Knee osteoarthritis image intelligent analysis teaching system based on computer vision

Abstract

The invention relates to the technical field of computer-aided teaching and medical image recognition, and particularly discloses an intelligent knee osteoarthritis image analysis teaching system based on computer vision. The system comprises an image data management module, a student interaction and behavior acquisition module, a multi-dimensional skill quantitative analysis module, a personalized knowledge graph construction and gap analysis module, a self-adaptive teaching path planning module and a teaching feedback and report generation module. By collecting and quantitatively analyzing operation behaviors and sight tracks of students, a personal knowledge graph is constructed to accurately position skill defects, and a personalized teaching path is planned according to the skill defects, so that refined evaluation and adaptive teaching guidance of image analysis practical skills are realized.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided teaching and medical image recognition technology, specifically involving a computer vision-based intelligent image analysis teaching system for knee osteoarthritis. Background Technology

[0002] Artificial intelligence (AI) technology is increasingly being applied in the healthcare field, showing great potential, especially in the intelligent analysis and assisted diagnosis of medical images. Computer vision, as a core branch of AI, uses deep learning models to automatically extract and analyze key features in medical images, providing a technological foundation for improving diagnostic efficiency and accuracy.

[0003] Clinical skills teaching and assessment based on medical imaging is an important direction in medical education. Traditional teaching models rely on teachers explaining standard case images, with students learning disease characteristics through observation and memorization. The core objective is to help students master the ability to identify pathological signs from images and make correct diagnoses.

[0004] Using standardized image databases for teaching and assessing students through multiple-choice or short-answer questions has significant limitations: its assessment method relies heavily on static, pre-defined answer options, failing to effectively capture and quantify students' dynamic operational skills, observational pathways, and diagnostic reasoning logic during actual image interpretation, resulting in superficial and one-sided skills assessment. Furthermore, uniform teaching content and assessment standards ignore individual differences in students' knowledge mastery and skill weaknesses, making it difficult to achieve personalized training tailored to individual needs. These problems are particularly pronounced in teaching diseases like knee osteoarthritis, which rely heavily on detailed image interpretation, making it difficult to accurately measure and effectively improve teaching effectiveness. Summary of the Invention

[0005] The purpose of this invention is to provide a computer vision-based intelligent image analysis teaching system for knee osteoarthritis, in order to solve the problems in existing teaching assessment methods that cannot quantify dynamic operational skills and diagnostic reasoning logic, and that make it difficult to achieve personalized teaching guidance.

[0006] The technical solution of the present invention includes: The image data management module is used to store and manage a teaching image database containing knee joint images, image annotation information, standard diagnostic reports, and image metadata. The student interaction and behavior acquisition module is used to present the knee joint images to be analyzed in the graphical user interface and to collect the student's operation behavior sequence, gaze focus trajectory and timestamp information on the interactive interface in real time. The multi-dimensional skills quantification analysis module is used to receive raw behavioral data collected by the student interaction and behavior collection module, and to conduct multi-dimensional quantitative evaluation of students' image analysis skills based on the preset quantification model, generating a skills quantification report that includes operational standardization score, observation integrity score, and diagnostic logic consistency score. The personalized knowledge graph construction and gap analysis module is used to dynamically construct a student's personal knowledge graph that represents the individual student's cognitive state based on standard knowledge in the skills quantification report and teaching video database. By comparing it with the standard expert knowledge graph, the module can calculate the student's knowledge weaknesses and skill deficiencies. The adaptive teaching path planning module receives the gap analysis results output by the personalized knowledge graph construction and gap analysis module, and generates a personalized sequence of subsequent learning content and training task instructions for the student based on the preset teaching resource library and teaching strategy rule library. The teaching feedback and report generation module integrates skills quantification reports and personalized teaching path planning instructions to generate visual learning reports and drive real-time feedback through the graphical user interface.

[0007] Furthermore, the multi-dimensional skill quantification analysis module includes an operational standardization analysis unit, an observation integrity analysis unit, and a diagnostic logic consistency analysis unit.

[0008] The operation standardization analysis unit is used to parse the operation behavior sequence, which consists of a series of discrete interactive actions. The operation standardization analysis unit compares each interactive action with the standard action template in the preset standard operation procedure library, calculates the action matching degree and records the action sequence deviation degree. Finally, the operation standardization score is determined by the weighted average of the action matching degree and the penalty coefficient of the action sequence deviation degree.

