An AI-based strabismus diagnosis and interpretation method, electronic device, and application product based on causality and large-scale models.

Abstract

This invention discloses an AI-based strabismus diagnosis and interpretation method, electronic device, and program product based on causality and a large-scale model. The method includes: obtaining the AI ​​diagnostic conclusion of a case to be interpreted; simultaneously extracting the SHAP value and corresponding features of the case; using the AI ​​diagnostic conclusion as an index, retrieving typical feature combinations of the strabismus subtype corresponding to the AI ​​diagnostic conclusion from an AI behavioral knowledge base; comparing the typical feature combinations with the feature with the highest SHAP value of the case to be interpreted; using the feature corresponding to the matching SHAP value as a starting point, performing a search in a clinical causal knowledge base to find a logical chain from the starting point to the diagnostic conclusion; using the found logical chain as instructions and context for a large-scale language model, and outputting the final interpretation text through the large-scale language model. This invention includes a deep interpretation of a complete causal logical chain, highly consistent with the diagnostic thinking of clinicians, and enhances professionalism.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical diagnostic technology, and in particular to an AI-based strabismus diagnostic interpretation method, electronic device, and program product based on causality and large model. Background Technology

[0002] Artificial intelligence (AI), particularly deep learning networks, has shown great potential in the automated diagnosis of ophthalmic diseases, such as strabismus detection and subtype classification. To enhance clinicians' trust in AI diagnostic conclusions, interpretable AI techniques have been introduced. For example, existing research has proposed causal-inspired graph neural networks (CI-GNN) for strabismus subtype classification and utilized SHAP (SHapley Additive exPlanations) technology to interpret the model's predictions, numerically quantifying the contribution of each visual feature to the diagnostic results.

[0003] However, the aforementioned existing technical explanations are not intuitive and lack clinical logic because the explanations provided by the SHAP method are numerical lists or charts of feature importance. This format is not intuitive enough for clinicians; it only answers "which features the model focuses on," but does not provide a complete "clinical reasoning process" from "evidence" (abnormal features) to "conclusion" (diagnosis) that aligns with the doctor's thinking habits. Directly translating the numerical SHAP results into natural language using a general Large Language Model (LLM) risks unreliable content. Large models may produce "illusions," generating explanations that seem reasonable but do not conform to the model's actual decision-making basis or clinical facts, thus compromising the accuracy of interpretability. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based strabismus diagnosis and interpretation method, electronic device, and program product based on causality and a large model, to solve the aforementioned technical problems existing in the prior art. The various technical effects of the preferred solutions among the many technical solutions provided by this invention are detailed below.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an AI-based strabismus diagnosis and interpretation method based on causality and a large model, comprising: Obtain the AI ​​diagnostic conclusion of the case to be explained, and simultaneously extract the SHAP value of the case to be explained and the features corresponding to the SHAP value; Using the AI ​​diagnostic conclusion as an index, retrieve the typical feature combination of the strabismus subtype corresponding to the AI ​​diagnostic conclusion from the AI ​​behavior knowledge base; The typical feature combination is compared with the feature with the highest SHAP value of the case to be explained. The feature corresponding to the matching SHAP value is used as the starting point to perform a search in the clinical causal knowledge base to find a logical chain from the starting point to the diagnostic conclusion. The found logical chain is used as the instruction and context of the large language model, and the final interpreted text is output through the large language model.

[0006] In some embodiments, the SHAP value includes the SHAP local contribution score, the baseline value, and / or the sequence symbol corresponding to the feature ranking; the typical feature combination includes the difference between the exotropia in the downward gaze position and the exotropia in the upward gaze position, the exotropia in the downward gaze position, the degree of outward shift of the corneal light point in the downward gaze position, the horizontal difference between the pupil centers of the two eyes in the downward gaze position, and the change in graph connectivity weight between the nodes of the upper and lower eyes.

