Intelligent evaluation system for children's pain based on AI facial expression analysis

Abstract

The present application relates to the field of medical auxiliary diagnosis, and specifically relates to a child pain intelligent evaluation system based on AI facial expression analysis, a face detection module acquires a child face image, and uses a multi-task cascaded convolutional neural network model for three-level cascaded detection to obtain face key point pixel coordinate information, a key point positioning module carries out noise filtering processing on the coordinate information to obtain more accurate face key point coordinate information, the information is transmitted to a key point heat map representation module, the module combines face key point coordinate information and a heat map mechanism to generate a face key point heat map representing the degree of child pain, a child pain intelligent evaluation module evaluates the degree of child pain based on the heat map and obtains an evaluation result, finally, a pain evaluation result visualization module visualizes the evaluation result, high-precision face detection and key point heat map technology improve the accuracy of pain representation, and provide an effective method for child pain evaluation.

Description

Technical Field

[0001] This invention relates to the field of medical auxiliary diagnosis, specifically to an intelligent pain assessment system for children based on AI facial expression analysis, which is particularly suitable for objectively assessing the pain level of child patients who are unable to accurately express their pain sensations. Background Technology

[0002] Pain, as a subjective experience, is of significant value in clinical diagnosis and treatment. Adult patients can usually accurately describe the intensity, location, and nature of their pain verbally. However, pain assessment in young children, especially infants and children with special needs who cannot express themselves accurately, has always been a challenge in the medical field.

[0003] Currently, clinical assessment of pain in children mainly relies on the following methods: first, observing the child's behavioral manifestations, such as crying and restlessness; second, using scales for assessment, such as the Facial Expression Pain Scale (FLACC) and the Wong-Baker Facial Expression Pain Scale; and third, assessing physiological indicators, such as changes in heart rate, blood pressure, and respiration. These methods all have certain limitations, such as strong subjectivity, low reliability of assessment results, and poor real-time performance.

[0004] With the development of artificial intelligence technology, especially the advancements in computer vision and deep learning, it has become possible to use AI technology to analyze children's facial expressions to assess pain levels. However, existing AI-based pain assessment systems have the following problems: first, inaccurate facial expression capture leads to inaccurate assessment results; second, the assessment models lack targeted analysis of children's unique facial expression characteristics; third, they cannot effectively handle interference from individual differences; and fourth, they lack intuitive display of assessment results.

[0005] Therefore, there is an urgent need for an intelligent system that can objectively, accurately, and in real time assess the degree of pain in children to assist medical staff in carrying out reasonable pain management and treatment. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the prior art by providing an intelligent pain assessment system for children based on AI facial expression analysis. This system achieves high-precision facial detection through a multi-task cascaded convolutional neural network, uses key point heatmap technology to represent the degree of pain, and achieves objective assessment through a dual-model comparison assessment mechanism. Finally, the assessment results are displayed in a visual manner, providing medical staff with objective, accurate, and real-time evidence for assessing children's pain.

[0007] This invention proposes an intelligent pain assessment system for children based on AI facial expression analysis, comprising:

[0008] The face detection module is used for:

[0009] Obtain images of children's faces;

[0010] The child's facial image is input into a multi-task cascaded convolutional neural network model for three-level cascaded detection to obtain the pixel coordinate information of key points on the child's face.

[0011] The key point localization module, which is communicatively connected to the face detection module, is used for:

[0012] Receive the pixel coordinate information of key points on the child's face sent by the face detection module;

[0013] The pixel coordinate information of the child's facial key points is subjected to noise filtering to obtain the facial key point coordinate information.

[0014] The key point heatmap representation module, which is communicatively connected to the key point localization module, is used for:

[0015] Receive facial key point coordinate information sent by the key point positioning module;

[0016] Based on the combination of the facial key point coordinate information and the heat map mechanism, a facial key point heat map representing the degree of pain in children is generated.

[0017] The intelligent pain assessment module for children is communicatively connected to the key point heatmap representation module and is used for:

[0018] Receive the facial key point heatmap sent by the key point heatmap representation module;

[0019] Based on the facial key point heat map, the child's pain was assessed, and the assessment results were obtained.

