified teaching management mode).

[0157] Figure 3 This diagram illustrates the dual-dimensional difficulty control principle provided by an embodiment of the present invention. The diagram uses a 3-row (basic / advanced / expert) × 3-column (low-difficulty subset / default subset / high-difficulty subset) matrix to show the hierarchical grouping structure of the questions. The diagram indicates the horizontal movement to the right driven by consecutive correct answers (≥3 times), the horizontal movement to the left driven by consecutive incorrect answers (≥2 times), and the vertical progression direction from bottom to top driven by consecutive high scores (≥0.85 on 5 questions).

[0158] In addition, embodiments of the present invention also provide a medical imaging question interaction system based on pre-annotated standard answers, the system comprising:

[0159] The question bank storage module is used to store medical image recognition questions and pre-annotated answers.

[0160] Specifically, the aforementioned question bank storage module manages and maintains all data assets of the pre-annotated question bank, persistently storing two parts of data on a non-volatile storage medium: the first part is a question bank metadata database, which stores the metadata records of all questions in a structured document format (such as a JSON document collection or a relational database table). The data items contained in the metadata of each question are completely consistent with the question bank metadata structure defined in step S1 of the method scheme of this invention; the second part is a medical image file library, which stores the associated image files of all questions in a file system directory structure, supporting PNG, JPEG, DICOM, GIF and common video container formats. The question bank storage module is also responsible for data integrity verification when questions are added to the database and query filtering functions when questions are retrieved, and provides a data access interface to the upper-layer server module.

[0161] The aforementioned question bank metadata includes the following data items: a unique question identifier (id), using the format of "category prefix number," ensuring global uniqueness within the system, used for referencing and locating specific questions between interfaces; a primary category, indicating the knowledge domain to which the question belongs, examples of which include BLUE protocol, cardiac function assessment, volume assessment, and lung ultrasound assessment; a secondary subcategory, an optional data item representing further subdivision under the primary category, such as further subdivision under the BLUE protocol category into A-line recognition, B-line recognition, lung slippage judgment, and PLAPS point assessment; a difficulty level, an enumerated type with values ​​of basic, advanced, or expert, corresponding to different depths of knowledge required; and an image file storage path (image_path), pointing to the associated medical image file (supporting PNG, JPEG, and DICOM). The relative path of the image in non-volatile storage media (e.g., format); image type (image_type), an enumeration type, with values ​​of static image, dynamic image (e.g., GIF), or video clip (video_clip), used by the client to select the appropriate image presentation method; question text (question_text), the interpretation question posed to the user, such as "Please describe what you see in this ultrasound image of the upper right lung and give your interpretation"; standard answer text (standard_answer), the full text of the standard interpretation of the question written in natural language by a domain expert, serving as the benchmark for subsequent semantic comparison; key points list (key_points), an array type, where each element is a key point object containing the following three fields: point text, describing the specific content of the point; weight type (wei... `ght`, with values ​​of `required` (indicating the point is a core point that the user must mention) or `bonus` (indicating the point is a bonus point; mentioning it adds points, not mentioning it deducts points); `keywords`, a string array containing standard medical terminology for the point, as well as common synonyms and alternative expressions, serves as the basis for keyword matching; `common_mistakes`, an optional array type, where each element is a common error pattern object containing the following three fields: error pattern description (`pattern`), describing the typical manifestation of the error; `keywords`, containing typical keywords corresponding to the error pattern, used to detect the error in the user's response text; and correction text, a correction message used to provide corrective information to the interactive system when the user hits the error.The three-level progressive hint text array (hints) is a fixed-length string array of 3 characters. The 0th, 1st, and 2nd elements of the array correspond to the 1st, 2nd, and 3rd level hint texts of the question, respectively. The source annotation (source) is an optional string item that records the knowledge source of the standard answer, such as the DOI number of a reference, the chapter number of a textbook, or the name of the annotation expert, for knowledge tracing and quality auditing. The question creation time (created_at) and last update time (updated_at) are used for version management and change tracking. The annotation expert identifier (expert_id) is an optional string item that records the identity of the domain expert who performed this annotation.

