Adenoid facies screening method and device based on three-view detection and medical rule fusion

Abstract

The application discloses an adenoid facial screening method and equipment based on three-view detection and medical rule fusion, and belongs to the technical field of adenoid recognition. Three views of the face of a user are collected, a pre-trained target detection model is used for detection, and a detection result and an original confidence of a specific label are output. Clinical factors are obtained, each clinical factor is converted into a multiplicative coefficient by a rule engine and multiplied, and a risk factor is obtained. Only the abnormal detection results of the front part of the mouth and the left and right lower jaws are subjected to bounded monotonic enhancement. The maximum value of the confidence of the enhanced abnormal category in the three views is taken for threshold grading, a risk grading result is output, and adenoid facial screening is completed. The boundedness, monotonicity and interpretability are cooperated to effectively improve the robustness and clinical interpretability of the adenoid facial screening method, so that the accuracy and interpretability of the screening result of the actual demand of the medical scene are better met, and the risk of missed diagnosis is significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of adenoid recognition technology, specifically to an adenoid facial screening method and device based on the fusion of three-view detection and medical rules. Background Technology

[0002] Upper airway obstruction and mouth breathing syndrome caused by adenoid hypertrophy are common in children, with typical facial features including poor lip closure, mandibular retrusion, and chin morphology changes. Currently, adenoid facies is often detected late, making correction difficult. Therefore, promoting universal screening for adenoid facies in children is an urgent social and health issue that needs to be addressed.

[0003] Clinically, adenoid facies has always been a focus for otolaryngologists and orthodontists. Once adenoid facies is diagnosed, otolaryngologists primarily treat adenoid hypertrophy and chronic nasal inflammation through internal and surgical methods to restore nasal patency. Orthodontists mainly improve the patient's appearance by correcting mouth breathing habits, correcting dentofacial deformities, and combining orthognathic and orthodontic treatment for severe dentofacial deformities. As is well known, all of the above treatments begin with parents bringing their children to the hospital for examination. However, initial screening in outpatient clinics relies on facial examinations and imaging, which suffers from high subjectivity, inconsistent standards, and limited efficiency.

[0004] In addition, most patients with adenoid facies are often not discovered or are discovered very late due to parental neglect or lack of attention, missing the opportunity for early intervention and treatment, thus causing irreversible and typical "adenoid facies". Therefore, it is imperative to carry out early screening for adenoid facies. Early screening, early detection, and early treatment of adenoid facies in children can prevent the decline in the appearance of affected children in the early stages.

[0005] With the development of artificial intelligence, corresponding early screening technologies for adenoid facies have also been developed. For example, invention patent CN115206540A discloses an adenoid facies screening, monitoring, and processing system and method, specifically including: S1: collecting normal facial images and establishing normal value ranges; S2: distributing questionnaires; S3: assigning scores to the questionnaire results; S4: obtaining facial information of the child; S5: using AI to identify the child's facial information and comparing it with normal ranges or parental facial information. This allows for screening of adenoid facies in children without going to the hospital. However, this method is based solely on questionnaire surveys, resulting in low screening accuracy.

[0006] For example, invention patent CN115120227A discloses an adenoid identification system and method based on dynamic pattern recognition. This includes: employing deterministic learning theory, training a training nasal airflow dynamic pattern using an RBF neural network to obtain a constant RBF neural network; constructing a dynamic estimator from the constant RBF neural network; identifying the test nasal airflow dynamic pattern using the dynamic estimator; obtaining the identification result; acquiring the identification error between the identification result and the test nasal airflow dynamic pattern; selecting the training nasal airflow dynamic pattern with a small identification error; statistically analyzing the adenoid types corresponding to the selected nasal airflow dynamic patterns; and identifying the adenoid type with the largest proportion as the adenoid to be identified. This method improves the speed and accuracy of adenoid identification and is easy to operate. However, this method requires the identification of nasal airflow dynamic patterns, making it difficult to implement at home.