[0020] The pain assessment result visualization module, which is communicatively connected to the intelligent pain assessment module for children, is used for:

[0021] Receive the assessment results sent by the child pain intelligent assessment module;

[0022] The evaluation results are visualized.

[0023] Preferably, the intelligent pain assessment module for children includes:

[0024] The facial key point heatmap localization submodule is used to locate the detected facial region based on the facial key point heatmap sent by the key point heatmap representation module, and obtain the facial key point heatmap.

[0025] The key point heatmap score extraction submodule is communicatively connected to the facial key point heatmap positioning submodule and is used to obtain the scores of all key points heatmaps based on the facial key point heatmap sent by the facial key point heatmap representation module.

[0026] The child pain model training submodule is communicatively connected to the key point heatmap score extraction submodule and is used to train the child pain model based on all key point heatmap scores output by the key point localization module.

[0027] The child pain assessment submodule is communicatively connected to the key point heatmap score extraction submodule and the child pain model training submodule. It is used to input the output of the key point heatmap score extraction submodule into the child pain model for classification to obtain the assessment result.

[0028] As a preferred embodiment, the facial key point heatmap is represented as follows:

[0029] ,

[0030] in, Let i be the heatmap matrix of the i-th key point in the facial region. Indicates will Pixel values ​​normalized to (0,1); Point and points The Euclidean distance between them; The heatmap represents the first... Each region.

[0031] Preferably, the heatmap scores of all the key points are represented as follows:

[0032] ,

[0033] in, This is the heatmap fraction matrix for key points.

[0034] Preferably, the pediatric pain assessment model is constructed using the Dlib library; the pediatric pain assessment model includes: a pain assessment model and a no-pain assessment model; the pain assessment model is used to assess whether a child is in pain; the no-pain assessment model is used to assess whether a child is not in pain; the pain assessment model and the no-pain assessment model are obtained using a preset number of training set data from the pediatric pain model training submodule, the preset number being 50-100; the training set data includes the child's facial key point heatmap scores.

[0035] Preferably, the output of the key point heatmap score extraction submodule includes: a child pain score matrix and a child no-pain score matrix; the child pain score matrix and the child no-pain score matrix are subtracted using heatmaps to obtain the child pain assessment score, expressed as:

[0036] ,

[0037] in, For children's pain assessment scores, For children's pain score matrix, A pain-free score matrix for children.

[0038] As a preferred method, the assessment results of facial pain evaluation in children are expressed as follows:

[0039] ,

[0040] in, For the results of pain assessment in children, The preset pain score threshold; when When the child is in pain, the intelligent pain assessment module outputs an assessment result indicating that the child is in pain; when When the time is specified, it indicates that the child is not in pain, and the assessment result output by the child pain intelligent assessment module is "no pain".

[0041] Preferably, the face detection module includes a face detection submodule and an eye detection submodule; the face detection module includes a face detection model, an eye detection model, and an eye detection model; the face detection model is used to perform face detection on the child's face image acquired by the face detection module and output a face image; the eye detection model is used to perform eye detection on the child's face image acquired by the face detection module and output an eye image; the eye detection model is used to perform eye detection on the child's face image acquired by the face detection module and output an eye image; the face detection model adopts the MTCNN model, which includes a first-level detector, a second-level detector, and a third-level detector; the first-level detector is used to perform first-level face detection on the child's face image acquired by the face detection module; the second-level detector is used to perform second-level face detection on the child's face image acquired by the face detection module; the third-level detector is used to perform third-level face detection on the child's face image acquired by the face detection module; the MTCNN model uses a downsampling mechanism to convert the intermediate results obtained from the first-level face detection of the child's face image into the input image size of the second-level detector and the third-level detector.

[0042] Preferably, in the MTCNN model, the trained eye detection model is used as the child's facial key point data processing model to obtain the child's facial key point pixel coordinate information, represented as:

[0043]

[0044] in, The weights are constants, typically ranging from 0.5 to 1.5. The two-dimensional pixel coordinates of the eye center output by the eye detection model; The output of the eye detection model is a two-dimensional pixel coordinate matrix of three key points, where each column represents the inner corner of the left eye. The outer corner of the left eye and the outer corner of the right eye The coordinates; This is the standardized coordinate matrix of key eye points obtained after transformation, where each column represents the standardized inner canthus of the left eye. The outer corner of the left eye and the outer corner of the right eye The coordinates.