[0009] The observation integrity analysis unit is used to analyze the gaze focus trajectory, which consists of a series of coordinate points in the image coordinate system and their corresponding timestamps. The observation integrity analysis unit performs spatiotemporal correlation analysis on the gaze focus trajectory and the key pathological area annotation information of the image in the teaching image database. It calculates the percentage of total dwell time of the gaze in the key pathological areas and identifies whether there are key areas not covered by the gaze. Finally, the observation integrity score is determined by the percentage of total dwell time and the number of key areas not covered.

[0010] The diagnostic logic consistency analysis unit is used to correlate and analyze the sequence of operational behaviors, the trajectory of gaze focus, and the diagnostic conclusion text submitted by the student. This unit first extracts the set of image features that the student actually focuses on from the behavior sequence and gaze trajectory, and then uses natural language processing technology to parse the set of inferred features that the student relies on from the diagnostic conclusion text. Next, it calculates the intersection ratio between the actual focus feature set and the inferred feature set, and analyzes whether the reasoning chain from the focus features to the diagnostic conclusion conforms to a pre-defined medical diagnostic logic rule base. Finally, the diagnostic logic consistency score is jointly determined by the intersection ratio of the feature sets and the degree to which the reasoning chain conforms to the logical rules.

[0011] Furthermore, the personalized knowledge graph construction and gap analysis module performs the following process: This module maps the scores in the skill quantification report to the mastery level values ​​of the corresponding nodes in the knowledge graph. Nodes represent specific pathological sign recognition skills or diagnostic reasoning rules. Simultaneously, based on the student's historical performance data for the same type of node in multiple training sessions, this module uses an exponentially weighted moving average algorithm to update the mastery level value of that node, thereby dynamically constructing and updating the student's personal knowledge graph. The standard expert knowledge graph is predefined by domain experts and includes all knowledge nodes required for knee arthritis image analysis and the logical relationships between nodes. Gap analysis is achieved by comparing the mastery level values ​​of the corresponding nodes in the student's personal knowledge graph with those in the standard expert knowledge graph node by node. When the mastery level value of a node is lower than a preset mastery threshold, that node is marked as a knowledge weakness. At the same time, the logical paths associated with weak nodes are analyzed to identify the core skill deficiencies that lead to the weakness.

[0012] Furthermore, the adaptive teaching path planning module includes a teaching resource indexing unit and a strategy rule engine.

[0013] The teaching resource index unit is used to manage a structured teaching resource repository. Each resource in the repository is labeled with the knowledge node tag it is for, the resource type, and the difficulty level.

[0014] The strategy rule engine is used to load a pre-defined teaching strategy rule base. This rule base contains a series of production rules, each consisting of a precondition and an action. The preconditions are set based on the type, severity, and historical learning records of knowledge gaps and skill deficiencies in the gap analysis results. The action is to retrieve and combine teaching resources with specific tags, types, and difficulty levels from the teaching resource index unit to form an ordered sequence of learning content and training tasks.

[0015] Furthermore, the rule priorities in the teaching strategy rule base are dynamically adjusted according to the severity of the deficiencies. For deficiencies corresponding to key knowledge nodes whose mastery level is below the first threshold, a high-priority rule is triggered, and the training task instruction generated by this rule will be forcibly inserted at the forefront of the current learning process. For deficiencies corresponding to non-key knowledge nodes whose mastery level is between the first and second thresholds, a medium-priority rule is triggered, and the training task generated by this rule will be arranged as recommended content in subsequent learning. For knowledge nodes whose mastery level is above the second threshold, a consolidation rule is triggered, and this rule will arrange spaced repetitive training tasks to prevent skill forgetting.

[0016] Furthermore, the system operates within a closed-loop workflow that includes four core components: learning, evaluation, feedback, and planning.

[0017] During the learning phase, the student interaction and behavior collection module presents teaching videos and collects behavioral data.

[0018] During the assessment phase, the multi-dimensional skills quantification analysis module processes the behavioral data and generates a skills quantification report.

[0019] During the feedback phase, the teaching feedback and report generation module instantly presents students with key scores and visual analysis charts from the skills quantification report.

[0020] In the planning stage, the personalized knowledge graph construction and gap analysis module and the adaptive teaching path planning module work together based on the assessment results to generate and update the learning plan for the next stage, thereby driving the learning process into a new round of iteration.

[0021] Furthermore, the system also includes a teaching effectiveness tracking and model optimization module.