[0162] When question bank data is added to the database, the following data integrity verification rules are executed: Verify that the format of the question's unique identifier conforms to the preset regular expression rules and that the identifier is globally unique within the question bank; verify that the value of the difficulty level field is within the range of the three valid enumeration values: basic, advanced, and expert; verify that the key knowledge point list (key_points) contains at least one key point with a weight type of required, ensuring that each question has at least one core knowledge point that must be tested; verify that the length of the hint text array (hints) is strictly equal to 3, ensuring that all three levels of hints for each question are filled in completely; verify that the file pointed to by the image file storage path (image_path) actually exists on the storage medium. If any verification rule fails, the question is not written to the question bank metadata database, and the specific reason for the verification failure is returned to the annotator.

[0163] The data receiving module is used to receive the user's interpretation and answer text for medical image recognition questions, as well as the unique identifier of the question.

[0164] The semantic comparison engine module includes a text preprocessing unit, a keyword matching unit, a semantic similarity calculation unit, a common error detection unit, and a comprehensive scoring unit.

[0165] Specifically, the text preprocessing unit is used to preprocess the interpretation and answer text to obtain the preprocessed interpretation and answer text; the keyword matching unit is used to perform keyword matching on the preprocessed interpretation and answer text based on the list of key knowledge points of the pre-labeled answer corresponding to the unique identifier of the question, and obtain a keyword matching score; the semantic similarity calculation unit is used to input the interpretation and answer text and the pre-labeled answer into a pre-trained medical text embedding model to calculate the semantic similarity score; the common error detection unit is used to perform keyword matching on the preprocessed interpretation and answer text and the list of common error patterns of the pre-labeled answer, and obtain a common error deduction score; the comprehensive scoring unit is used to obtain a comprehensive answer score based on the keyword matching score, the semantic similarity score, and the common error deduction score, and to determine whether the interpretation and answer text is correct based on the comprehensive answer score.

[0166] The layered prompting module is used to receive prompt requests from users for incorrect interpretation of the response text, progressively return layered prompt text to the user, and receive the interpretation response text entered by the user based on the layered prompt text.

[0167] Specifically, the aforementioned hierarchical hint module is responsible for managing the request verification and return logic for three levels of progressive hints. This module maintains a history of hint requests for each user on each question in memory, indexed by a composite key {user_id, question_id}. The records contain a list of requested hint levels and the timestamp of each request. Upon receiving a request to the hierarchical hint interface, the module performs the following verification steps: verifying that the requested hint level parameter value is within the valid range of 1-3; querying the history of hint requests for that user on that question and performing order constraint verification; and verifying that the question has already configured the corresponding level of hint text in the question bank metadata. After successful verification, the module retrieves the pre-written hint text corresponding to that level from the question bank storage module, adds a conversion factor and information on whether there is a next-level hint, and returns it. Immediately after each successful return of hint text, the module updates the user's history of hint requests for that question, appending the level and timestamp of the current request.

[0168] Preferably, the medical imaging question interaction system based on pre-annotated standard answers provided in this embodiment of the invention further includes:

[0169] The standardized protocol server module runs as an independent service process and can be deployed on a single physical computing device or, via containerization technology (such as Docker), on a cluster of multiple computing devices. It exposes five utility interfaces to clients via HTTP / HTTPS or gRPC protocols: a category query interface (list_categories), a question retrieval interface (get_question), an answer comparison interface (check_answer), a hierarchical hint interface (get_hint), and a batch statistics interface (submit_batch_result). The input parameter structure, output response structure, and exception handling mechanism of these five interfaces are consistent with step S2 in this method. The question retrieval interface uses a field whitelist filtering mechanism to exclude standard answer text and related sensitive fields when returning question information, ensuring system security. Each interface follows a stateless design principle; each request independently carries all the parameter information required for that request. The server does not maintain persistent storage of client session state, thus supporting horizontal scaling deployment of the server.

[0170] Preferably, the medical imaging question interaction system based on pre-annotated standard answers provided in this embodiment of the invention further includes:

[0171] The statistical analysis module is responsible for analyzing and processing the input array of answer results records after receiving a call request from the batch statistics interface, according to the aggregation calculation rules and weak category identification rules defined in the method steps of this invention. This module contains three processing sub-processes: an aggregation calculation sub-process—calculating the accuracy rate, average number of attempts, average number of hints used, and average answer time item by item according to the category dimension; a frequency statistics sub-process—statistically calculating the frequency of missing key knowledge points and the hit frequency of common error patterns across all answer records; and a weak category identification sub-process—based on the results of aggregation calculation and frequency statistics, identifying weak categories and some weak categories according to three preset identification rules (accuracy threshold rule, key point missing rate rule, and error pattern frequency rule), and generating priority improvement suggestion text. The statistical analysis results are returned in a structured format.