[0045] In actual calculations, this formula represents the relative positional relationship between the coordinates of the eye's center and the coordinates of each corner of the eye. When calculating each key point separately, the formula expands to:

[0046] ,

[0047] This mapping method uses the stability of the eye region as a benchmark reference system and calculates more accurate facial key point locations through relative positional relationships, effectively solving the challenge of key point localization caused by facial expression changes in children under pain conditions.

[0048] Preferably, the key point localization module includes: a heatmap filtering submodule and a key point coordinate extraction submodule; the heatmap filtering submodule is used to perform mean filtering on the key point heatmap response values ​​output by the face detection module, using the following formula:

[0049] ,

[0050] In the formula, For heat map at location The response value at that location, The x-coordinate of the pixel in the heatmap. The vertical coordinate of the pixel in the heatmap. The filtered heatmap response value. This represents the number of pixels contained in the filter window. For the point The filtering window area centered on the center is typically... A square area of ​​size The recommended value is 3 or 5.

[0051] The key point coordinate extraction submodule is used to perform local maximum detection on the filtered heatmap to obtain stable key point coordinates of the child's face, represented as follows:

[0052] ,

[0053] In the formula, The coordinates of the detected key points, For The local neighborhood centered on the center is usually taken for or The window size; this formula represents the point A point whose heatmap response value is higher than that of all other points in its neighborhood is identified as a key point. To further improve the stability of key point localization, the system also performs weighted smoothing on the key point coordinates of consecutive frames along the time axis:

[0054] ,

[0055] In the formula, Indicates time The smoothed keypoint coordinates, For a moment The coordinates of the key points were directly detected. For a moment The smoothed keypoint coordinates, The smoothing coefficient ranges from (0,1], typically between 0.3 and 0.7. A larger value is used for high confidence levels, and a smaller value for low confidence levels. This two-stage processing method (first performing spatial domain filtering on the heatmap, then coordinate smoothing in the time domain) avoids the meaningless translation problems that may result from directly applying mean filtering to coordinate values. It also fully utilizes the spatial distribution information and temporal continuity of the heatmap, significantly improving the stability and robustness of key point localization. It is particularly suitable for handling facial tremors and expression changes that may occur in children under pain conditions.

[0056] The beneficial effects of this invention include:

[0057] 1. Through a three-level cascaded detection mechanism of multi-task cascaded convolutional neural network (MTCNN), high-precision detection of children's faces was achieved. Even under different lighting, angles and expression conditions, the facial regions and key points can be accurately located, providing a reliable foundation for subsequent analysis.

[0058] 2. Using key point heatmap technology to represent changes in children's facial expressions, discrete key point information is transformed into a continuous heat map distribution, which captures facial expression changes more comprehensively and improves the accuracy of pain representation.

[0059] 3. An innovative dual-model comparison mechanism of "pain assessment model" and "no pain assessment model" was designed. The pain level was judged by the difference between the output results of the two models, which effectively eliminated the interference caused by individual differences and improved the accuracy of the assessment.

[0060] 4. The system displays the assessment results in a visual manner, including a pain intensity heatmap and pain / no pain judgment results, which makes it easier for medical staff to understand the assessment results intuitively and improves the efficiency of clinical decision-making.

[0061] 5. The system adopts a modular design with clearly defined interfaces for each functional module, which facilitates functional expansion or optimization based on clinical needs and adapts to the requirements of different medical environments. Attached Figure Description

[0062] Figure 1 This is an overall architecture diagram of the intelligent pain assessment system for children based on AI facial expression analysis, as described in this invention.

[0063] Figure 2 This is a structural diagram of the face detection module of the present invention;

[0064] Figure 3 This is a flowchart of the three-level detector of the MTCNN model of this invention;

[0065] Figure 4 This is a structural diagram of the key point positioning module of the present invention;

[0066] Figure 5 This is a schematic diagram illustrating the working principle of the heat map representation module, a key feature of this invention.

[0067] Figure 6 This is a structural diagram of the intelligent pain assessment module for children in this invention;

[0068] Figure 7 This is a schematic diagram illustrating the working principle of the dual-model comparison and evaluation mechanism of this invention.