[0022] This module is used to collect and store anonymized skill quantification reports, personal knowledge graph evolution sequences, and final teaching outcome data for all students over a long period of time. The module regularly performs statistical analysis on the collected data and calculates the effectiveness indicators of different teaching strategy rules for improving specific types of skill deficiencies. When the effectiveness indicator of a certain teaching strategy rule is consistently lower than the preset optimization threshold, the system will issue an alert to the administrator, prompting the rule to be reviewed and optimized.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention captures students' operational behavior sequences and gaze trajectory in real time during image analysis through a student interaction and behavior acquisition module. This data is then analyzed in depth by a multi-dimensional skills quantification module, transforming previously invisible and dynamic image interpretation skills into quantifiable scores for operational standardization, observational completeness, and diagnostic logic consistency. This assessment method transcends traditional static, right-or-wrong-answer-based evaluation models, providing a comprehensive and refined measurement of students' practical skills across three core dimensions: operational process, observation path, and reasoning logic. This enables a deep assessment of clinical image analysis capabilities.

[0024] 2. This invention, through a personalized knowledge graph construction and gap analysis module, maps and integrates discrete skill scores into a structured student-specific knowledge graph. By comparing it with standard expert knowledge graphs, it accurately identifies each student's unique knowledge weaknesses and skill deficiencies. This process enables dynamic modeling and diagnosis of individual student cognitive states, transforming instructional intervention from uniform, extensive content delivery to precise, targeted training addressing specific individual weaknesses. This provides a reliable technological foundation for truly individualized instruction.

[0025] 3. This invention, through its adaptive teaching path planning module and built-in strategy rule engine, can automatically generate and dynamically adjust subsequent learning content and training tasks based on real-time gap analysis results. The system implements differentiated teaching strategies according to the severity of deficiencies, providing mandatory priority remediation for severe deficiencies, recommended reinforcement for general deficiencies, and arranging consolidation training to prevent forgetting. This constructs a complete intelligent teaching loop of assessment, feedback, planning, and learning, making the teaching system not only a passive assessment tool but also an intelligent tutor that actively guides the learning process and continuously optimizes teaching effectiveness, significantly improving the efficiency and personalization of teaching knee osteoarthritis image analysis. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-dimensional skill quantification analysis module in this invention; Figure 3 This is a logical flowchart of the personalized knowledge graph construction and gap analysis module in this invention; Figure 4 This is a schematic diagram illustrating the interaction and data flow between the adaptive teaching path planning module, the teaching resource library, and the strategy rule library in this invention. Figure 5 This is a schematic diagram of the closed-loop workflow of system learning-evaluation-feedback-planning in this invention. Detailed Implementation

[0027] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 This invention proposes a computer vision-based intelligent image analysis teaching system for knee osteoarthritis. The system consists of six core functional modules: image data management, student interaction and behavior acquisition, multi-dimensional skill quantification analysis, personalized knowledge graph construction and gap analysis, adaptive teaching path planning, and teaching feedback and report generation. These modules communicate and synchronize their states through standardized data interfaces, forming a closed-loop intelligent teaching entity. The system is deployed on a computing terminal with a graphical user interface, supporting local or cloud deployment modes. All data interactions employ encrypted transmission protocols to ensure the security and privacy compliance of medical teaching data.

[0028] The image data management module serves as the system's data foundation, responsible for storing and managing a structured teaching image database. This database contains several knee joint image files, each associated with complete annotation information, a standard diagnostic report, and image metadata. The knee joint images are sourced from an ethically approved clinical imaging archive system, covering osteoarthritis cases at different disease stages (early, middle, and late). Image modalities include anteroposterior and lateral X-ray films, T1 / T2-weighted magnetic resonance imaging (MRI) sequences, and proton density-weighted sequences. Annotation information was independently completed by three radiologists with over 10 years of experience, and the final annotation results were agreed upon at a consensus meeting. The annotations include pixel-level masks or bounding box coordinates of key pathological features such as joint space narrowing areas, osteophyte formation locations, subchondral bone sclerosis extent, cystic area coordinates, and synovial thickening degree. The standard diagnostic report follows the imaging assessment guidelines of the Osteoarthritis Research International Society (OARSI), clearly recording the Kellgren-Lawrence classification, the degree of involvement of various anatomical structures, and key points for differential diagnosis. Image metadata includes technical parameters such as patient anonymity ID, image acquisition device model, slice thickness, window width and level parameters, imaging position, and timestamp. All data undergoes format standardization before being imported into the database, being uniformly converted to DICOM format and indexed to support rapid retrieval by multi-dimensional tags such as pathology type, disease stage, and imaging modality. The image data management module provides a read-only API interface, allowing other modules to access raw images and their associated information as needed, ensuring the consistency and authority of the data source.