[0172] Preferably, the medical imaging question interaction system based on pre-annotated standard answers provided in this embodiment of the invention further includes:

[0173] The difficulty-based dual-dimensional control module includes a vertical dimension controller, a horizontal dimension controller, and an upgrade suggestion trigger.

[0174] The aforementioned vertical dimension controller is used to limit the range of selectable question groups based on the target difficulty level. The aforementioned horizontal dimension controller, within the same level, determines whether to select questions from a low-difficulty subset, a default subset, or a high-difficulty subset based on the values ​​of consecutive correct answers (C) and consecutive incorrect answers (E) and their corresponding preset thresholds. The aforementioned upgrade suggestion trigger monitors the user's continuous performance within a certain level and triggers an upgrade suggestion when the performance meets preset upgrade conditions.

[0175] Compared with the prior art, the present invention has the following beneficial effects:

[0176] This invention does not rely entirely on AI image recognition. The construction and operation of the entire system do not require GPU computing resources, large-scale medical image annotation training datasets, or the deployment and maintenance of deep learning inference frameworks. When adding new questions to the question bank, only qualified domain experts need to complete the text annotation before it can be put into use. The deployment cycle for new questions is shortened from several weeks, which is usually required by AI-based solutions, to minutes.

[0177] In this invention, the pre-annotated standard answers are always consistent with clinical standards. Because the standard answers are directly written and reviewed by domain experts with clinical practice experience, rather than generated by machine learning models based on statistical patterns in training data, the technical risk of inconsistency between AI interpretation standards and actual clinical standards is completely avoided. This prevents interpretation errors caused by AI models due to training data bias or insufficient model generalization ability from being amplified and spread during the teaching and evaluation process.

[0178] This invention achieves decoupling between system modules through a standardized protocol. The client-side AI dialogue agent system, the front-end application system, and the back-end question bank server communicate through five clearly defined tool interfaces. Updates to the question bank content (e.g., adding questions, revising standard answers, supplementing common error patterns) do not affect the normal operation of the client; upgrades or replacements to the client's interaction logic and user interface do not affect the stability of the question bank data structure. The stateless design of each interface allows the server to be horizontally scaled by increasing the number of instances to support the growth of concurrent users without fundamentally altering the system architecture.

[0179] The free-response format employed in this invention more accurately reflects the user's interpretive abilities. Users observe real medical images with their own eyes and describe their observations and reasoning using their own natural language, rather than passively selecting from a few preset options. Because users cannot arrive at the correct answer through elimination, guessing, or semantic analysis of the option text, the assessment conclusions provided by this invention are closer to the user's actual ability to independently interpret images in real clinical situations.

[0180] This invention provides data support for personalized learning path planning by automatically locating weak knowledge domains and offering priority improvement suggestions. Instead of simply outputting a general "accuracy rate" figure, it identifies specific knowledge domain categories where weaknesses exist, which key knowledge points are repeatedly missed, and which error patterns are repeatedly triggered. Users or educational administrators can then develop differentiated review and reinforcement plans based on this targeted information.

[0181] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

[0182] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, or alterations made by those skilled in the art using the disclosed technical content shall fall within the protection scope of the present invention.

Claims

1. A method for interactive medical imaging questions based on pre-annotated standard answers, characterized in that, The method includes: Receive the user's interpretation and answer text for medical image recognition questions, along with the question's unique identifier; The judgment response text is preprocessed to obtain the preprocessed judgment response text; Based on the list of key knowledge points of the pre-annotated answer corresponding to the unique identifier of the question, keyword matching is performed on the preprocessed interpretation and answer text to obtain a keyword matching score. The interpreted response text and the pre-labeled answer are input into a pre-trained medical text embedding model to calculate a semantic similarity score; The preprocessed interpretation and answer text is matched with the list of common error patterns in the pre-labeled answers using keywords to obtain the common error deductions; The overall answer score is obtained based on the keyword matching score, the semantic similarity score, and the common error deductions. The system determines whether the interpreted answer text is correct based on the overall score of the answer. If the interpreted answer text is incorrect, it progressively returns hierarchical prompt text to the user and receives the interpreted answer text input by the user based on the hierarchical prompt text.

2. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 1, characterized in that, Receive the user's interpretation of medical image recognition questions, along with the question's unique identifier, including: After receiving a question request from a user on the client side, the server sends the medical image recognition question data to the interception layer. In the interception layer, the medical image recognition question data is compared with a preset whitelist, preset fields are removed, and the medical image recognition question data that has passed the interception layer is sent to the client. The preset fields include standard answer text, key knowledge point list, common error pattern list and prompt text. It receives the user's interpretation of medical image recognition questions and the question's unique identifier from the client.

3. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 1, characterized in that, The judgment response text is preprocessed to obtain the preprocessed judgment response text, including: The interpretation and response text is cleaned by removing punctuation marks and common stop words to obtain pure content text; The plain text is segmented by calling a word segmenter loaded with a medical dictionary to obtain segmented text; The system identifies negative expressions and paraphrasing prompts in the segmented text, marks text segments containing negative expressions with negative context, and marks text segments containing paraphrasing prompts with paraphrasing prompts.

4. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 3, characterized in that, Based on the list of key knowledge points in the pre-annotated answer corresponding to the unique identifier of the question, keyword matching is performed on the preprocessed interpretation text to obtain a keyword matching score, including: For each key knowledge point object in the list of key knowledge points in the pre-annotated answer corresponding to the unique identifier of the question, based on the preset hierarchical matching rules, the key knowledge point objects are used to perform hierarchical matching in the pre-processed interpretation and answer text. If all key keywords in the key knowledge point object match the preprocessed interpretation and answer text, then the corresponding key keywords are defined as a complete match, and the matching weight of the current key knowledge point object is assigned the first preset matching weight. The key keywords include core key keywords and bonus key keywords. The core key keywords correspond to the core type weight, and the bonus key keywords correspond to the bonus type weight. If at least one key keyword in the key knowledge point object matches the preprocessed interpretation and response text, then the corresponding key keyword is defined as a partial match, and the matching weight of the current key knowledge point object is assigned the second preset matching weight. If the key keywords in the key knowledge point object do not appear in the preprocessed judgment and answer text, the corresponding key keywords are defined as unmatched, and the matching weight of the current key knowledge point object is assigned to the third preset matching weight, wherein the first preset matching weight is greater than the second preset matching weight, and the second preset matching weight is greater than the third preset matching weight. Based on the matching weight and type weight of the key keywords in each key knowledge point object, the comprehensive keyword matching score of the interpretation and response text is obtained.

5. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 4, characterized in that, The method further includes: Based on the paraphrase guidance mark, determine whether the key keywords that are successfully matched in the key knowledge point object only appear in the paraphrase or retelling text in the interpretation and answer text, and the key keywords that are successfully matched are key keywords that are completely matched or partially matched; If the key knowledge point object that is successfully matched only appears in the paraphrased or retelled text in the interpretation and answer text, then the key knowledge point keyword is downgraded to unmatched; If the key knowledge point object that is successfully matched appears not only in the paraphrased or retelled text in the interpretation and answer text, then the key knowledge point keyword will not be downgraded to unmatched. Based on the negative context marker, it is determined whether the key keywords that are successfully matched in the key knowledge point object appear within the scope of the negative context marker. The scope of the negative context marker is from the beginning of the negative expression word in the clause to the end of the clause. If a key keyword that is successfully matched in the key knowledge point object appears within the scope of the negative context marker, then the key keyword is downgraded to unmatched. If the key knowledge point object that is successfully matched does not appear in the scope of the negative context marker, then the key knowledge point keyword will not be downgraded to unmatched; Determine whether the character length of high-density sentences in the response text is less than a preset length threshold. The high-density sentences are those in which the number of successfully matched key words is greater than or equal to the preset key word number threshold. If the character length of a high-density sentence in the response text is less than a preset length threshold, then the exact matching keywords in the high-density sentence will be downgraded to partial matching, while the partial matching keywords in the high-density sentence will remain unchanged. If the character length of a high-density sentence in the response text is greater than or equal to a preset length threshold, then the key keywords in that high-density sentence will not be downgraded.

6. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 1, characterized in that, The interpreted response text and the pre-labeled answer are input into a pre-trained medical text embedding model to calculate a semantic similarity score, including: The interpretation response text and the pre-labeled answer are input into the same pre-trained medical text embedding model. The cosine similarity between the semantic vectors corresponding to the interpretation response text and the pre-labeled answer input is calculated. The medical text embedding model is a medical-specific text embedding model obtained by domain fine-tuning with a large number of Chinese medical literature and clinical texts on the basis of a general Chinese pre-trained language model. If the keyword matching score is within the preset passing range, the semantic weight will be assigned a preset high semantic weight value. If the keyword matching score is not within the preset passing threshold range, the semantic weight will be assigned a preset low semantic weight value. The semantic similarity score is obtained based on the semantic weight and the cosine similarity.

7. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 1, characterized in that, The preprocessed interpretation response text is matched with the list of common error patterns in the pre-annotated answers using keywords to obtain common error deductions, including: For each common error pattern object in the list of common error patterns of the pre-annotated answer corresponding to the unique identifier of the question, the common error keywords in the common error pattern object are matched in the preprocessed interpretation and answer text, and the number of matched error keywords is counted. Calculate the hit rate of incorrect keywords based on the number of incorrect keywords hit; Determine whether the hit rate of the erroneous keywords is greater than or equal to a preset hit rate threshold; If the hit rate of the erroneous keyword is greater than or equal to the preset hit rate threshold, then the common error pattern object corresponding to the hit rate of the erroneous keyword is hit; If the hit rate of the erroneous keyword is less than the preset hit rate threshold, then the common error pattern object corresponding to the hit rate of the erroneous keyword is not hit; The common error deduction is obtained based on the number of common error pattern objects hit and the preset deduction coefficient.

8. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 1, characterized in that, The accuracy of the interpreted response text is determined based on the overall score of the response, including: If the overall score of the answer is greater than or equal to the preset correctness threshold, and all the current core keywords are completely matched, and no common error pattern objects are hit, then the user's answer text for the medical image recognition question will be judged as correct. If the overall score of the answer is less than the preset correctness threshold, or if the current core keyword is not a complete match, or if a common error pattern is hit, then the user's answer to the medical image recognition question will be judged as incorrect.

9. The interactive method for medical imaging questions based on pre-annotated standard answers according to claim 1, characterized in that, If the interpretation response text is incorrect, a progressively layered prompt text is returned to the user, and the interpretation response text input by the user based on the layered prompt text is received, including: For interpretations of the response text that are deemed incorrect, receive prompt requests sent by the user; Determine whether the level parameter of the current prompt request is within the preset valid level range; If the level parameter of the current prompt request is not within the preset valid level range, the prompt request will be rejected. If the level parameter of the current prompt request is within the preset valid level range, then determine whether all the preceding requests of the current prompt request have been completed; If not all the prerequisite requests for the current prompt request have been completed, then the prompt request is rejected; If all prerequisite requests for the current prompt request have been completed, then determine whether there is a prompt text corresponding to the current prompt request level in the prompt text array that uniquely identifies the question. If the prompt text array that uniquely identifies the question does not contain the prompt text corresponding to the current prompt request level, then the prompt request is rejected; If the prompt text array that uniquely identifies the question contains a prompt text corresponding to the current prompt request level, then the prompt text corresponding to the current prompt request level will be output to the client.

10. A medical imaging question-and-answer interactive system based on pre-annotated standard answers, characterized in that, The system includes: The question bank storage module is used to store medical image recognition questions and pre-annotated answers; The data receiving module is used to receive the user's interpretation and answer text for medical image recognition questions and the unique identifier of the question; The semantic comparison engine module includes a text preprocessing unit, a keyword matching unit, a semantic similarity calculation unit, a common error detection unit, and a comprehensive scoring unit; The text preprocessing unit is used to preprocess the judgment response text to obtain the preprocessed judgment response text. The keyword matching unit is used to perform keyword matching on the preprocessed interpretation and answer text based on the list of key knowledge points of the pre-annotated answer corresponding to the unique identifier of the question, and to obtain a keyword matching score. The semantic similarity calculation unit is used to input the interpreted answer text and the pre-labeled answer into a pre-trained medical text embedding model to calculate the semantic similarity score; The common error detection unit is used to perform keyword matching between the preprocessed interpretation and answer text and the list of common error patterns of the pre-labeled answer to obtain common error deductions; The comprehensive scoring unit is used to obtain a comprehensive answer score based on the keyword matching score, the semantic similarity score, and the common error deductions, and to determine whether the interpreted answer text is correct based on the comprehensive answer score. The layered prompting module is used to receive prompt requests from users for incorrect interpretation of the response text, progressively return layered prompt text to the user, and receive the interpretation response text entered by the user based on the layered prompt text.

Images