[0007] In some embodiments, the construction of the AI ​​behavior knowledge base includes: Eye position image features or eye position coordinate features are constructed as graph nodes. A clinical adjacency matrix is ​​generated according to the clinical diagnosis rules of strabismus. The clinical adjacency matrix is ​​used as the initial topology for message passing in the graph neural network. The edge weights are initialized according to the clinical correlation strength between different eye position nodes. During model training, the edge connection relationship is adaptively updated according to the edge weights, attention coefficients and / or feature contributions. Edges below a preset threshold are dynamically pruned. Collect the SHAP analysis results or interpretable integral gradient algorithm analysis results of the trained CK-GNN model on the labeled dataset, and identify the missing features corresponding to the classification error of the CK-GNN model based on the analysis results; For each strabismus subtype, the CK-GNN model, after training, is used to perform clustering statistics to extract the typical feature combination corresponding to the highest SHAP value.

[0008] In some embodiments, before clustering statistics, an intermediate neuron that explicitly maps clinical indicators is inserted after the GNN message passing layer and before the final Softmax classification layer. Through causal consistency loss, the CK-GNN model is forced to calculate the indicator corresponding to the intermediate neuron before making the final diagnosis.

[0009] In some embodiments, the causal consistency loss is obtained through the following steps: The set of clinical indicators related to strabismus subtypes is determined according to the clinical diagnostic rules for strabismus. The set of clinical indicators includes the horizontal deviation in the primary gaze position, the difference in horizontal deviation in the vertical gaze position, the difference in horizontal deviation in the left and right gaze position, the difference in vertical deviation, the offset of the corneal light point, the relative offset of the pupil center, and / or the binocular motion coordination index. The clinical indicator set is mapped to multiple intermediate neurons in the causal bottleneck layer, so that each intermediate neuron outputs a corresponding clinical indicator. Based on manually labeled clinical measurements or regular indicators calculated from eye position coordinates, the difference between the output value of the interneuron and the true value of the corresponding clinical indicator is calculated, and this difference is used as the causal consistency loss. The causal consistency loss and the strabismus subtype classification loss are jointly optimized.

[0010] In some embodiments, the construction of the clinical causality knowledge base includes: Automated entity extraction and attribute parsing are performed on strabismus diagnosis and treatment imaging data. The parsed terms are aligned to medical standard ontology, and the causal relationship direction is defined. The imaging features corresponding to the aligned terms are used as cause nodes, and the diagnostic conclusions are used as result nodes to obtain a causal graph.

[0011] In some embodiments, the causal graph constructed includes an observation layer, a mechanism layer, and a conclusion layer corresponding to each node.

[0012] In some embodiments, the edges in the constructed causal graph represent causal logic, and are assigned explanatory weights based on clinical criteria to quantify the contribution of corresponding imaging features to diagnosis.

[0013] In some embodiments, an error message is also output when the comparison does not match.

[0014] According to another aspect of the present invention, an electronic device is also provided, comprising: At least one processor; At least one memory used to store processor-executable instructions; The at least one processor executes the executable instructions to implement the AI ​​strabismus diagnosis and interpretation method based on causality and large models as described above.

[0015] According to another aspect of the present invention, a computer program product is also provided, comprising a computer program or instructions that, when executed by a processor, implement the AI ​​strabismus diagnosis and interpretation method based on causality and large model as described above.

[0016] Implementing one of the above-described technical solutions of the present invention has the following advantages or beneficial effects: This invention employs a two-step strategy: pre-constructing a logical chain using a causal reasoning engine and then having a large model perform language synthesis. This effectively avoids the "illusion" that might arise from solely using a large model, ensuring that the explanation faithfully reflects both the AI's decision-making evidence and the objective facts of clinical medicine. Therefore, this method generates not just a simple list of features, but a deep explanation containing a complete causal logical chain, highly consistent with the diagnostic thinking of clinicians and enhancing professionalism. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart illustrating an AI strabismus diagnosis and interpretation method based on causality and a large model, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of an architecture for constructing an AI behavioral knowledge base using a clinical knowledge-driven graph neural network, according to an embodiment of the present invention. Figure 3 This is a schematic diagram of an electronic device structure according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments, illustrating various exemplary embodiments that may be used to implement the present invention. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of the present invention disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of the present invention.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," etc., indicate the orientation or positional relationship based on the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the referred element must have a specific orientation, or be constructed and operated in a specific orientation. The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The term "multiple" means two or more. The terms "connected" and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, integral connections, mechanical connections, electrical connections, communication connections, direct connections, indirect connections through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more of the related listed items. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0020] To illustrate the technical solution described in this invention, specific embodiments are described below, showing only the parts related to the embodiments of this invention.