[0069] Figure 8 This is a display effect diagram of the pain assessment result visualization module of the present invention. Detailed Implementation

[0070] Please refer to the attached document. Figure 1-8 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0071] like Figure 1 As shown, the AI-based facial expression analysis-based intelligent pain assessment system for children provided by this invention includes: a facial detection module 1, a key point localization module 2, a key point heatmap representation module 3, an intelligent pain assessment module for children 4, and a pain assessment result visualization module 5. These modules are interconnected through a data communication interface to form a complete pain assessment process.

[0072] The face detection module 1 acquires images of the child's face and inputs them into a multi-task cascaded convolutional neural network (MTCNN) model for three-level cascaded detection to obtain the pixel coordinates of key facial points. The key point localization module 2 communicates with the face detection module 1 and receives the pixel coordinates of the key facial points from the face detection module 1. It performs noise filtering on this information to obtain more accurate key facial point coordinates. The key point heatmap representation module 3 communicates with the key point localization module 2 and receives the facial key point coordinates from the key point localization module 2. Based on this coordinate information and a heatmap mechanism, it generates a facial key point heatmap representing the child's pain level. The child pain intelligent assessment module 4 communicates with the key point heatmap representation module 3 and receives the facial key point heatmap from the key point heatmap representation module 3. Based on this heatmap, it assesses the child's pain and obtains the assessment result. The pain assessment result visualization module 5 communicates with the children's pain intelligent assessment module 4 to receive the assessment results sent by the children's pain intelligent assessment module 4 and display the assessment results in a visual way to facilitate medical staff's understanding and use.

[0073] Preferably, such as Figure 2 As shown, the face detection module 1 includes a face detection submodule 11 and an eye detection submodule 12. The face detection submodule 11 includes a face detection model 111, an eye detection model 112, and an eye detection model 113. The face detection model 111 performs face detection on the child's face image acquired by the face detection module 1 and outputs a face image; the eye detection model 112 performs eye detection on the child's face image acquired by the face detection module 1 and outputs an eye image; the eye detection model 113 performs eye detection on the child's face image acquired by the face detection module 1 and outputs an eye image.

[0074] In one embodiment of the present invention, the face detection model employs the MTCNN model, such as... Figure 3 As shown, the model includes a first-level detector 114, a second-level detector 115, and a third-level detector 116. The first-level detector 114 performs first-level face detection on the child's face image acquired by the face detection module 1; the second-level detector 115 performs second-level face detection on the child's face image acquired by the face detection module 1; and the third-level detector 116 performs third-level face detection on the child's face image acquired by the face detection module 1. The MTCNN model uses a downsampling mechanism to convert the intermediate results obtained from the first-level face detection of the child's face image into the input image size for the second-level and third-level detectors.

[0075] Specifically, the three-stage detection process of the MTCNN model is as follows: First, the first-stage detector uses a fully convolutional network to quickly scan the input image and generate candidate windows; second, the second-stage detector performs fine-grained detection on these candidate windows, filtering out most non-face regions; finally, the third-stage detector further refines the detection of face regions and outputs the locations of facial key points. This cascaded structure significantly improves the accuracy and efficiency of face detection. In practical applications, the scanning stride of the first-stage detector is typically set to 2 pixels, while the scanning strides of the second and third-stage detectors are set to 4 pixels and 8 pixels, respectively, to improve processing speed while ensuring detection accuracy.

[0076] Furthermore, in the MTCNN model, the trained eye detection model is used as the data processing model for children's facial key points to obtain the pixel coordinate information of children's facial key points, represented as:

[0077]

[0078] in, The weights are constants, typically ranging from 0.5 to 1.5. The two-dimensional pixel coordinates of the eye center output by the eye detection model; The output of the eye detection model is a two-dimensional pixel coordinate matrix of three key points, where each column represents the inner corner of the left eye. The outer corner of the left eye and the outer corner of the right eye The coordinates; This is the standardized coordinate matrix of key eye points obtained after transformation, where each column represents the standardized inner canthus of the left eye. The outer corner of the left eye and the outer corner of the right eye The coordinates.