[0007] Therefore, developing an early adenoid facies screening method for screening children with adenoid facies is an urgent problem to be solved. Summary of the Invention

[0008] The purpose of this invention is to provide a method and device for screening adenoid facial features based on the fusion of three-view detection and medical rules. After detecting and calculating risk factors based on the three-view facial features and clinical factors, bounded monotonic enhancement is performed only on abnormal categories. Then, the maximum enhancement confidence value is taken in the three-view facial features for threshold classification. Finally, the image-by-image and rule results are output for adenoid facial feature screening.

[0009] To achieve the above-mentioned objectives, an embodiment provides an adenoid facial screening method based on three-view detection and medical rule fusion, comprising the following steps: Collect three-view images of the user's face to establish a three-view sample set. The three-view images of the face include front, left and right side images. At the same time, acquire clinical factors, including age, gender, weight, height, and mouth breathing level. The three-view sample set is input into the pre-trained target detection model to obtain the category information and original confidence score of each view as the detection result. The category information includes abnormality-related categories. Based on clinical factors, the risk factor is calculated and output through the rule engine. Based on the detection results output by the target detection model, if the category belongs to the abnormal related category, the original confidence is enhanced according to the risk factor and the preset category enhancement coefficient to obtain the enhanced confidence. From the enhanced confidence scores, the maximum value is selected as the final diagnostic confidence score; The final diagnostic confidence level is compared with the preset grading threshold, and the corresponding risk grading result is output to complete the adenoid facies screening.

[0010] In one embodiment, after establishing the three-view sample set and before inputting it into the pre-trained target detection model, a validity verification step for the three-view face is also included. Specifically, the three-view face is verified by one or more of the following: face confidence, detection density, and overall validity score. If the validity verification fails, a retake suggestion is issued to the user.

[0011] In one embodiment, the labeling system used by the pre-trained object detection model includes the following nine categories: Overall frontal view of the face, normal frontal view of the mouth, abnormal frontal view of the mouth, overall right side of the face, normal right chin, abnormal right chin, overall left side of the face, normal left chin, and abnormal left chin.

[0012] In one embodiment, the calculation and output of risk factors based on clinical factors through a rule engine includes: Each clinical factor is mapped to a multiplicative coefficient within a preset safety range, and the multiplicative coefficient is used to quantify the relative impact of a clinical factor on the risk of abnormality. Multiply the multiplicative coefficients of all clinical factors to obtain an initial risk value; The initial risk value is clipped to a preset global boundary. The final risk factors are obtained, among which This is the upper bound of the risk factor. This represents the lower bound of the risk factor.

[0013] By designing multiplicative coefficients, the synergistic amplification or synergistic attenuation of effects between factors can be naturally reflected, clearly quantifying the relative impact of a clinical factor on abnormal risk and improving the accuracy of detection.

[0014] Furthermore, the global boundary interval is [0.8, 1.8].

[0015] By cropping the initial risk value to a predefined global boundary, extreme values ​​can be avoided. This prevents the original confidence level from being over-amplified, reflecting the principle of boundedness. Furthermore, it prevents the high-confidence anomalies already detected by the model from being excessively weakened, protecting the model's valid evidence and avoiding false negatives. Through the synergistic effect of the global boundary safety interval design, the influence from clinical factors is limited to a predefined, reasonable range, thereby ensuring the robustness and reliability of the entire "bounded monotonic enhancement" system.

[0016] In one embodiment, the rule engine also outputs a contribution detail for each clinical factor, which includes the clinical factor name, input value, and corresponding multiplicative coefficient; wherein the input value is clinical data provided by the user.

[0017] In one embodiment, the post-enhancement confidence is calculated using the following formula, where the enhancement operation satisfies monotonicity and boundedness: , In the formula, To enhance post-confidence; This represents the original confidence level. This is a truncation function used to restrict the result to the interval [0,1]. This is the preset category enhancement coefficient; As a risk factor.

[0018] In one embodiment, the category enhancement coefficient is configured to a value greater than zero only for anomalously related categories, and to zero for non-anomalously related categories.