[0021] Example 1: like Figure 1 As shown, this invention provides an AI-based strabismus diagnosis and explanation method based on causality and a large model, including: S100. Obtain the AI ​​diagnostic conclusion of the case to be explained, and simultaneously extract the SHAP value of the case to be explained and the features corresponding to the SHAP value.

[0022] In some embodiments, the SHAP value includes the SHAP local contribution score, the baseline value, and / or the sequence symbol corresponding to the feature sorting.

[0023] It should be noted that the features corresponding to the SHAP value include, but are not limited to: iris center point distance, left eye angle, right eye angle, and binocular distance difference.

[0024] Understandably, AI diagnostic conclusions can be obtained through AI agents, such as diagnostic robots and other intelligent devices. SHAP values ​​and their corresponding features are obtained using existing SHAP methods, and will not be elaborated upon here.

[0025] S200. Using the AI ​​diagnostic conclusion as an index, retrieve the typical feature combination of the strabismus subtype corresponding to the AI ​​diagnostic conclusion from the AI ​​behavior knowledge base.

[0026] like Figure 2As shown, in some embodiments, the construction of the AI ​​behavioral knowledge base can be implemented using a Clinical Knowledge-Driven Graph Neural Network (CK-GNN) model architecture, which includes a Knowledge-Guided Interaction Layer and a Causal Bottleneck Layer connected sequentially. Specifically, The construction of the AI ​​behavior knowledge base includes: The steps performed in the knowledge-guided interaction layer include: constructing graph nodes from eye position image features or eye position coordinate features; generating a clinical adjacency matrix based on strabismus clinical diagnosis rules (refer to existing technologies); using the clinical adjacency matrix as the initial topology for message passing in the graph neural network; initializing edge weights based on the clinical correlation strength between different eye position nodes; adaptively updating edge connections during model training based on edge weights, attention coefficients, and / or feature contributions; and dynamically pruning edges below a preset threshold.

[0027] In a specific embodiment, the "clinically relevant eye positions, eye features, or measurement indicators" that the doctor deems relevant are pre-written into a relationship table. This table informs the GNN which nodes should prioritize information transmission and which nodes have weaker relationships or should not be connected. In a typical GNN, the adjacency matrix usually indicates whether nodes in the graph are connected; however, in this embodiment, the clinical adjacency matrix is ​​not an arbitrary connection but a pre-designed medical relationship structure based on strabismus diagnosis experience.

[0028] It should be noted that dynamic pruning of invalid edges can reduce noise interference and enhance causal consistency.

[0029] To put it simply, embedding GNN (Graph Neural Network) message passing can improve the interpretability and decision rigidity of AI models. The aforementioned clinical association strength can be represented by the angular difference weight between the upper and lower gaze points. This weight can be initially assigned a value by clinical rules and then corrected by labeled case statistics and model training results.

[0030] Furthermore, the steps performed at the causal bottleneck layer include: Before clustering statistics, after the GNN message passing layer and before the final Softmax classification layer, interneurons that explicitly map clinical indicators (such as eye position angle difference and left-right eye coordination indicators) are inserted. Through causal consistency loss, the CK-GNN model is forced to calculate the indicators corresponding to the interneurons before making the final diagnosis.