[0029] The student interaction and behavior acquisition module presents the knee joint images to be analyzed, provided by the image data management module, in a graphical user interface, and simultaneously initiates a high-precision behavior data capture mechanism. This module integrates an eye-tracking hardware interface and an operation log recorder, enabling real-time acquisition of two core types of behavior data: operational behavior sequences and gaze focus trajectories. Operational behavior sequences consist of all interactive actions performed by the student during image analysis, including but not limited to adjusting image window width and level, switching zoom levels, panning operations, activating measurement tools, drawing annotation tools, confirming pathological signs, and submitting diagnostic conclusions. Each interactive action is recorded as a structured event object, containing an action type identifier, execution timestamp, action coordinates, parameter values ​​(such as zoom factor and window width), and contextual state (such as the currently active tool). Gazing focus trajectories are acquired through an external or embedded eye tracker, with a sampling frequency of at least 60 Hz. The output is a time-series data stream, with each frame containing x and y coordinates in the screen coordinate system, fixation duration, pupil diameter, and confidence score. This module preprocesses the raw eye-tracking data, including noise filtering, drift correction, and gaze point clustering, transforming continuous eye-tracking signals into discrete sequences of gaze events. Each event includes a start time, end time, center coordinates, and coverage area radius. All behavioral data is aligned to a unified time base and appended with a unique identifier for the currently displayed image, forming a complete behavioral log with spatiotemporal context. This log is transmitted in real-time to the multi-dimensional skill quantification analysis module via a secure channel, while a local copy is cached for subsequent retrospective analysis.

[0030] The multi-dimensional skills quantification analysis module receives behavior logs from the student interaction and behavior collection module and, based on a pre-set quantification model, performs a refined assessment of students' image analysis skills across three orthogonal dimensions. Please refer to the appendix. Figure 2 This module is internally divided into an operational standardization analysis unit, an observation integrity analysis unit, and a diagnostic logic consistency analysis unit. The three units process the input data in parallel and output individual scores respectively, and finally merge them to generate a comprehensive skills quantitative report.

[0031] The operational standardization analysis unit first parses the sequence of operational behaviors and dynamically compares it with a pre-defined standard operating procedure library. This library, built by a team of domain experts based on clinical best practices, defines standardized operational step templates for different imaging modalities and disease stages. For example, when analyzing X-ray anteroposterior views, the standard procedure requires the following sequence: global observation of joint alignment → assessment of medial / lateral joint space → examination of tibial plateau and femoral condyle osteophytes → examination of the patellofemoral joint surface → and finally, overall grading. Each step corresponds to a standardized action template, which defines the allowed set of action types, a reasonable parameter range (e.g., window width should be between 800-1200 HU), the expected execution order, and the minimum dwell time. The operational standardization analysis unit uses a dynamic time warping algorithm to align the student's actual operational sequence with the best-matching template and calculates a matching score for each action. The matching score is calculated based on action type consistency, parameter deviation, and time reasonableness, ranging from 0 to 1. Simultaneously, the unit records the action sequence deviation, i.e., the number of times the template order constraints are violated in the actual sequence. Finally, an operational standardization score is generated. Determined by the following formula:

[0032] Where N is the total number of standardized steps, The matching score for the i-th step. The weights are assigned by experts based on clinical importance, d is the normalized order deviation (ranging from 0 to 1), and α is the penalty coefficient, typically set to 0.3. This formula ensures that sequences with high matching scores but serious operational logic violations will still receive significant deductions.

[0033] The observation integrity analysis unit processes gaze focus trajectories, focusing on assessing students' coverage of key pathological regions. This unit first retrieves the annotation information of key pathological regions from the current image data management module and maps it to the screen coordinate system. Subsequently, it performs spatiotemporal overlay analysis on the preprocessed gaze event sequence and these regions. For each key region... Calculate the total duration of the line of sight within it. And sum them up to get the total gaze duration of all key regions. Simultaneously, the number U of key regions not covered by any observed events was counted. Observation integrity score. Defined as:

[0034] in, M represents the total duration of this analysis task, and M represents the total number of critical regions. βThis is the duration weighting coefficient, with a typical value of 0.6. This design balances gaze depth (time percentage) and gaze breadth (coverage integrity), preventing students from simply scanning all areas quickly without in-depth observation.

