OSA discrimination method based on face and neck contour image features

Abstract

The invention discloses an OSA judgment method based on face and neck contour image features, and relates to the technical field of OSA judgment. Comprising the following steps: acquiring a face and neck contour image of a to-be-detected individual through an image acquisition module, extracting face and neck contour feature values of an OSA risk from the face and neck contour image based on a feature extraction module, analyzing the eight face and neck contour feature values through a risk assessment module, and judging the risk level of the to-be-detected individual suffering from the OSA, the method has the advantages that the anatomical features, namely the palatopharyngeal length and the mandibular body length, which are most related to the OSA severity are converted into the human body external features, diagnosis can be performed by photographing the human body appearance image, the requirement on equipment is low, the diagnosis process is simplified, and the diagnosis accuracy is improved by combining the edge and skin color information. The face recognition, the region segmentation and the proportion measurement can be efficiently realized, the most convenient and accurate measurement is realized, and a step-by-step judgment mode is adopted, so that the accuracy is improved, the updating logic is optimized, and the subsequent maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of OSA determination, and in particular to an OSA determination method based on face and neck contour image features. Background Art

[0002] Obstructive sleep apnea (OSA) refers to a sleep-related breathing disorder caused by intermittent or persistent partial or complete obstruction of the upper airway during sleep, which interferes with normal ventilation and sleep patterns. OSA patients often have symptoms such as snoring, daytime sleepiness, fatigue and persistent morning headaches, and the incidence of motor vehicle accidents increases. At the same time, OSA can cause a series of health problems. Untreated OSA increases the risk of hypertension, coronary heart disease, heart failure, myocardial infarction, stroke, arrhythmia and insulin resistance. According to the Sleep and Heart Health Study, after controlling for major potential confounders (sex, age, and body mass index) and other related variables (drinking, smoking), patients with higher levels of AHI (apnea-hypopnea index) or nocturnal hypoxia (the time with the lowest blood oxygen saturation < 90% accounts for ≥ 12% of the total sleep time) had a higher prevalence of systemic hypertension, and compared with treated OSA patients and various control groups, severe untreated OSA was associated with a significantly increased incidence of fatal (death from myocardial infarction or stroke) and non-fatal cardiovascular events (non-fatal myocardial infarction, stroke, acute coronary insufficiency). The results of a prospective longitudinal epidemiological study by SHHS showed that men with AHI ≥ 30 times / hour were 58% more likely to develop heart failure than men with AHI < 5 times / hour. The SHHS and Community Atherosclerosis Risk Study confirmed that after adjusting for confounders (body mass index, waist circumference), patients with severe OSA had a greater risk of developing diabetes compared with normal people, indicating a causal relationship between severe OSA and the risk of diabetes.

[0003] In the prior art, polysomnography (PSG) technology is used to diagnose OSA. It can provide a large amount of information to determine the type and severity of the disease and is considered to be the gold standard for diagnosing OSA. The test simultaneously monitors sleep and respiratory parameters, records electroencephalogram, electrooculogram and mandibular electromyogram to monitor whether the patient is in a sleeping or awake state, and uses chest and abdominal respiratory induction plethysmography, nasal air pressure and thermal sensors, and finger blood oxygen sensors to monitor respiratory airflow and arterial oxygen saturation. In addition, posture and leg movement sensors are included to monitor the patient's body position and abnormal leg movements. However, the cost of PSG testing is expensive, and in many places around the world, there are no conditions to purchase PSG testing equipment and establish a professional testing team. In addition, PSG testing cannot show the specific location of the obstruction. Summary of the invention

[0004] The purpose of the present invention is to provide an OSA identification method based on face and neck contour image features, which solves the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: an OSA identification method based on face and neck contour image features, comprising the following identification method: Step S1: acquiring a face and neck contour image of the individual to be tested through an image acquisition module; Step S2: extracting eight facial and neck contour feature values ​​of OSA risk from the facial and neck contour image based on the feature extraction module; The eight facial and neck contour characteristic values ​​are: lower 1 / 3 height of face, chin-neck depth, neck depth, chin-neck angle, mandibular soft tissue width, neck width, chin-neck height and neck height; Step S3: Analyze the eight facial and neck contour feature values ​​through the risk assessment module and determine the risk level of OSA of the individual to be tested.