[0031] In a specific embodiment, during model training, a causal consistency loss is constructed based on clinical indicators obtained through manual annotation or rule calculation. Specifically, firstly, a set of clinical indicators related to strabismus subtypes is determined according to strabismus clinical diagnosis rules. The set of clinical indicators includes the horizontal deviation in the first eye position, the difference in horizontal deviation between vertical and horizontal fixation positions, the difference in horizontal deviation between left and right fixation positions, the difference in vertical deviation, the offset of the corneal light-reflection point, the relative offset of the pupillary center, and / or binocular motion coordination indicators. Secondly, the set of clinical indicators is mapped to multiple intermediate neurons in the causal bottleneck layer, so that each intermediate neuron outputs a corresponding clinical indicator. Then, based on the manually annotated clinical measurements or rule indicators calculated from eye position coordinates, the difference between the output value of the intermediate neuron and the true value of the corresponding clinical indicator is calculated, and this difference is used as the causal consistency loss. Finally, the causal consistency loss and the strabismus subtype classification loss are jointly optimized so that while the model predicts strabismus subtypes, its internal intermediate representation remains consistent with the causal path of clinical diagnosis.

[0032] The true values ​​of the above clinical indicators are either manually annotated clinical measurements or regular indicators calculated from eye position coordinates.

[0033] It should be noted that the causal consistency loss is introduced to ensure that the CK-GNN model calculates intermediate indicators before making a final diagnosis, so that the reasoning path is consistent with the doctor's thinking.

[0034] The trained CK-GNN model is analyzed using SHAP (Shadows and Pulses) or Integrated Gradients (Interpretable Integral Gradients) methods on labeled datasets. Based on these analyses, the missing features corresponding to classification errors in the CK-GNN model are identified. To make this easier to understand, this step not only records positively contributing features but also analyzes counterfactual SHAPs. That is, it identifies "which features, when missing, cause the model to misclassify," thereby determining the rigidly dependent features of the decision.

[0035] Furthermore, for each strabismus subtype, the typical feature combination corresponding to the highest SHAP value is extracted by clustering statistics using the trained CK-GNN model.

[0036] It should be noted that, for each strabismus subtype, cluster statistics are used to extract the "typical high SHAP value feature combination" that the model relies on most (e.g., the model's judgment of esotropia mainly depends on the difference in eyeball centroid coordinates and the changes in GNN topological connections).

[0037] In this embodiment, the typical feature combination refers to a set of key image / eye position features that the AI ​​model repeatedly relies on in a large number of labeled cases for a specific strabismus subtype, has a consistently high SHAP contribution value, and conforms to clinical diagnostic logic. The typical feature combination includes the difference between exotropia in downward gaze and exotropia in upward gaze, exotropia in downward gaze, the degree of corneal light point lateralization in downward gaze, the horizontal difference in the center of the pupils of both eyes in downward gaze, and the change in graph connectivity weights between nodes in the upper and lower eye positions.

[0038] S300. Compare the typical feature combination with the feature with the highest SHAP value of the case to be explained. Take the feature corresponding to the matching SHAP value as the starting point and perform a search in the clinical causal knowledge base to find a logical chain from the starting point to the diagnostic conclusion.

[0039] In a specific embodiment, if the AI ​​diagnosis of the case to be explained is Type A exotropia, then using "Type A exotropia" as an index, the typical feature combination corresponding to this subtype is retrieved from the AI ​​behavior knowledge base. Subsequently, the features with the highest SHAP values ​​in the case to be explained are extracted and compared with the typical feature combination. The comparison process includes feature name or semantic alignment, feature direction consistency verification, and contribution threshold verification.

[0040] For example, in the case to be explained, the exotropia in the downward gaze position is 32△, and the exotropia in the upward gaze position is 16△, with a difference of 16△. The SHAP value of this difference for the feature is 0.31, ranking first among the SHAP features of the case to be explained. Since this feature belongs to the typical feature combination of type A exotropia, and its direction satisfies the clinical diagnostic logic of "exotropia in the downward gaze position is greater than exotropia in the upward gaze position," and its SHAP value exceeds the preset contribution threshold, this feature is determined to be consistent with the typical feature combination of type A exotropia.