[0019] By enhancing only the confidence level of abnormal categories, the system can amplify these weak or uncertain abnormal signals, directly targeting the core objectives of screening to detect abnormalities, effectively reducing missed diagnoses caused by model uncertainty. Furthermore, setting the normal category to zero means that regardless of clinical risk, the judgment of normality is entirely based on visual evidence itself, remaining unchanged, protecting the originality and stability of the model's judgment of normal samples, preventing excessive interference from clinical information that would cause the system to "think everything is normal," thereby avoiding the risk of false positives introduced by false enhancement.

[0020] In one instance, the abnormality-related categories include: frontal mouth abnormalities, left chin abnormalities, and right chin abnormalities.

[0021] In one embodiment, the preset grading thresholds include a high threshold and a middle threshold, and the grading rule is as follows: If the final diagnosis confidence is greater than or equal to the high threshold, then output "high confidence probability". If the final diagnostic confidence level is between the high threshold and the middle threshold, the output is "medium confidence questionable". If the final diagnosis confidence is less than the intermediate threshold, then output "low confidence, impossible".

[0022] Furthermore, the high threshold is greater than or equal to 0.8, and the middle threshold is less than 0.6.

[0023] By aggregating the maximum values ​​within the aforementioned parameter range and dividing the data into three levels, a risk filter is constructed. A high threshold of 0.8 or higher is used to lock in highly suspected cases, ensuring the accuracy of alerts. An intermediate threshold between 0.6 and 0.8 is used to capture all suspicious cases, ensuring no cases are missed, and these cases are then submitted to doctors for final decision. Finally, a low threshold of less than 0.6 is used to efficiently release people who are likely to be healthy. In this way, the risks of missed diagnoses and misdiagnoses are balanced to the maximum extent, forming a diagnostic closed loop that is both sensitive and reliable and conforms to clinical practice.

[0024] The present invention also provides an adenoid facial screening device based on three-view detection and medical rule fusion, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the adenoid facial screening method based on three-view detection and medical rule fusion when the computer program is executed.

[0025] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention discloses an adenoid facial screening method and device based on the fusion of three-view detection and medical rules. It collects three-view images of the user's face, performs detection using a pre-trained target detection model, and outputs detection results containing nine specific labels and raw confidence scores. Simultaneously, it acquires clinical factors, converts each clinical factor into multiplicative coefficients using a rule engine, and multiplies them to obtain risk factors. Subsequently, it performs bounded monotonic enhancement only on three types of abnormality detection results: frontal mouth abnormalities, left and right chin abnormalities. Finally, it takes the maximum confidence score of the abnormality category after enhancement in the three-view images for threshold classification, outputting high, medium, and low risk conclusions, and simultaneously outputting the risk factors and their contribution details. This invention ensures output stability through "boundedness," improves logical consistency through "monotonicity," and enhances the transparency and credibility of the adenoid facial screening method through "interpretability." These three characteristics work together to effectively improve the robustness and clinical interpretability of the adenoid facial screening method, thereby better meeting the actual needs of medical scenarios in terms of the accuracy and interpretability of screening results and significantly reducing the risk of missed diagnoses. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0027] Figure 1 This is a flowchart illustrating the adenoid facial screening method based on three-view detection and medical rule fusion provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the accompanying drawings and... The embodiments further illustrate the present invention in detail. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0029] To achieve early screening for adenoid facies in children, this embodiment provides a method for screening adenoid facies based on three-view detection and medical rule fusion, such as... Figure 1 As shown, it includes the following steps: S1. Collect the user's three-view facial images and establish a three-view sample set. The three-view facial images include front, left and right side images. Simultaneously acquire clinical factors, including age, gender, weight, height, and mouth breathing level.

[0030] The system collects three-view images and optional clinical factors (age, gender, weight, height, and open-mouth breathing level, where weight and height are used to calculate BMI) via an H5 scale or guided page. The terminal device only performs a three-view integrity and basic validity check and provides retake suggestions; it does not store the original images for a long time, adhering to the principles of data minimization and privacy protection from the system design level.