[0006] Optionally, the feature extraction module includes: Face recognition submodule: Use OpenCV and dlib to perform face recognition and target parameter extraction on the face and neck contour image; Feature point alignment submodule: detect facial feature points and align them to standard positions; Eigenvalue calculation submodule: Apply edge detection and skin color detection, combine edge information and skin color information to segment and measure the relevant parameters of facial parts, and obtain eight facial and neck contour eigenvalues.

[0007] Optionally, the risk assessment module uses a multi-layer perceptron MLP model, and the MLP model includes an input layer, a hidden layer and an output layer; Input layer: used to receive eight facial and neck contour feature values; Hidden layer: used for nonlinear transformation and feature extraction of input feature values; Output layer: used to output the risk level of OSA of the individual to be tested; The risk assessment module includes: a characteristic value assessment submodule, a prediction submodule and an adjustment submodule.

[0008] Optionally, the calculation formula of the eigenvalue evaluation submodule is as follows:

[0009] Among them: FFUA refers to the comprehensive evaluation index of characteristic values, Ha refers to the lower 1 / 3 height of the face, Wf refers to the facial width, Dc refers to the chin-neck depth, Hn refers to the neck height, Wj refers to the mandibular soft tissue width, Wn refers to the neck width, HC refers to the chin-neck height, Dn refers to the neck depth, KT refers to the chin-neck angle, A1, A2, A3, A4 and A5 refer to the weight coefficient one, weight coefficient two, weight coefficient three, weight coefficient four and weight coefficient five respectively; The processing process of the eigenvalue evaluation submodule is as follows: the lower 1 / 3 height Ha, facial width Wf, chin-neck depth Dc, neck height Hn, mandibular soft tissue width Wj, neck width Wn, chin-neck height HC, neck depth Dn and chin-neck angle KT are input into the eigenvalue evaluation submodule, and the eigenvalue comprehensive evaluation index FFUA is output.

[0010] Optionally, the calculation formula of the prediction submodule is as follows:

[0011] Among them: RVLU refers to the OSA risk prediction value, RAG refers to age, RGD refers to gender, the output for males is 1 and the output for females is 0, BMI refers to body mass index, B1, B2, B3 and B4 refer to model parameter 1, model parameter 2, model parameter 3 and model parameter 4 respectively; The processing process of the prediction submodule is as follows: based on the characteristic value comprehensive evaluation index FFUA, age RAG, gender RGD and body mass index BMI are input into the prediction submodule, and the OSA risk prediction value RVLU is output.

[0012] Optionally, the calculation formula of the adjustment submodule is as follows: AHIA=C1×RVLU+C2×(RVLU-RVLUa) 2 +V0; Where: AHIA refers to the adjusted predicted value, RVLUa refers to the mean RVLU, which is used to adjust the center position of the square term, V0 refers to the constant term, which is used to adjust the overall offset of the predicted value, C1 and C2 refer to the first model parameter and the second model parameter, respectively; The processing process of the adjustment submodule is as follows: the OSA risk prediction value RVLU is input into the adjustment submodule, and based on the calculation of the OSA risk prediction value RVLU and the RVLU average RVLUa, the adjusted prediction value AHIA is output.

[0013] Optionally, the face recognition submodule includes: an image reading unit, an image conversion to grayscale unit, a face detection unit and a detection result drawing unit.