[0041] Further, a combined matching score is calculated based on the number of typical features that are successfully matched, feature weights, and directional consistency. If the combined matching score is greater than or equal to a preset matching threshold, it is determined that the AI ​​decision-making path of the case to be explained is consistent with the typical AI behavior path of this strabismus subtype, and the matching feature with the highest SHAP value and clear clinical semantics is selected as the starting point for retrieval in the clinical causal knowledge base. In this embodiment, "the difference between exotropia in the downward gaze position and exotropia in the upward gaze position is 16△" is used as the retrieval starting point. Subsequently, in the clinical causal knowledge base, a path retrieval is performed with this retrieval starting point as the cause node and "Type A exotropia" as the conclusion node, obtaining the following logical chain: the exotropia in the downward gaze position is significantly greater than the exotropia in the upward gaze position, causing the difference in horizontal deviation between the upward and downward gaze positions to exceed the threshold for determining Type A exotropia, thus conforming to the diagnostic logic of Type A exotropia, ultimately pointing to the diagnostic conclusion of Type A exotropia. This logical chain is output in a structured manner and serves as the factual basis and reasoning skeleton for the subsequent generation of explanatory text by the large language model.

[0042] If the above comparisons are consistent, meaning the confidence level is high, then it can be determined that the AI ​​decision-making follows the model's usual reasonable path.

[0043] A mismatch indicates low confidence. When a mismatch occurs, an anomaly message is output, indicating that the interpretation may be biased or that the model adopted a special decision path for this case.

[0044] In this step, high SHAP value image features that pass verification (such as "abnormal difference in horizontal skew angle between downward and upward gaze positions") are used as the starting point. In the clinical causal knowledge base, a shortest path search or conditional traversal is performed to find a logically coherent path to the diagnostic conclusion node. The found path is then solidified into a logical chain and output in a structured manner.

[0045] In some embodiments, the construction of a clinical causality knowledge base includes: Automated entity extraction and attribute parsing are performed on strabismus diagnosis and treatment imaging data. The parsed terms are aligned to medical standard ontology, and the causal relationship direction is defined. The imaging features corresponding to the aligned terms are used as cause nodes, and the diagnostic conclusions are used as result nodes to obtain the constructed causal graph.

[0046] In specific embodiments, pre-trained medical models such as Clinical-Longformer or BioBERT are used to perform automated entity extraction (NER) and attribute parsing on strabismus diagnosis and treatment imaging data, such as authoritative strabismus diagnosis and treatment guidelines and textbooks. The parsed terms are aligned (e.g., using dictionary matching alignment, rule template alignment, semantic vector similarity alignment, ontology encoding alignment, etc.) to medical standard ontology such as SNOMED CT or ICD-11, and causal relationship directions are defined.

[0047] Furthermore, in the constructed causal graph, the layers corresponding to the nodes include the observation layer (such as...). The data is categorized into three layers: difference, mechanism (e.g., oblique hypertrophy), and conclusion (e.g., A-pattern (Type A strabismus)). The edges represent causal logic, and explanatory weights are assigned based on clinical criteria to quantify the contribution of corresponding imaging features to the diagnosis.

[0048] An example of a core rule: "If (If the difference in horizontal skewness between the upward and downward gaze positions is greater than 10°, the A-pattern diagnostic logic will be activated.)

[0049] S400: The found logical chain is used as the instruction and context of the large language model, and the final interpreted text is output through the large language model.

[0050] This step leverages the capabilities of Large Language Models (LLMs) to transform machine-readable logical chains into human-readable professional text. Specifically, it includes: (1) Constructing a controlled context: The structured causal reasoning chain generated in step S300 is used as the core facts and logical skeleton.

[0051] (2) Calling the large language model: The logical skeleton is used as a strict instruction and context and input into a large language model. The task of LLM is limited to "language polishing and organization", that is, it must strictly follow each link of the input logical chain, "translate" and organize it into a fluent, professional text that conforms to the doctor's language habits, and cannot add any information outside the logical chain or make free play.