[0031] Shooting guidelines for facial three-view images: For a frontal view, the face should be centered, the eyes should be level and unobstructed; the left and right side angles should be approximately 80-90 degrees, avoiding extreme tilt angles; it is recommended to shoot under natural light or uniform indoor light; the image format should be JPG or PNG, and the size of a single image should not exceed 5 MB.

[0032] The validity of the collected facial three-view images is verified: the lowest confidence level of the detected face (not less than 0.5), the average detection density (not less than 1 target per image), and the overall validity score (not less than 0.3) are used as reference standards.

[0033] Quality control feedback: Immediately return samples that do not meet the requirements for retaking.

[0034] S2. Input the three-view sample set into the pre-trained target detection model to obtain the category information and original confidence score of each view as the detection result. The category information includes abnormal related categories. Based on clinical factors, the risk factor is calculated and output through the rule engine.

[0035] In this embodiment, the pre-trained target detection model is the YOLO model, which uses a 9-class labeling system and employs clinically and everyday language that is easy to understand: front_face: The overall frontal view of the face (including visible facial features and facial contours); front_mouth_normal: The mouth appears normal from the front (naturally closed, with no obvious teeth showing or cleft lips); front_mouth_abnormal: Abnormalities in the frontal mouth (such as open mouth, showing teeth, obvious gap between the upper and lower lips, etc.); right_face: The entire right side of the face; right_chin_normal: The right chin is normal (the proportion and contour of the chin are in line with the normal age). right_chin_abnormal: Abnormality of the right chin (such as receding jaw, short chin, or protruding mouth); left_face: The entire left side of the face; left_chin_abnormal: Abnormality in the left chin (same as the right side, determined from the left-side perspective); left_chin_normal: The left side of the chin is normal.

[0036] Each image is accompanied by a corresponding text annotation file, recording the position and size of the annotation box line by line (using relative coordinates, normalized to the 0-1 range). The selection guidelines emphasize covering key functional areas, avoiding excessive background, and prioritizing the annotation of identifiable areas in occluded or blurred images.

[0037] A dual-person cross-checking approach (focusing on IoU and consistency rate) is employed, combined with rule-based script self-checking (verifying category names, coordinate ranges, file alignment, etc.). Problematic samples are re-labeled and the reasons are recorded, forming a closed-loop management system to improve labeling consistency.

[0038] Next, a dataset for the object detection model is established, organized according to the structure yolo_dataset / {train,val,test} / {images,labels}. The training set (train) is used for object detection model learning, the validation set (val) is used for validation and parameter tuning during training, and the test set (test) is used only for final independent evaluation. The labels are the class labels. When partitioning the dataset, "patients" are used as the smallest unit to ensure that different view images of the same subject are in the same set, preventing data leakage.

[0039] Before training the object detection model, the dataset is lightly cleaned: damaged images are removed, and continuous shots or highly similar samples are removed to ensure statistical independence between sets.

[0040] Model Training: The model was trained using the YOLO general object detection framework for approximately 320 epochs with a batch size of 4 and an input resolution of 576. The AdamW optimizer was used in conjunction with a cosine learning rate scheduler; mosaic enhancement was disabled at the end of training to ensure evaluation objectivity. To ensure reproducibility, the random seed and number of worker threads were fixed. Data augmentation primarily employed multi-scale transformations and random affine transformations, balancing appearance diversity with medical recognizability. To address class imbalance, weighted sampling and loss weight adjustment were used, with resampling of rare classes as needed to reduce model bias. Training outputs included optimal weights, final weights, training logs, and evaluation curves.

[0041] Model Inference: Objectives and Scope: a. Objective: To enhance the detection confidence of abnormality-related categories in a bounded, monotonic, and interpretable manner; without adding new detection boxes, excessively amplifying confidence, or affecting the normal category. b. Abnormality-related categories: Including front mouth abnormalities, left chin abnormalities, and right chin abnormalities.