[0079] In actual calculations, this formula represents the relative positional relationship between the coordinates of the eye's center and the coordinates of each corner of the eye. When calculating each key point separately, the formula expands to:

[0080] ,

[0081] This mapping method uses the stability of the eye region as a benchmark reference system and calculates more accurate facial key point locations through relative positional relationships, effectively solving the challenge of key point localization caused by facial expression changes in children under pain conditions.

[0082] Preferably, the key point localization module includes: a heatmap filtering submodule 21 and a key point coordinate extraction submodule 22; the heatmap filtering submodule 21 is used to perform mean filtering on the key point heatmap response values ​​output by the face detection module, using the following formula:

[0083] ,

[0084] In the formula, For heat map at location The response value at that location, The x-coordinate of the pixel in the heatmap. The vertical coordinate of the pixel in the heatmap. The filtered heatmap response value. This represents the number of pixels contained in the filter window. For the point The filtering window area centered on the center is typically... A square area of ​​size The recommended value is 3 or 5.

[0085] The key point coordinate extraction submodule 22 is used to perform local maximum detection on the filtered heatmap to obtain stable key point coordinates of the child's face, represented as:

[0086] ,

[0087] In the formula, The coordinates of the detected key points, For The local neighborhood centered on the center is usually taken for or The window size; this formula represents the point A point whose heatmap response value is higher than that of all other points in its neighborhood is identified as a key point. To further improve the stability of key point localization, the system also performs weighted smoothing on the key point coordinates of consecutive frames along the time axis:

[0088] ,

[0089] In the formula, Indicates time The smoothed keypoint coordinates, For a moment The coordinates of the key points were directly detected. For a moment The smoothed keypoint coordinates, The smoothing coefficient ranges from (0,1], typically between 0.3 and 0.7. A larger value is used for high confidence levels, and a smaller value for low confidence levels. This two-stage processing method (first performing spatial domain filtering on the heatmap, then coordinate smoothing in the time domain) avoids the meaningless translation problems that may result from directly applying mean filtering to coordinate values. It also fully utilizes the spatial distribution information and temporal continuity of the heatmap, significantly improving the stability and robustness of key point localization. It is particularly suitable for handling facial tremors and expression changes that may occur in children under pain conditions.

[0090] Mean filtering works by replacing the current pixel value with the average value of all pixels in the current pixel's neighborhood. This effectively removes high-frequency noise from the image, making the position of facial key points more stable. In practical applications, noise filtering is crucial for improving the stability of key point localization, especially for pediatric patients, as factors such as facial tremors and changes in lighting may cause noise interference in the original key point positions.

[0091] By employing noise filtering and keypoint coordinate extraction, the accuracy and stability of facial keypoint localization can be significantly improved. This is crucial for subsequent pain assessment, as even minor deviations in the location of facial keypoints can lead to significant changes in pain assessment results.

[0092] Preferably, the key point heatmap representation module 3 receives facial key point coordinate information sent by the key point positioning module 2, and generates a facial key point heatmap representing the child's pain level based on this coordinate information. The facial key point heatmap is represented as follows:

[0093] ,

[0094] in, Let be the heatmap matrix of the i-th key point in the facial region, representing the i-th key point relative to its coordinates. Thermodynamic influence value; Point and key points The Euclidean distance between them is calculated using the following formula: ,in It is the first The coordinates of the key points; This is a parameter that controls the diffusion range of the heat map, and its value is usually between 1 and 5. In this embodiment, a value of 2 is preferred, as a smaller value is acceptable. Values ​​that make the heatmap more concentrated, larger values The value will make the heatmap more dispersed; This represents the natural exponential function.

[0095] The core idea of ​​heatmap technology is to transform discrete key points into a continuous heat map distribution, thus capturing facial expression changes more comprehensively. In practice, a two-dimensional Gaussian distribution centered on each key point and decreasing with distance is generated. Then, the Gaussian distributions of all key points are superimposed to form the final heatmap. This method can effectively represent subtle changes in facial expressions, especially features commonly seen in expressions of pain, such as frown lines and closed eyes.

[0096] Preferably, such as Figure 6 As shown, the intelligent assessment module 4 for children's pain includes a facial key point heatmap localization submodule 41, a key point heatmap score extraction submodule 42, a children's pain model training submodule 43, and a children's pain assessment submodule 44.