[0052] (3) Output final explanation: Output a complete explanation text containing "diagnostic conclusion", "key evidence" and "detailed reasoning process".

[0053] Based on the above embodiments, a specific example output is: "Based on the model, the patient is diagnosed with Type A strabismus (A-pattern). Key evidence captured shows that the difference in horizontal strabismus angle between the patient's upward and downward gaze reaches 15°. In clinical logic, this significant cross-directional difference reflects the asymmetry of the patient's extraocular muscle function, meeting the gold standard for clinical diagnosis of the A-pattern. In summary, this AI decision demonstrates extremely high professional reliability."

[0054] In summary, this method generates not just a simple list of features, but a deep explanation containing a complete causal logic chain, which is highly consistent with the diagnostic thinking of clinicians and enhances professionalism. The "two-step" strategy of pre-constructing the logic chain through a causal inference engine and then having a large model perform language synthesis effectively avoids the "illusion" that may arise from using a large model alone, ensuring that the explanation is both faithful to the AI's decision-making evidence and consistent with the objective facts of clinical medicine.

[0055] Furthermore, automating the entire process from obtaining AI diagnostic results to generating professional interpretation reports frees doctors from the tedious task of interpreting AI results, significantly improving clinical work efficiency. By providing a transparent and traceable reasoning process, it greatly enhances clinicians' trust in AI diagnostic conclusions, removing key obstacles to the widespread application of AI technology in clinical practice.

[0056] Example 2: like Figure 3 As shown, this embodiment also provides an electronic device, including: One or more processors; One or more memories are used to store one or more computer programs, and one or more processors are used to execute the one or more computer programs stored in the memories, so that the one or more processors perform the features / steps of an AI strabismus diagnosis and interpretation method based on causality and large models as described in Embodiment 1.

[0057] Furthermore, at the hardware level, the device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other necessary hardware. One or more embodiments of this application can be implemented in software, for example, by the processor reading the corresponding computer program from the non-volatile memory into memory and then running it.

[0058] Of course, in addition to the software implementation, one or more embodiments of this application do not exclude other implementation methods, such as logic devices or a combination of hardware and software. That is to say, the execution subject of the above processing flow is not limited to each logic module, but can also be hardware or logic devices.

[0059] Example 3: As will be understood by those skilled in the art, all or part of the features / steps of the above method embodiments can be implemented by methods, data processing systems, or computer programs. These features can be implemented without hardware, entirely in software, or in a combination of hardware and software. The aforementioned computer program can be stored in one or more computer-readable storage media. When the computer program is executed (e.g., by a processor), it performs the steps of the above-described embodiment of an AI strabismus diagnosis and interpretation method based on causality and a large model.

[0060] The aforementioned storage media capable of storing program code include: solid-state drives, random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), optical storage devices, magnetic storage devices, flash memory, magnetic disks or optical disks and / or combinations thereof, that is, they can be implemented by any type of volatile or non-volatile storage devices or combinations thereof.

[0061] Example 4: This example provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps of the AI ​​strabismus diagnosis and interpretation method based on causality and large models as described in Example 1.

[0062] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the processes of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0063] It should be noted that the method in this embodiment is the same as that in Embodiment 1, and any technical features not described herein are described in Embodiment 1.