[0042] Input and output definitions: a. Input A (Visual Evidence): Detection results of the three views (category, location, original confidence level). b. Input B (Clinical Factors): Age, gender, BMI calculated from weight and height, and open-mouth breathing level (all optional). c. Output: i. Enhanced detection results for each view (only abnormally related categories may be enhanced); ii. Risk factor values ​​calculated by the rule engine, details of the contribution of each clinical factor, and rule version number; iii. Graded diagnostic results based on the maximum enhanced confidence level of the abnormal category in the three views.

[0043] The inference parameters (input resolution, detection threshold, NMS settings, etc.) remain consistent with the configuration at the end of training.

[0044] Rule parameters (risk factor boundaries) , This is the upper bound of the risk factor. Lower bound of risk factor, category enhancement coefficient This only applies to exception-related categories.

[0045] The classification is based on the "maximum confidence level of anomaly categories in three-view diagrams": a high threshold corresponds to "high confidence probability", a medium threshold corresponds to "medium confidence doubt", and a low threshold corresponds to "low confidence impossibility".

[0046] Validity thresholds (minimum face confidence, detection density, etc.) are used for input quality control and interception.

[0047] Factor mapping and risk factors ( )calculate: a. Each clinical factor is mapped to a multiplicative coefficient ( This reflects its relative impact on abnormal risks. Values ​​are taken within a preset safe range; a neutral value of 1.0 is taken when a value is missing or invalid.

[0048] b. Risk Factors It is the product of the coefficients of each factor, and is clipped to the global boundary. By cropping the initial risk value to a predefined global boundary, extreme values ​​can be avoided. This prevents the original confidence level from being over-amplified, reflecting the principle of boundedness. Furthermore, it prevents the high-confidence anomalies already detected by the model from being excessively weakened, protecting the model's valid evidence and avoiding false negatives. Through the synergistic effect of the global boundary safety interval design, the influence from clinical factors is limited to a predefined, reasonable range, thereby ensuring the robustness and reliability of the entire "bounded monotonic enhancement" system.

[0049] c. The rules engine synchronously outputs contribution details (including factor name, input value, and corresponding coefficient) and version number for traceability and calibration.

[0050] d. Example Mapping (calibrable): i. Age: Unit: years, common range 3-12 years. Default is near neutral, e.g.: 3-5 years → 1.02, 6-9 years → 1.00, 10-12 years → 1.03; boundary [0.95, 1.05]. ii. Sex: Value is male / female. Direct impact is weak, default is neutral, e.g.: male → 1.00, female → 1.00; boundary [0.98, 1.02]. iii. BMI: Calculated from height and weight. It is recommended to stratify by age or use Z-score mapping to a mild coefficient, e.g., normal → 1.00, slight deviation → 0.98-1.02, significant deviation → 0.95-1.05; boundary [0.95, 1.05]. iv. Mouth Breathing Grade: Classified as 0 / 1 / 2 / 3 (none / mild / moderate / severe). The weights are relatively high but still controlled, such as: 0→1.00, 1→1.15, 2→1.35, 3→1.55; boundary [1.00, 1.60].

[0051] e. Preprocessing and missing value handling: All factors are first validated and standardized, and out-of-bounds values ​​are pruned to a safe range; when there is no value or the value cannot be parsed, a neutral value of 1.0 is taken; contribution details record the original value, standardized value and final coefficient.

[0052] S3. Based on the detection results output by the target detection model, if the category belongs to the abnormal correlation category, the original confidence is enhanced according to the risk factor and the preset category enhancement coefficient to obtain the enhanced confidence. The enhancement operation satisfies monotonicity and boundedness.

[0053] The specific process is as follows: a. Read the three-view detection results and (optional) clinical factors; b. Perform validity verification and standardization on the clinical factors, setting missing values ​​to 1.0; c. Calculate the coefficients of each factor according to the rule table and multiply them to obtain R, then trim to... d. Traverse each detection box: If it is an anomaly-related category, sort by category coefficient. e. Perform bounded linear augmentation on its confidence level; f. Retain the augmented detection set within each viewpoint; g. Take the maximum augmentation confidence level of anomaly-related categories across the three views. g. Based on the classification threshold Perform diagnostic grading and output structured results and rule traceability information.