[0097] The facial key point heatmap localization submodule 41 is used to locate the detected facial regions using the facial key point heatmap sent by the key point heatmap representation module 3, thereby obtaining a facial key point heatmap. The key point heatmap score extraction submodule 42 is communicatively connected to the facial key point heatmap localization submodule 41 and is used to obtain the scores for all key point heatmaps based on the facial key point heatmap sent by the key point heatmap representation module 3. The scores for all key point heatmaps are represented as follows:

[0098] ,

[0099] in, For the first The heatmap score of each key point indicates the contribution of that key point to the expression of pain. This indicates the operation of retrieving the maximum value; This represents the coordinates of key points on the heatmap; Indicates the first The response area of ​​a key point is defined as a rectangular area centered on the key point, usually a 5×5 or 7×7 area around the key point. For the first Key points at coordinates The heat value at that location.

[0100] The key point heatmap score extraction process involves extracting the maximum response value of each key region from the heatmap, which serves as the pain feature score for that region. This method can accurately capture key information in facial expression changes, providing reliable feature input for subsequent pain assessment.

[0101] The pediatric pain model training submodule 43 is communicatively connected to the keypoint heatmap score extraction submodule 42, and is used to train the pediatric pain model based on all keypoint heatmap scores output by the keypoint localization module 2. The pediatric pain assessment submodule 44 is communicatively connected to both the keypoint heatmap score extraction submodule 42 and the pediatric pain model training submodule 43, and is used to input the results output by the keypoint heatmap score extraction submodule 42 into the pediatric pain model for classification to obtain the assessment results.

[0102] In one embodiment of the present invention, such as Figure 7 As shown, the pediatric pain assessment model is built using the Dlib library. This assessment model includes a pain assessment model and a no-pain assessment model. The pain assessment model is used to assess whether a child is in pain; the no-pain assessment model is used to assess whether a child is not in pain. Both models are obtained from the pediatric pain model training submodule 43 using a preset number of training set data, which is 50-100 sets. The training set data includes facial key point heatmap scores for children, and each set of training data contains a child's facial image and its labeled pain state (pain or no pain).

[0103] The advantage of using a dual-model structure is that it can assess pain from both positive and negative perspectives, improving the accuracy of the judgment. In practical applications, the training dataset typically contains 50-100 sets of labeled facial expression samples from children, with pain samples and non-pain samples each accounting for about half. This amount of data ensures the effectiveness of model training without leading to overfitting. For children of specific age groups, the composition of the training samples can be appropriately adjusted. For example, for infants and toddlers, the proportion of crying samples can be increased; for preschool children, the proportion of facial features such as frowning and tightly closed eyes can be increased.

[0104] The key point heatmap score extraction submodule 42 outputs a child pain score matrix and a child no-pain score matrix. These two matrices are subtracted using heatmap subtraction to obtain the child's pain assessment score, represented as:

[0105] ,

[0106] Among them, Score is the child's pain assessment score, which represents the probability of the child's current pain state, and the value range is usually between [-1, 1]. The pain score matrix for children is the score output by the pain assessment model, representing the degree to which pain features are detected. The pain-free score matrix for children is the score output by the pain-free assessment model, representing the degree to which pain-free features are detected.

[0107] The core idea of ​​this difference calculation method is to eliminate interference from individual differences and improve the accuracy of the assessment by comparing the outputs of two assessment models in opposite directions. In practice, when the assessment score is positive and the absolute value is large, it indicates that the child is more likely to be in pain; when the assessment score is negative or close to zero, it indicates that the child is more likely to be in a pain-free state.

[0108] The assessment results of facial pain assessment in children are expressed as follows:

[0109] ,

[0110] in, The pain assessment results for children are the final binary classification results, which can be "pain present" or "no pain present"; The preset pain score threshold is usually between 0.3 and 0.7. In this embodiment, a value of 0.5 is preferred, which is the optimal judgment threshold determined based on clinical verification. The pain assessment score for children calculated above. When When the child is in pain, the intelligent pain assessment module 4 outputs an assessment result indicating that the child is in pain; when When the time is specified, it indicates that the child is not in pain. The assessment result output by the child pain intelligent assessment module 4 is "no pain".