[0064] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0065] The above description is merely a preferred embodiment of the present invention. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An AI-based strabismus diagnosis and interpretation method based on causality and a large model, characterized in that, include: Obtain the AI ​​diagnostic conclusion of the case to be explained, and simultaneously extract the SHAP value of the case to be explained and the features corresponding to the SHAP value; Using the AI ​​diagnostic conclusion as an index, retrieve the typical feature combination of the strabismus subtype corresponding to the AI ​​diagnostic conclusion from the AI ​​behavior knowledge base; The typical feature combination is compared with the feature with the highest SHAP value of the case to be explained. The feature corresponding to the matching SHAP value is used as the starting point to perform a search in the clinical causal knowledge base to find a logical chain from the starting point to the diagnostic conclusion. The found logical chain is used as the instruction and context of the large language model, and the final interpreted text is output through the large language model. The construction of the AI ​​behavior knowledge base includes: Eye position image features or eye position coordinate features are constructed as graph nodes. A clinical adjacency matrix is ​​generated according to the clinical diagnosis rules of strabismus. The clinical adjacency matrix is ​​used as the initial topology for message passing in the graph neural network. The edge weights are initialized according to the clinical correlation strength between different eye position nodes. During the training of the CK-GNN model, the edge connection relationship is adaptively updated according to the edge weights, attention coefficients and / or feature contributions. Edges below a preset threshold are dynamically pruned. Collect the SHAP analysis results or interpretable integral gradient algorithm analysis results of the trained CK-GNN model on the labeled dataset, and identify the missing features corresponding to the classification error of the CK-GNN model based on the analysis results; For each strabismus subtype, the CK-GNN model, after training, is used to perform clustering statistics to extract the typical feature combination corresponding to the highest SHAP value; Before clustering statistics, an intermediate neuron that explicitly maps clinical indicators is inserted after the GNN message passing layer and before the final Softmax classification layer. Through causal consistency loss, the CK-GNN model is forced to calculate the indicator corresponding to the intermediate neuron before making the final diagnosis. The causal consistency loss is obtained through the following steps: The set of clinical indicators related to strabismus subtypes is determined according to the clinical diagnostic rules for strabismus. The set of clinical indicators includes the horizontal deviation in the primary gaze position, the difference in horizontal deviation in the vertical gaze position, the difference in horizontal deviation in the left and right gaze position, the difference in vertical deviation, the offset of the corneal light point, the relative offset of the pupil center, and / or the binocular motion coordination index. The clinical indicator set is mapped to multiple intermediate neurons in the causal bottleneck layer, so that each intermediate neuron outputs a corresponding clinical indicator. Based on manually labeled clinical measurements or regular indicators calculated from eye position coordinates, the difference between the output value of the interneuron and the true value of the corresponding clinical indicator is calculated, and this difference is used as the causal consistency loss. The causal consistency loss and the strabismus subtype classification loss are jointly optimized.

2. The AI ​​strabismus diagnosis and interpretation method based on causality and a large model as described in claim 1, characterized in that, The SHAP value includes the SHAP local contribution score, the baseline value, and / or the sequence symbol corresponding to the feature ranking; The typical feature combination includes the difference between exotropia in downward gaze and exotropia in upward gaze, exotropia in downward gaze, the degree of outward shift of the corneal light-reflecting point in downward gaze, the horizontal difference between the pupil centers of the two eyes in downward gaze, and the change in graph connectivity weight between nodes in the upper and lower eye positions.

3. The AI ​​strabismus diagnosis and interpretation method based on causality and a large model as described in claim 1, characterized in that, The construction of the clinical causality knowledge base includes: Automated entity extraction and attribute parsing are performed on strabismus diagnosis and treatment imaging data. The parsed terms are aligned to medical standard ontology, and the causal relationship direction is defined. The imaging features corresponding to the aligned terms are used as cause nodes, and the diagnostic conclusions are used as result nodes to obtain a causal graph.

4. The AI ​​strabismus diagnosis and interpretation method based on causality and a large model as described in claim 3, characterized in that, In the constructed causal graph, the layers corresponding to the nodes include the observation layer, the mechanism layer, and the conclusion layer.

5. The AI ​​strabismus diagnosis and interpretation method based on causality and a large model as described in claim 3, characterized in that, In the constructed causal graph, the edges represent causal logic, and explanatory weights are assigned according to clinical criteria to quantify the contribution of corresponding imaging features to diagnosis.

6. An electronic device, characterized in that, include: At least one processor; At least one memory used to store processor-executable instructions; The at least one processor implements an AI strabismus diagnosis and interpretation method based on causality and large models as described in any one of claims 1 to 5 by running the executable instructions.

7. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement an AI strabismus diagnosis and interpretation method based on causality and a large model as described in any one of claims 1 to 5.

Images