[0054] Bounded monotonic augmentation only for anomaly-related categories: a. For each detection box (category) Original confidence level If category If it belongs to the abnormal correlation category, then it is calculated according to the category enhancement coefficient. Linear amplification is applied; no enhancement is applied to normal or overall categories. b. The enhancement magnitude is subject to two constraints: i. Monotonicity: When it increases, ii. Not reduced; boundedness: Not exceeding the upper limit of 1.0, the reinforcement strength is... Control. c. Enhancement operations do not generate new detection boxes or change their positions; they only adjust the target confidence level. Boundary and anomaly handling: a. Missing factor: The corresponding coefficient is set to 1.0, which does not affect... a. Neutral baseline; b. Out-of-bounds values: pruned according to the upper and lower bounds of each factor before being included in the calculation; c. Low-quality samples: if the validity check is not passed, priority is given to suggesting re-shooting; d. Inconsistent views: the maximum value of the three views is used as the standard to avoid weak views lowering strong evidence; e. Normal category: no enhancement is performed to prevent the introduction of false negatives.

[0055] Design principles: a. Monotonicity: When clinical risk factors When augmented, the overall evidence related to the anomaly category should not be weakened, and the confidence level after augmentation should not be lower than before augmentation. Implementation method: Using... linear response and matching ≥0 ensures the slope of the response curve is non-negative. Verification is performed using offline heatmaps and online sampling audits.

[0056] b. Boundedness: The confidence level is maintained within the [0,1] interval after enhancement to avoid excessive amplification of small priors. Implementation: The multiplicative response calculation results are pruned with an upper bound and numerically stabilized. It is within a mild range (e.g., 0.4-0.8). Verification was performed through full-coverage scanning and online distributed monitoring.

[0057] c. Explainability: System output risk factors The rules include the source of the data, details of the contribution of each clinical factor, key thresholds, and rule version numbers, facilitating physician understanding and project review. Implementation: Calculations are performed in the rule engine module, generating readable output. Validation is achieved through traceability sampling, clinical review, and A / B evaluation.

[0058] Monotonicity ensures correct direction, boundedness ensures controllable magnitude, and interpretability ensures auditability; these three elements together form a closed-loop engineering approach of "steady-state enhancement." Compared to unconstrained enhancement or black-box retraining, this approach significantly reduces clinical and compliance risks, facilitating rapid rollback and incremental calibration. Enhancing only abnormality-related categories significantly reduces the risk of missed diagnoses without altering the model structure; aggregation of maximum values ​​from three views suppresses the negative impact of weakly priced views, consistent with the clinical logic of "strong abnormalities prompting a warning"; inference-phase fusion maintains engineering flexibility, facilitating rapid calibration and rollback, and reducing maintenance costs.

[0059] In this embodiment, the enhanced confidence level is calculated using the following formula, which is used to linearly adjust the detection results of the anomaly category using the risk factor R: First, calculate the amplification coefficient ( )(when ,constant; ,according to Enlarged proportionally; ,according to (Reduced proportionally), then multiply the coefficient by the original confidence level. Finally, the results are truncated to the interval [0,1] (1 for values ​​greater than 1 and 0 for values ​​less than 0) to obtain the enhanced confidence score. The specific calculation is performed using the following formula: , In the formula, To enhance post-confidence; The original confidence level (YOLO model output, interval [0,1]); This is a truncation function used to restrict the result to the interval [0,1], clamping the value with upper and lower bounds to ensure boundedness; This is the preset category enhancement coefficient; The risk factor is obtained by multiplicative combination and trimming of clinical factors, within the range of... ;default , .

[0060] The category enhancement coefficient mentioned above is only positive for abnormally related categories; for non-abnormally related categories, the category enhancement coefficient is zero.

[0061] Monotonicity: when When it increases, No reduction ( ≥0).

[0062] Boundedness: Always ∈ [0,1], and The gain naturally converges when it approaches 1.0.