[0014] Optionally, the image reading unit is: using OpenCV to read an image file that needs to be subjected to face detection, and loading the image data from a storage medium into a memory for subsequent processing; The image conversion to grayscale unit is: converting the image data into a grayscale image; The face detection unit is: using a dlib face detector to detect the position of a face in a grayscale image; The detection result drawing unit is to draw a rectangular frame on the image data and mark the detected face position.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention converts the anatomical features of velopharyngeal length and mandibular body length, which are most relevant to the severity of OSA, into external features of the human body. Since the velopharyngeal length and mandibular body length may require professional equipment to be photographed, which is costly and bulky, it is not suitable for patients to self-test or doctors to go out for medical treatment. The eight feature values ​​of this method only require taking pictures of the human body's appearance, and have low requirements for equipment. With the combination of mini-programs and other software, patients can complete self-testing at home.

[0016] 2. The present invention utilizes OpenCV and dlib to perform face recognition and target parameter extraction on images. By combining edge and skin color information, face recognition, region segmentation and proportion measurement can be efficiently realized. Feature point alignment and data standardization are used to free the measurement process from the influence of shooting angle and shooting distance, thus realizing the most convenient and accurate measurement.

[0017] 3. The present invention introduces the sleep apnea hypopnea syndrome assessment standard at the OSA risk point, combines the concept of the disease with the specific manifestations in daily life, improves credibility and can well reflect the patient's condition. In addition, a step-by-step judgment method is adopted to improve accuracy while optimizing the update logic and reducing subsequent maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the steps of the OSA identification method based on face and neck contour image features; Figure 2 This is a schematic diagram of face position identification of the OSA identification method based on face and neck contour image features; Figure 3 Schematic diagram of facial edge detection of the OSA discrimination method based on face and neck contour image features; Figure 4 This is a schematic diagram of target position extraction of the OSA discrimination method based on face and neck contour image features; Figure 5This is a schematic diagram of the sleep apnea hypopnea syndrome assessment standard of the OSA discrimination method based on the face and neck contour image features; Figure 6 Schematic diagram of the facial and neck contour features of the OSA discrimination method based on facial and neck contour image features Figure 1 ; Figure 7 Schematic diagram of the facial and neck contour features of the OSA discrimination method based on facial and neck contour image features Figure 2 ; Figure 8 This is the MLP structure diagram of the OSA discrimination method based on face and neck contour image features; Fig. 9 This is the MLP data flow diagram of the OSA discrimination method based on face and neck contour image features; Fig.10 This is the MLp optimization update flow chart of the OSA discrimination method based on face and neck contour image features; Fig.11 This is a comparison chart of the velopharyngeal length and mandibular body length for the OSA discrimination method based on face and neck contour image features. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] It should be noted that: The etiology of OSA is still unclear. It is currently believed to be multifactorial, including anatomical and non-anatomical factors. Anatomical factors refer to anatomical narrowing or easy collapse of the upper airway, which is mainly caused by obesity or abnormal craniofacial morphology. It is a key determinant of OSA. Related studies have explored the relationship between three-dimensional craniofacial and upper airway morphological variables and the severity of OSA based on large sample sizes of CBCT imaging data. The results show that three-dimensional craniofacial and upper airway morphology play an important role in the severity of OSA. The anatomical features most relevant to the severity of OSA are the velopharyngeal length and mandibular body length. The length comparison between patients and normal people, for example, Fig.11 As shown, these two variables can be used to help identify severe,OSA patients and serve as evaluation indicators for selecting the most appropriate treatment for OSA patients; However, as parameters of the internal structure of the oral cavity, the velopharyngeal length and mandibular body length are usually measured by CBCT, which is a cone beam computed tomography device that takes pictures and then builds images. CBCT equipment is expensive and bulky, so the measurement cost is high and the usage scenarios are limited. Therefore, the external manifestations of velopharyngeal length and mandibular body length in the human body are analyzed. After comparing the craniofacial morphological data of patients with and without OSA, the present invention selects fifteen features closely related to velopharyngeal length and mandibular body length, namely, skull length, skull width, facial width, lower 1 / 3 height of face, nose width, nose height, chin-neck depth, neck length, neck height, neck width, neck depth, cervical curvature, chin-neck angle, mandibular soft tissue width, chin-neck height, and then performs correlation analysis on them through random forest; Random forest is an ensemble learning method based on decision trees. It constructs multiple decision trees for classification and finally decides the final result by voting or averaging. The normalized training data is fed into the model and the training set is used to train the random forest model. Each tree uses the bootstrap sampling method to extract samples from the training set during training and introduces randomness in feature selection. In each tree, the contribution of each feature to the node split is calculated and the feature contributions of all trees in the forest are averaged to obtain the overall importance score of each feature. After multiple calculations of the existing data, eight groups of features with greater correlation were selected, namely, lower 1 / 3 height of face, chin-neck depth, neck depth, chin-neck angle, mandibular soft tissue width, neck width, chin-neck height and neck height, as external features reflecting the degree of OSA. Their specific locations are as follows: Figure 6 and Figure 7 As shown, by analyzing the facial and neck contour features, the velopharyngeal length and mandibular body length can be inferred to determine the probability of OSA; Regarding the OSA discrimination method based on face and neck contour image features, different from the existing OSA judgment method, the unique innovation of this method is: the present invention proposes a new judgment method for the key factors affecting OSA, namely, the velopharyngeal length and mandibular body length, that is, a method of reflecting the internal state through external features, and selects fifteen features. After correlation analysis, eight external manifestations that can reflect its state are obtained, namely, the lower 1 / 3 height of the face, the chin neck depth, the neck depth, the chin neck angle, the mandibular face soft tissue width, the neck width, the chin neck height and the neck height. This kind of judgment method can greatly reduce the cost and complexity of conventional CT; In the extraction of eight feature quantities, since the feature points involved include not only the conventional feature points such as eyes and mouth corners, but also the neck and jaw junction, this paper proposes a regional segmentation method that combines edge and skin color information. Based on OpenCV and dlib, face recognition and target parameter extraction are performed on the image. On the basis of conventional feature point and edge detection, the binary image of edge information and the mask of skin color information are ANDed, and then the relevant parameters of the face parts are segmented and measured. In judging whether a patient suffers from OSA based on feature ratio, the present invention proposes a new step-by-step multi-layer perceptron scheme, which divides the judging process into two steps, can be optimized and updated in layers, can effectively reduce the cost of retraining the model, introduces the sleep apnea hypopnea syndrome assessment standard, and uses AHI as a quantitative indicator of the disease.