[0035] The diagnostic logic consistency analysis unit performs cross-modal correlation analysis, integrating operational behaviors, gaze trajectories, and the diagnostic conclusion text submitted by the student. This unit first extracts the set of actual attention features from the behavioral data. By analyzing the usage records of annotation tools, the application locations of measurement tools, and the areas of prolonged fixation, pathological signs that students explicitly or implicitly focused on were identified, such as medial joint space narrowing and femoral condyle osteophytes. Each feature was accompanied by a confidence level. Secondly, a pre-trained medical natural language processing model was used to semantically parse the diagnostic conclusion text, extracting a set of inference features. This refers to the visual evidence that students claim to use to support their conclusions. Then, the ratio of the intersection size to the union size of the two sets is calculated to obtain the feature consistency index. In addition, the unit loads a pre-defined medical diagnostic logic rule library, which is stored in the form of a knowledge graph. Nodes represent pathological signs and diagnostic conclusions, and edges represent logical implications (e.g., severe joint space stenosis + osteophyte formation → KL grade 3). Unit verification starts from... Does the reasoning path to the diagnostic conclusion exist in the rule base? If so, what is the logical conformity? A score of 1 is given; otherwise, partial compliance is calculated based on the shortest edit distance. Finally, a diagnostic logical consistency score is determined. , The typical value is 0.5.

[0036] The three scores are normalized and then weighted and merged to generate a skills quantification report, which includes scores for each dimension, radar chart visualization, and details of specific points lost.

[0037] The personalized knowledge graph construction and gap analysis module receives skills quantification reports and dynamically constructs individual knowledge graphs for each student by combining them with the standard knowledge system in the teaching video database. Please refer to the attached document. Figure 3 The core of this module lies in mapping discrete skill scores to structured cognitive states. Nodes in the knowledge graph represent specific pathological sign recognition abilities (such as recognizing subchondral cysts) or diagnostic reasoning rules (such as the mapping relationship between joint space narrowing and KL grading), while edges represent dependencies or combinatorial relationships between concepts. Each node... Associated with a level of mastery value The initial value is directly assigned from the initial evaluation result. In subsequent training iterations, if the student encounters tasks related to the same node again, the module updates the value using an exponentially weighted moving average algorithm. :

[0038] in, The sub-item score corresponding to this node in the t-th training iteration (extracted from the skill quantification report) is λ, where λ is the learning rate, typically 0.4, to ensure that new evidence has a moderate impact on cognitive state. The standard expert knowledge graph was jointly constructed by five senior radiologists, covering all 128 core knowledge points required for knee osteoarthritis image analysis, with a preset mastery threshold for each node. (Typically 0.85). The gap analysis process traverses all nodes, if... If the value is set to 0.7, it is marked as a knowledge gap. Further, a graph traversal algorithm is used to analyze the upstream dependent nodes of the gap to identify the root cause of the defect (e.g., an incorrect KL classification due to failure to identify early osteophytes). This module outputs structured gap analysis results, including a list of gaps, defect type (perceptual / reasoning), severity level, and associated paths.

[0039] Based on the gap analysis results above, the adaptive learning path planning module generates personalized follow-up learning instructions. Please refer to the appendix. Figure 4 This module consists of a teaching resource index unit and a strategy rule engine. The teaching resource library contains resources such as video tutorials, interactive cases, 3D anatomical models, comparative image sets, and simulated diagnostic tasks. Each resource is labeled with a target knowledge node tag, resource type (explanation / practice / test), and difficulty level (levels 1-5). The strategy rule engine loads a teaching strategy rule library, which contains several production rules. For example, a high-priority rule is defined as: IF Weakness type = Key pathological sign recognition AND Mastery level < 0.6 THEN Generate a mandatory task: Push an interactive osteophyte recognition training at difficulty level 2, placing it at the top of the learning queue. Rule preconditions can be combined with multiple Boolean expressions, involving defect severity, historical performance trends, and learned resource types. The execution action specifies resource retrieval conditions (tags, types, difficulty) and insertion strategies (forced pre-placement / recommended subsequent placement / interval repetition). The rule priority is dynamically adjusted: when the mastery level of a node is below the first threshold (0.6), a high-priority rule is triggered; when it is between 0.6 and 0.8, a medium-priority rule is triggered; when it is above 0.8, a consolidation rule is triggered, and repeated training is arranged after 7 days and 14 days to counteract the forgetting curve. The engine sorts the rules according to priority, matches them in sequence, and generates an ordered task sequence.

[0040] The teaching feedback and report generation module integrates skills quantification reports and teaching path instructions, driving multimodal feedback through the graphical user interface. Students can see a dynamically generated learning report within 3 seconds of submitting a diagnosis. This report includes a 3D radar chart displaying the three-dimensional score, a heatmap overlay showing blind spots, a flowchart highlighting incorrect steps, and natural language-generated improvement suggestions (e.g., "You haven't fully observed the patellofemoral joint surface; we suggest strengthening your recognition training in this area"). Simultaneously, the right side of the interface displays the next stage's learning task card, with high-priority tasks highlighted in red.