[0063] S4. Select the maximum value from the enhanced confidence scores of the abnormality-related categories in the three-view sample set as the final diagnostic confidence score.

[0064] Within the three-view range, the maximum value of the "enhanced confidence score of the anomaly-related category" is taken as the final diagnostic criterion. Example of grading thresholds: a maximum value ≥ 0.80 is "high confidence possible", ∈ [0.60, 0.80) is "medium confidence questionable", and < 0.60 is "low confidence impossible".

[0065] This design ensures that a single perspective of strong evidence can drive the overall conclusion, while a weak perspective will not unduly diminish the strength of the evidence.

[0066] For example, a. Suppose an image on the right side detects an "abnormal right chin," with an initial confidence level of 0.62; b. Calculate... =1.40, this category =0.60; c. Enhanced confidence ≈0.62×(1+0.60×(1.40-1))≈0.62×1.24≈0.77 (<1.0); d. If the maximum value of the three views, 0.77, falls within the middle threshold range (e.g., 0.60-0.80), then the classification is "medium confidence doubtful".

[0067] Preset graded threshold calibration: Objective: To select a diagnostic grading threshold while satisfying the constraints of monotonicity and boundedness. With category enhancement coefficient This optimizes the overall performance indicators (such as the F1 score).

[0068] Constraints: a. Monotonicity: When the threshold increases, the confidence level does not decrease after enhancement; when the threshold is raised, the positive result does not increase. b. Bounded: the confidence level after enhancement ∈ [0,1]; ; ∈[0.4,0.8]; .

[0069] Calibration process: a. Fix the basic inference parameters; b. Generate on the validation set. Quadruple, where a. Represent the three views of the face and calculate the enhanced confidence and maximum value; b. Search on the candidate parameter grid, check the monotonicity and boundedness of each parameter group and calculate the performance index; c. Select the parameter group that satisfies the constraints and has the best comprehensive index; e. Verify the generalization ability on the independent test set; f. Generate a calibration report.

[0070] In this embodiment, the initial value is: =0.80, =0.60, =0.6; b. If the false positive rate is high, the level can be increased. Or narrow If the rate of missed diagnoses is high, the level can be lowered. Or relax c. Any adjustments must be re-verified for monotonicity and bounded constraints, and A / B testing and version rollback management must be performed.

[0071] For samples with abnormally low confidence or excessively low detection density, manual review is conducted according to validity rules to continuously improve data quality.

[0072] On the other hand, the embodiment also provides an adenoid facial screening device based on three-view detection and medical rule fusion, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the adenoid facial screening method based on three-view detection and medical rule fusion when the computer program is executed.

[0073] To better illustrate the effectiveness of the method provided by this invention, experimental verification was conducted, as shown in Table 1.

[0074] Table 1

[0075] The results in Table 1 show that the model accurately defines the normal and abnormal states of the face as a whole and in parts (mouth and chin) from three perspectives: front, left, and right. It also achieves high-precision recognition of all nine types of targets, with an average accuracy (mAP@0.5) of 97.8%, demonstrating excellent classification performance and robustness.

[0076] The adenoid facial screening method based on the fusion of three-view detection and medical rules has achieved the following technical breakthroughs: (1) Multi-view stable aggregation: After enhancing the abnormal related categories during the inference period, the aggregation method of “taking the maximum value of the three-view enhancement confidence” is adopted, which not only ensures monotonicity but also has robustness, effectively combating the uncertainty caused by the difference in angle and illumination.

[0077] (2) Consistency integration of medical rules and visual evidence: Clinical factors are mapped into bounded, monotonic, and interpretable risk factors that only apply to abnormally related categories. The enhancement magnitude is controlled by upper and lower bounds and category coefficients, taking into account both medical rationality and engineering stability.

[0078] (3) Class imbalance and distribution drift: Weighted loss and data augmentation are used to alleviate imbalance at the training end, while the parameters at the inference end are kept consistent with those at the end of training and an equivalent scheme of risk weighting is introduced, which jointly improves the generalization ability of the model.