[0111] threshold The threshold setting significantly impacts the accuracy of the assessment results. Setting the threshold too low may lead to too many false positives, while setting it too high may result in too many false negatives. In practical applications, the threshold can be adjusted based on the characteristics of children of different age groups and clinical needs. For example, for younger children, the threshold can be appropriately lowered to 0.3-0.4 to improve the pain detection rate; for older children, the threshold can be appropriately increased to 0.6-0.7 to reduce misjudgments.

[0112] Preferably, such as Figure 8 As shown, the pain assessment result visualization module 5 receives the assessment results sent by the pediatric pain intelligent assessment module 4 and displays the results in a visual manner. Visualization methods include, but are not limited to: facial key point connection diagrams, heat map displays, pain / no pain text labels, and pain intensity level displays. This intuitive visualization method facilitates quick understanding of the assessment results by medical staff, improving the efficiency of clinical decision-making.

[0113] In practical applications, the system can select different visualization methods according to different application scenarios. For example, in an emergency environment, the pain / no pain assessment results and pain level can be displayed first to enable medical staff to make quick decisions; in long-term pain management, pain trend graphs can be added to help medical staff track changes in pain; and in pediatric wards, a more user-friendly interface design can be used to reduce children's fear.

[0114] The workflow of this invention is as follows: First, the face detection module 1 acquires a child's facial image and performs a three-level cascaded detection using the MTCNN model to obtain the pixel coordinate information of facial key points; then, the key point localization module 2 performs noise filtering on these coordinate information to obtain more accurate facial key point coordinate information; next, the key point heatmap representation module 3 generates a facial key point heatmap representing the child's pain level based on these coordinate information; then, the child pain intelligent assessment module 4 assesses the child's pain based on the facial key point heatmap using a dual-model comparison mechanism to obtain the assessment result; finally, the pain assessment result visualization module 5 displays the assessment result in a visual manner to assist medical staff in pain management and treatment decisions.

[0115] In this embodiment, the system can be deployed in hospital pediatric wards, emergency departments, operating rooms, and other locations requiring pain assessment for children. The hardware required for system operation includes, but is not limited to, high-definition cameras, computing servers, and display terminals. The system software can adopt a modular design, facilitating customization and optimization to meet the needs of different medical environments.

[0116] The AI-based facial expression analysis-based intelligent pain assessment system for children provided by this invention allows medical staff to obtain objective, accurate, and real-time pain assessment results for children, effectively solving the problems of strong subjectivity and low reliability of traditional assessment methods, and providing a powerful tool for children's pain management.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A child pain intelligent assessment system based on AI facial expression analysis, characterized in that, include: The face detection module is used for: Obtain images of children's faces; The child's facial image is input into a multi-task cascaded convolutional neural network model for three-level cascaded detection to obtain the pixel coordinate information of key points on the child's face. The key point localization module, which is communicatively connected to the face detection module, is used for: Receive the pixel coordinate information of key points on the child's face sent by the face detection module; The pixel coordinate information of the child's facial key points is subjected to noise filtering to obtain the facial key point coordinate information. The key point heatmap representation module, which is communicatively connected to the key point localization module, is used for: Receive facial key point coordinate information sent by the key point positioning module; Based on the combination of the facial key point coordinate information and the heat map mechanism, a facial key point heat map representing the degree of pain in children is generated. The intelligent pain assessment module for children is communicatively connected to the key point heatmap representation module and is used for: Receive the facial key point heatmap sent by the key point heatmap representation module; Based on the facial key point heat map, the child's pain was assessed, and the assessment results were obtained. The pain assessment result visualization module, which is communicatively connected to the intelligent pain assessment module for children, is used for: Receive the assessment results sent by the child pain intelligent assessment module; Visualize the evaluation results; The key point localization module includes: a heatmap filtering submodule and a key point coordinate extraction submodule; The heatmap filtering submodule is used to perform mean filtering on the key point heatmap response values ​​output by the face detection module, and the formula is: ， In the formula, For heat map at location The response value at that location, The x-coordinate of the pixel in the heatmap. The vertical coordinate of the pixel in the heatmap. The filtered heatmap response value. This represents the number of pixels contained in the filter window. For the point The filtering window area centered on the center.

2. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 1, characterized in that, The intelligent pain assessment module for children includes: The facial key point heatmap localization submodule is used to locate the detected facial region based on the facial key point heatmap sent by the key point heatmap representation module, and obtain the facial key point heatmap. The key point heatmap score extraction submodule is communicatively connected to the facial key point heatmap positioning submodule and is used to obtain the scores of all key points heatmaps based on the facial key point heatmap sent by the facial key point heatmap representation module. The child pain model training submodule is communicatively connected to the key point heatmap score extraction submodule and is used to train the child pain model based on all key point heatmap scores output by the key point localization module. The child pain assessment submodule is communicatively connected to the key point heatmap score extraction submodule and the child pain model training submodule. It is used to input the output of the key point heatmap score extraction submodule into the child pain model for classification to obtain the assessment result.

3. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 2, characterized in that, The facial key point heatmap is represented as follows: ， in, Let be the heatmap matrix of the i-th key point in the facial region; Indicates will Pixel values ​​normalized to (0,1); Point and points The Euclidean distance between them; The heatmap represents the first... Each region Parameters for controlling the diffusion range of the heat map, It is an exponential function.

4. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 2, characterized in that, The heatmap scores for all key points are represented as follows: ， in, The heatmap fraction matrix for key points. The first one represented in the heat map The scores for each region.

5. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 2, characterized in that, The child pain assessment model is constructed using the Dlib library; the child pain assessment model includes: a pain assessment model and a no-pain assessment model; the pain assessment model is used to assess whether a child is in pain; the no-pain assessment model is used to assess whether a child is not in pain; the pain assessment model and the no-pain assessment model are obtained using a preset number of training set data from the child pain model training submodule, the preset number being 50-100; the training set data includes the child's facial key point heatmap scores.

6. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 5, characterized in that, The output of the key point heatmap score extraction submodule includes: a child pain score matrix and a child no-pain score matrix; the child pain score matrix and the child no-pain score matrix are subtracted using a heatmap subtraction method to obtain the child pain assessment score, represented as: ， in, For children's pain assessment scores, For children's pain score matrix, A pain-free score matrix for children.

7. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 5, characterized in that, The assessment results of facial pain assessment in children are expressed as follows: ， in, For the results of pain assessment in children, The preset pain score threshold; when When the child is in pain, the intelligent pain assessment module outputs an assessment result indicating that the child is in pain; when When the time is specified, it indicates that the child is not in pain, and the assessment result output by the child pain intelligent assessment module is "no pain".

8. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 1, characterized in that, The face detection module includes a face detection submodule and an eye detection submodule. The face detection module includes a face detection model, an eye detection model, and an eye detection model. The face detection model performs face detection on the child's face image acquired by the face detection module and outputs a face image. The eye detection model performs eye detection on the child's face image acquired by the face detection module and outputs an eye image. The eye detection model performs eye detection on the child's face image acquired by the face detection module and outputs an eye image. The face detection model uses the MTCNN model, which includes a first-level detector, a second-level detector, and a third-level detector. The first-level detector performs first-level face detection on the child's face image acquired by the face detection module. The second-level detector performs second-level face detection on the child's face image acquired by the face detection module. The third-level detector performs third-level face detection on the child's face image acquired by the face detection module. The MTCNN model uses a downsampling mechanism to convert the intermediate results obtained from the first-level face detection of the child's face image into the input image size for the second-level and third-level detectors.

9. The intelligent pain assessment system for children based on AI facial expression analysis according to claim 8, characterized in that, In the MTCNN model, the trained eye detection model is used as the child's facial key point data processing model to obtain the pixel coordinate information of the child's facial key points, which is represented as: ， in, The weights are constants, typically ranging from 0.5 to 1.

5. The two-dimensional pixel coordinates of the eye center output by the eye detection model; The output of the eye detection model is a two-dimensional pixel coordinate matrix of three key points, where each column represents the inner corner of the left eye. The outer corner of the left eye and the outer corner of the right eye The coordinates; This is the standardized coordinate matrix of key eye points obtained after transformation, where each column represents the standardized inner canthus of the left eye. The outer corner of the left eye and the outer corner of the right eye The coordinates.

Images