[0041] The entire system runs on the attached Figure 5 The learning-assessment-feedback-planning closed-loop workflow is shown below. In the learning phase, students encounter new images and generate behavioral data; in the assessment phase, quantitative reports are generated in real time; in the feedback phase, results are presented instantly; and in the planning phase, the knowledge graph is updated and new tasks are generated, seamlessly connecting to the next round of learning. Each iteration of this closed loop takes no more than 5 minutes, supporting high-intensity, high-frequency skills training.

[0042] In addition, the system includes a teaching effectiveness tracking and model optimization module. This module aggregates anonymized data from all students over a long period, including skill scoring sequences, knowledge graph evolution trajectories, and final assessment pass rates. A strategy effectiveness analysis is performed quarterly: for each teaching rule, the average improvement in students' relevant weak areas after applying the rule is calculated. If a rule has two consecutive cycles If the value falls below the optimization threshold (set to 0.15), the system sends an alert to the administrator, prompting a review of the rule logic or resource quality. This mechanism ensures that teaching strategies continuously evolve to align with real learning patterns.

[0043] Example 2: Building upon Example 1, this example enhances the diagnostic logic consistency analysis unit within the multi-dimensional skill quantification analysis module by introducing a deep diagnostic text parsing mechanism based on a large language model to improve the ability to capture complex reasoning logic. This enhancement is particularly suitable for advanced training scenarios for senior medical students or resident physicians, whose diagnostic conclusions often include multi-level causal chains and differential diagnostic arguments.

[0044] In this embodiment, the diagnostic conclusion text submitted by students is no longer limited to structured options, but supports free text input with a length of up to 500 characters. The diagnostic logic consistency analysis unit first calls a large language model fine-tuned from the medical image diagnostic corpus to perform multi-level semantic parsing on the text. The model first identifies the core diagnostic conclusion in the text (e.g., conforming to Kellgren-Lawrence grade 2 osteoarthritis), then extracts direct evidence supporting the conclusion (e.g., mild medial joint space narrowing and small osteophytes are visible), and further uncovers implicit exclusionary arguments (e.g., no obvious subchondral bone sclerosis is seen, therefore grade 3 is not supported). This process generates a structured diagnostic argument tree, with the root node being the final conclusion, child nodes being direct evidence, and leaf nodes being original image features.

[0045] Meanwhile, the actual attention feature set F_actual, extracted from the sequence of operational behaviors and gaze trajectories, has also been upgraded. In addition to basic pathological signs, the unit also identifies whether students have performed differential diagnosis-related operations, such as switching to a lateral view to rule out patellofemoral joint disease, or adjusting the window level to differentiate between cystic changes and vascular shadows. These advanced operations are mapped to differential intent nodes and associated with corresponding exclusionary features.

[0046] The calculation of diagnostic logical consistency score is therefore expanded to two levels: feature-level consistency and logical-level consistency. Feature-level consistency It still uses set intersection and union comparison, but the set elements include basic characteristics and distinguishing features. Logical layer consistency. Then, a graph isomorphism algorithm is used to compare the student's argument tree with the expected reasoning path in the standard rule base. If the two trees are completely matched in node type and parent-child relationship, then... If there are missing or redundant subtrees, the similarity is calculated based on the size of the largest common subgraph. Final score. It places greater emphasis on the rigor of logical structure.

[0047] This enhancement mechanism enables the system to effectively distinguish between two scenarios: correct answers but incorrect reasoning, and reasonable reasoning but biased conclusions. The former will receive a low logical consistency score, while the latter may receive a higher score, thus guiding students to focus on the clinical reasoning process rather than merely pursuing correct conclusions. The knowledge graph nodes in this embodiment are also expanded accordingly, adding nodes related to differential diagnostic abilities, whose mastery level is updated driven by the logical consistency score. Based on this, the adaptive teaching path planning module pushes cases emphasizing reasoning training, such as providing comparative cases with similar imaging manifestations but different etiologies, strengthening students' differential thinking.

[0048] Example 3: This example focuses on a lightweight deployment solution for the system in resource-constrained environments, particularly suitable for primary healthcare institutions or mobile teaching scenarios. In this solution, eye-tracking hardware is replaced by a camera-based gaze estimation software module, while some calculations for multi-dimensional skill quantification analysis are migrated to edge devices.