[0079] (4) Sample validity and quality control: The lowest confidence level of the face, detection density and validity score are used as input gates, and combined with manual review closed loop, the noise introduced in the collection and labeling process is continuously reduced.

[0080] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for screening adenoid facial features based on three-view detection and medical rule fusion, characterized in that, Includes the following steps: Collect three-view images of the user's face, including front, left and right side images, establish a three-view sample set, and simultaneously acquire clinical factors, including age, gender, weight, height, and mouth breathing level. The three-view sample set is input into the pre-trained target detection model, and the category information and original confidence score of each view are obtained as the detection result. The category information includes anomaly-related categories. Based on clinical factors, a rule engine calculates and outputs risk factors, including: mapping each clinical factor to a multiplicative coefficient within a preset safety range, whereby the multiplicative coefficient quantifies the relative impact of a clinical factor on abnormal risk; multiplying the multiplicative coefficients of all clinical factors to obtain an initial risk value; and cropping the initial risk value to a preset global boundary. Within this process, the final risk factors are obtained, among which... This is the upper bound of the risk factor. This represents the lower bound of the risk factor. Based on the detection results output by the target detection model, if the category belongs to the anomaly-related category, the original confidence level is enhanced according to the risk factor and the preset category enhancement coefficient to obtain the enhanced confidence level. The enhanced confidence level is calculated using the following formula, and the enhancement operation satisfies monotonicity and boundedness: In the formula, To enhance post-confidence; This represents the original confidence level. This is a truncation function used to restrict the result to the interval [0,1]. This is the preset category enhancement coefficient; As a risk factor; From the enhanced confidence scores, the maximum value is selected as the final diagnostic confidence score; The final diagnostic confidence level is compared with the preset grading threshold, and the corresponding risk grading result is output to complete the adenoid facies screening.

2. The adenoid facies screening method based on three-view detection and medical rule fusion according to claim 1, characterized in that, After establishing the three-view sample set and before inputting it into the pre-trained target detection model, there is also a validity verification step for the three-view face. Specifically, the three-view face is verified by one or more of the following: face confidence, detection density and overall validity score. If the validity verification fails, a retake suggestion is sent to the user.

3. The adenoid facies screening method based on three-view detection and medical rule fusion according to claim 1, characterized in that, The labeling system used in the pre-trained object detection model includes the following 9 categories: Overall frontal view of the face, normal frontal view of the mouth, abnormal frontal view of the mouth, overall right side of the face, normal right chin, abnormal right chin, overall left side of the face, normal left chin, and abnormal left chin.

4. The adenoid facies screening method based on three-view detection and medical rule fusion according to claim 1, characterized in that, The rules engine also outputs a contribution detail for each clinical factor, which includes the clinical factor name, input value, and corresponding multiplicative coefficient; where the input value is clinical data provided by the user.

5. The adenoid facies screening method based on three-view detection and medical rule fusion according to claim 1, characterized in that, The category enhancement coefficient is configured to a value greater than zero only for abnormally related categories; for non-abnormally related categories, the category enhancement coefficient is configured to zero.

6. The adenoid facies screening method based on three-view detection and medical rule fusion according to claim 5, characterized in that, The abnormality categories mentioned include: abnormalities of the frontal mouth, abnormalities of the left chin, and abnormalities of the right chin.

7. The adenoid facies screening method based on three-view detection and medical rule fusion according to claim 1, characterized in that, The preset grading thresholds include a high threshold and a middle threshold, and the grading rules are as follows: If the final diagnosis confidence is greater than or equal to the high threshold, then output "high confidence probability". If the final diagnostic confidence level is between the high threshold and the middle threshold, the output is "medium confidence questionable". If the final diagnosis confidence is less than the intermediate threshold, then output "low confidence, impossible".

8. An adenoid facial screening device based on three-view detection and medical rule fusion, comprising a memory and a processor, wherein the memory is used to store a computer program, characterized in that, The processor is used to implement the adenoid facial screening method based on the fusion of three-view detection and medical rules as described in any one of claims 1 to 7 when executing the computer program.

Images