[0021] Example 1: Please refer to Figures 1 to 11 This embodiment provides an OSA identification method based on face and neck contour image features, including the following identification methods: Step S1: acquiring a face and neck contour image of the individual to be tested through an image acquisition module; Step S2: extracting eight facial and neck contour feature values ​​of OSA risk from the facial and neck contour image based on the feature extraction module; The eight facial and neck contour characteristic values ​​are: lower 1 / 3 height of face, chin-neck depth, neck depth, chin-neck angle, mandibular soft tissue width, neck width, chin-neck height and neck height; Step S3: Analyze the eight facial and neck contour feature values ​​through the risk assessment module, and determine the risk level of OSA of the individual to be tested; The feature extraction module includes: Face recognition submodule: Use OpenCV and dlib to perform face recognition and target parameter extraction on the face and neck contour image; Among them: face recognition and target parameter extraction of images based on OpenCV and dlib. OpenCV, or Open Source Computer Vision Library, is a cross-platform open source software library for computer vision processing. Dlib uses HOG, or Histogram of Oriented Gradients, combined with SVM, or Support Vector Machine, for face detection. Compared with traditional Haar features and cascade classifiers, HOG-SVM is more accurate in face detection, especially for frontal and slightly non-frontal face detection. Dlib's face detection algorithm runs fast on the CPU and is suitable for real-time application scenarios. Feature point alignment submodule: detect facial feature points and align them to standard positions; Eigenvalue calculation submodule: Apply edge detection and skin color detection, combine edge information and skin color information to segment and measure the relevant parameters of the face parts, and obtain eight facial and neck contour eigenvalues; Among them: the feature extraction module is specifically: First, feature point detection; Based on the grayscale image of the face obtained in the first step, the dlib algorithm is used to obtain the absolute coordinates of 68 facial feature points. This process outputs a two-dimensional matrix of coordinates. Second, feature point alignment; The position coordinates of facial feature points are used as reference points, and facial feature points such as the corners of the eyes and mouth are aligned to standard positions to reduce the influence of head posture. The feature points are aligned to ensure that their positions in the image are standardized, and then the facial image is converted into a standard coordinate system to achieve subsequent detection, providing a standard facial image for subsequent detection. The output of this process is an image; Third, edge detection; The Canny edge detector is applied to the face area in the standard coordinate system to identify the facial contour and generate a binary image, including the eyes, mouth and facial edges. The influence of the environment and background is eliminated, and the recognition effect is as follows: Figure 3 As shown, this process extracts a binary image of the human body; Fourth, skin color detection; Convert the face area in the standard coordinate system to the HSV color space, and create a skin color mask based on the approximate range of skin color in the HSV space. The output of this process is the color mask of each pixel. Fifth, combine edge and skin color information; The Canny edge detection result is logically ANDed with the skin color mask to enhance the edge of the face. Since the face has been aligned and modulated according to the feature points, the corresponding position features are intercepted to obtain eight feature quantity images. The effect is as follows: Figure 4 As shown, this process decomposes the image into eight feature quantities; The risk assessment module uses a multi-layer perceptron MLP model, which includes an input layer, a hidden layer, and an output layer; Input layer: used to receive eight facial and neck contour feature values; Hidden layer: used for nonlinear transformation and feature extraction of input feature values; Output layer: used to output the risk level of OSA of the individual to be tested; The risk assessment module includes: a characteristic value assessment submodule, a prediction submodule and a regulation submodule.