[0049] The student interaction and behavior acquisition module integrates a lightweight gaze estimation algorithm, utilizing the device's front-facing camera to capture student facial video streams. This algorithm uses MobileNetV3 as its backbone network to detect pupil center and head posture in real time, combining this with screen geometric parameters to infer the gaze's location on the image. Although the accuracy is slightly lower than professional eye trackers, by increasing the gaze point cluster radius and extending the minimum gaze determination time, it can still effectively identify the coverage of key areas. The acquisition logic for the operational behavior sequence remains unchanged.

[0050] To reduce server load, some functions of the multi-dimensional skill quantification analysis module have been moved to the client side. The operational standardization analysis unit and the observation integrity analysis unit run on local devices, uploading only the calculation results (not the original behavior logs) to the central server. The diagnostic logic consistency analysis unit, which relies on a large language model, remains in the cloud, but uses model distillation technology, replacing the original large model with Tiny BERT, which improves inference speed and reduces memory usage.

[0051] The personalized knowledge graph construction and gap analysis module employs an incremental synchronization strategy: it locally caches the graph snapshots from the last five training iterations, triggering cloud synchronization only when the mastery level value changes by more than 0.1. The adaptive teaching path planning module preloads commonly used teaching resources to local storage, ensuring that basic training tasks can continue to be provided even during network interruptions.

[0052] This embodiment significantly lowers the system deployment threshold by co-optimizing software and hardware, while ensuring core teaching functions, enabling medical students in remote areas to also have an intelligent knee joint image analysis training experience.

Claims

1. A computer vision-based intelligent image analysis teaching system for knee osteoarthritis, characterized in that, include: The image data management module is used to store and manage a teaching image database containing knee joint images, image annotation information, standard diagnostic reports, and image metadata. The student interaction and behavior acquisition module is used to present the knee joint images to be analyzed in the graphical user interface and to collect the student's operation behavior sequence, gaze focus trajectory and timestamp information on the interactive interface in real time. The multi-dimensional skills quantification analysis module is used to receive raw behavioral data collected by the student interaction and behavior collection module, and to conduct multi-dimensional quantitative evaluation of students' image analysis skills based on the preset quantification model, generating a skills quantification report that includes operational standardization score, observation integrity score, and diagnostic logic consistency score. The personalized knowledge graph construction and gap analysis module is used to dynamically construct a student's personal knowledge graph that represents the individual student's cognitive state based on standard knowledge in the skills quantification report and teaching video database. By comparing it with the standard expert knowledge graph, the module can calculate the student's knowledge weaknesses and skill deficiencies. The adaptive teaching path planning module receives the gap analysis results output by the personalized knowledge graph construction and gap analysis module, and generates a personalized sequence of subsequent learning content and training task instructions for the student based on the preset teaching resource library and teaching strategy rule library. The teaching feedback and report generation module integrates skills quantification reports and personalized teaching path planning instructions to generate visual learning reports and drive real-time feedback through the graphical user interface.

2. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 1, characterized in that, The multi-dimensional skills quantification analysis module includes an operational standardization analysis unit, an observation integrity analysis unit, and a diagnostic logic consistency analysis unit. The operation standardization analysis unit is used to parse the operation behavior sequence, which consists of a series of discrete interactive actions. The operation standardization analysis unit compares each interactive action with the standard action template in the preset standard operation process library, calculates the action matching degree and records the action sequence deviation degree. The operation standardization score is jointly determined by the weighted average of the action matching degree and the penalty coefficient of the action sequence deviation degree. The observation integrity analysis unit is used to analyze the gaze focus trajectory, which consists of a series of coordinate points in the image coordinate system and corresponding timestamps. The observation integrity analysis unit performs spatiotemporal correlation analysis between the gaze focus trajectory and the key pathological area annotation information of the image in the teaching image database. It calculates the proportion of the total dwell time of the gaze in the key pathological area and identifies the key areas not covered by the gaze. The observation completeness score is determined by the percentage of total dwell time and the number of key areas not covered; The diagnostic logic consistency analysis unit is used to correlate and analyze the sequence of operational behaviors, the trajectory of gaze focus, and the diagnostic conclusion text submitted by the student. This unit first extracts the set of image features that the student actually focuses on from the behavior sequence and gaze trajectory, and then uses natural language processing technology to parse the set of inferred features that the student relies on from the diagnostic conclusion text. Next, it calculates the intersection ratio between the actual focus feature set and the inferred feature set, and analyzes whether the reasoning chain from the focus features to the diagnostic conclusion conforms to a preset medical diagnostic logic rule base. The diagnostic logic consistency score is jointly determined by the intersection ratio of the feature sets and the degree to which the reasoning chain conforms to the logical rules.

3. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 1, characterized in that, The personalized knowledge graph construction and gap analysis module performs the following process: This module maps the scores in the skill quantification report to the mastery level values ​​of the corresponding nodes in the knowledge graph. Nodes represent specific pathological sign recognition skills or diagnostic reasoning rules. Simultaneously, based on the student's historical performance data for the same type of node in multiple training sessions, the module uses an exponentially weighted moving average algorithm to update the mastery level value of that node, thereby dynamically constructing and updating the student's personal knowledge graph. The standard expert knowledge graph is predefined by domain experts and includes all knowledge nodes required for knee arthritis image analysis and the logical relationships between nodes. Gap analysis is achieved by comparing the mastery level values ​​of the corresponding nodes in the student's personal knowledge graph with those in the standard expert knowledge graph node by node. When the mastery level value of a node is lower than a preset mastery threshold, that node is marked as a knowledge weakness. Simultaneously, the logical paths associated with weak nodes are analyzed to identify the core skill deficiencies leading to the weakness.

4. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 1, characterized in that, The adaptive teaching path planning module includes a teaching resource indexing unit and a strategy rule engine; The teaching resource indexing unit is used to manage a structured teaching resource database, in which each resource is labeled with the knowledge node tag it is for training, the resource type, and the difficulty level. The strategy rule engine is used to load a preset teaching strategy rule base. The rule base contains a series of production rules, each rule consisting of a precondition and an action. The preconditions are set based on the type, severity, and historical learning records of knowledge gaps and skill deficiencies in the gap analysis results. The action is to retrieve and combine teaching resources with specific tags, types, and difficulty levels from the teaching resource index unit to form an ordered sequence of learning content and training tasks.

5. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 4, characterized in that, The rule priority in the teaching strategy rule base is dynamically adjusted according to the severity of the deficiencies. For deficiencies corresponding to key knowledge nodes whose mastery level is below the first threshold, a high-priority rule is triggered, and the training task instruction generated by this rule will be forcibly inserted at the forefront of the current learning process. For deficiencies corresponding to non-key knowledge nodes whose mastery level is between the first and second thresholds, a medium-priority rule is triggered, and the training task generated by this rule will be arranged as recommended content in subsequent learning. For knowledge nodes whose mastery level is above the second threshold, a consolidation rule is triggered, and this rule will arrange spaced repetitive training tasks to prevent skill forgetting.

6. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 1, characterized in that, The system operates within a closed-loop workflow that includes four core components: learning, evaluation, feedback, and planning. During the learning phase, the student interaction and behavior collection module presents teaching videos and collects behavioral data. During the assessment phase, the multi-dimensional skills quantification analysis module processes the behavioral data and generates a skills quantification report; During the feedback phase, the teaching feedback and report generation module instantly presents students with key scores and visual analysis charts from the skills quantification report. In the planning stage, the personalized knowledge graph construction and gap analysis module and the adaptive teaching path planning module work together based on the assessment results to generate and update the learning plan for the next stage, thereby driving the learning process into a new round of iteration.

7. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 1, characterized in that, The system also includes a teaching effectiveness tracking and model optimization module; This module is used to collect and store anonymized skill quantification reports, personal knowledge graph evolution sequences, and final teaching outcome data for all students over a long period of time. The module regularly performs statistical analysis on the collected data and calculates the effectiveness indicators of different teaching strategy rules for improving specific types of skill deficiencies. When the effectiveness indicator of a certain teaching strategy rule is consistently lower than the preset optimization threshold, the system will issue an alert to the administrator, prompting the rule to be reviewed and optimized.

8. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 2, characterized in that, The operational standardization analysis unit uses a dynamic time warping algorithm to align the student's actual operation sequence with the standardized action templates in the standard operation procedure library and calculates the matching score for each action. The operational standardization score is determined by the weighted average of the matching scores of all steps and the result of multiplying the normalized action sequence deviation by the penalty coefficient.

9. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 2, characterized in that, The observation integrity analysis unit calculates the observation integrity score as follows: calculates the proportion of the total fixation time in all key pathological areas to the total time of this analysis task, and counts the number of key areas not covered by any fixation events; The observation integrity score is determined by the aforementioned proportion and the number of uncovered critical areas.

10. The computer vision-based intelligent image analysis teaching system for knee osteoarthritis according to claim 3, characterized in that, The process of updating the mastery level value of a node using the exponentially weighted moving average algorithm is as follows: obtain the sub-item score corresponding to the node in this training, and combine the mastery level value of the node at the previous moment with the preset learning rate to perform iterative updates and obtain the mastery level value of the node at the current moment.

Images