[0022] In this embodiment: Through the continuous calculation of the three submodules in the risk assessment module, an accurate assessment of OSA risk can be achieved. Starting from the facial and neck contour features, it gradually goes into risk prediction and severity assessment. The operation is simple and low-cost. This method is mainly based on image processing and simple mathematical calculations. It does not require complex equipment and operations, which reduces costs and is easy to promote in primary medical institutions. Early diagnosis and treatment support. This method can provide strong support for the early diagnosis and treatment of OSA, help doctors to timely detect and intervene in OSA patients, and improve their quality of life. This method can be continuously optimized and updated. With the increase in data volume and continuous training of the model, the accuracy and reliability of this method will be continuously improved, and it can meet the needs of different populations and application scenarios. The risk assessment modules are as follows: First, pixel length The pixel length of the feature is calculated based on the captured feature image. The specific method is to enter the queue and calculate in columns according to the corresponding screenshot of the feature. Figure 3 The coordinates of the feature points and contours in the pixel matrix are selected, and the maximum and minimum values ​​on the pixel matrix column are selected, and the difference is the length of the corresponding feature value; Second, data standardization Since pixel length is not the actual distance, it represents the size of the area occupied in the image. In order to compare the proportions of facial features of different individuals, the measurement results need to be standardized. This is achieved by dividing all pixel lengths by the width of the face to determine the proportional relationship between facial features and eliminate the influence of the distance of the photo. Among them: In terms of evaluation criteria, in order to have a quantifiable indicator of the patient's probability of illness, the sleep apnea hypopnea syndrome assessment standard is introduced, which stipulates the apnea hypopnea index AHI and defines the severity of the symptoms, such as Figure 5 As shown; The evaluation process is divided into two steps. The first step is to cluster the data into non-disease, mild, moderate and severe. The second step is to calculate the deviation value, subdivide the AHI times and improve the accuracy. The evaluation model uses the multilayer perceptron MLP, which is a feedforward neural network consisting of an input layer, three hidden layers and an output layer. The aforementioned eight feature quantities are used as the input of the multilayer perceptron. Its structure is as follows: Figure 8 As shown; The two-step evaluation process uses two MLP models to analyze the data separately. foundation The disease level is obtained, corresponding to AHI of 0, 10, 20 and 40 times / hour; MLP deviation Get the AHI deviation at various levels, sum it up to get the final AHI value, the input-output relationship diagram is as follows Fig. 9 As shown; During use, as the amount of data increases, the model is optimized and retrained regularly. Based on the logic of distribution judgment, the optimization is divided into regular optimization and quantitative optimization. The flow chart is as follows: Fig.10 As shown, the regular optimization is to perform MLP deviation The model is updated, and the quantitative optimization is to update the two models when the number of newly added samples increases exponentially; When this risk assessment module performs OSA assessment based on the constructed model, a data set is established. The data set is patient data provided by the dental hospital, with a training set of 400 people and a test set of 160 people, of which OSA patients account for about 40%. The recognition accuracy after training is 96.4%; See also Figures 1 to 11 , the processing process of the eigenvalue evaluation submodule is as follows:

[0023] Among them: FFUA refers to the comprehensive evaluation index of characteristic values, Ha refers to the lower 1 / 3 height of the face, Wf refers to the facial width, Dc refers to the chin-neck depth, Hn refers to the neck height, Wj refers to the mandibular soft tissue width, Wn refers to the neck width, HC refers to the chin-neck height, Dn refers to the neck depth, KT refers to the chin-neck angle, A1, A2, A3, A4 and A5 refer to the weight coefficient one, weight coefficient two, weight coefficient three, weight coefficient four and weight coefficient five respectively; The processing process of the eigenvalue evaluation submodule is as follows: the lower 1 / 3 height of the face Ha, facial width Wf, chin-neck depth Dc, neck height Hn, mandibular soft tissue width Wj, neck width Wn, chin-neck height HC, neck depth Dn and chin-neck angle KT are input into the eigenvalue evaluation submodule, and the eigenvalue comprehensive evaluation index FFUA is output.

[0024] In this embodiment: FFUA comprehensively evaluates OSA risk by calculating the weighted sum of multiple facial and neck contour features, taking into account multiple relevant factors, improving the comprehensiveness and accuracy of the assessment. Different weight coefficients are assigned to each feature value according to its relevance to OSA risk. This weighted sum method can highlight the impact of key features on risk assessment. FFUA converts complex facial and neck contour features into a single numerical indicator, which is convenient for subsequent risk prediction and level judgment. FFUA can be regarded as a preliminary comprehensive assessment of the facial and neck contour characteristics, which reflects multiple aspects of OSA risk. foundation In the model, this preliminary assessment is used to predict the disease level, so the calculation of FFUA can be regarded as MLP foundation Part of model input preparation; See also Figures 1 to 11 , the prediction submodule processing process is as follows:

[0025] Among them: RVLU refers to the OSA risk prediction value, RAG refers to age, RGD refers to gender, the output for males is 1 and the output for females is 0, BMI refers to body mass index, B1, B2, B3 and B4 refer to model parameter 1, model parameter 2, model parameter 3 and model parameter 4 respectively; The processing of the prediction submodule is as follows: based on the eigenvalue comprehensive evaluation index FFUA, age RAG, gender RGD and body mass index BMI are input into the prediction submodule, and the OSA risk prediction value RVLU is output.

[0026] In this embodiment: RVLU is not only based on FFUA, but also combines other physiological or pathological factors, such as age, gender, BMI, etc., so that risk assessment is more comprehensive and personalized. The nonlinear relationship in the form of an exponential function is used for risk prediction, which can capture the complex relationship between variables and improve the accuracy of prediction. RVLU, as a continuous value between 0 and 1, provides a fine measurement of OSA risk, which helps to more accurately determine the risk level; RVLU is calculated based on FFUA and other physiological or pathological factors, such as age, gender, and BMI, through a nonlinear functional relationship. FFUA is conceptually similar to MLP. foundation The model output is closely related to the disease level, and RVLU provides the basis for risk prediction; See also Figures 1 to 11 , the processing of the adjustment submodule is as follows: AHIA=C1×RVLU+C2×(RVLU-RVLUa) 2 +V0; Where: AHIA refers to the adjusted predicted value, RVLUa refers to the mean RVLU, which is used to adjust the center position of the square term, V0 refers to the constant term, which is used to adjust the overall offset of the predicted value, C1 and C2 refer to the first model parameter and the second model parameter, respectively; The processing process of the adjustment submodule is as follows: the OSA risk prediction value RVLU is input into the adjustment submodule, and based on the calculation of the OSA risk prediction value RVLU and the RVLU average RVLUa, the adjusted prediction value AHIA is output.

[0027] In this embodiment: AHIA is an apnea-hypopnea index predicted by another nonlinear function relationship based on RVLU. AHIA is an important indicator for evaluating the severity of OSA. According to the value of AHIA, the OSA risk level of an individual is judged with reference to the sleep apnea-hypopnea syndrome assessment standard. This submodule makes full use of the information in the previous steps to improve the accuracy of AHIA prediction. The model parameters in this submodule are obtained through training, so that the model can automatically adjust the parameters according to the actual data, thereby enhancing the generalization ability of the model. AHIA is an important indicator for evaluating the severity of OSA. AHIA provides a direct prediction of the severity of OSA, which helps doctors formulate treatment plans. AHIA is predicted based on RVLU through another nonlinear function relationship, which directly corresponds to the predicted value of apnea-hypopnea index AHI using MLP deviation The model obtains the AHIA deviation at various levels, which means that after the basic AHIA prediction value has been obtained, the MLPdeviation model is used to further adjust this prediction value to take into account the AHIA deviation at different disease levels. Therefore, the calculation of AHIA can be regarded as the implementation of this step, that is, obtaining the final AHIA value. In general, FFUA and RVLU are calculated as MLP foundation The model provides the inputs and the AHI is calculated similarly to MLP deviation The process by which the model adjusts the base AHI prediction value; Example 2: Please refer to Figures 1 to 11 The face recognition submodule includes: an image reading unit, an image conversion to grayscale unit, a face detection unit and a detection result drawing unit. The image reading unit is: Use OpenCV to read the image file that needs to be detected for face, and load the image data from the storage medium into the memory for subsequent processing; The image conversion to grayscale unit is: converting the image data into a grayscale image; Among them: grayscale images reduce the amount of data processed, and many visual features are easier to detect in grayscale space, and face detectors usually have better performance on grayscale images; The face detection unit is: using the dlib face detector to detect the position of the face in the grayscale image; Among them: the face detector analyzes the image content and finds areas that match the facial features in the training data, and these areas are considered to be faces; The detection result drawing unit is: drawing a rectangular frame on the image data and marking the detected face position; Among them: Drawing a rectangular frame is to visualize the detection results, so that users can intuitively see the accuracy and location of face detection, such as Figure 2 shown.

[0028] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An OSA discrimination method based on face and neck contour image features, characterized in that: The following judgment methods are included: Step S1: acquiring a face and neck contour image of the individual to be tested through an image acquisition module; Step S2: extracting eight facial and neck contour feature values ​​of OSA risk from the facial and neck contour image based on the feature extraction module; The eight facial and neck contour characteristic values ​​are: lower 1 / 3 height of face, chin-neck depth, neck depth, chin-neck angle, mandibular soft tissue width, neck width, chin-neck height and neck height; Step S3: Analyze the eight facial and neck contour feature values ​​through the risk assessment module and determine the risk level of OSA of the individual to be tested.

2. The OSA identification method based on face and neck contour image features according to claim 1, characterized in that: The feature extraction module comprises: Face recognition submodule: Use OpenCV tools to perform face recognition and target parameter extraction on the face and neck contour image; Feature point alignment submodule: detect facial feature points and align them to standard positions; Eigenvalue calculation submodule: Apply edge detection and skin color detection, combine edge information and skin color information to segment and measure the relevant parameters of facial parts, and obtain eight facial and neck contour eigenvalues.

3. The OSA identification method based on face and neck contour image features according to claim 2, characterized in that: The risk assessment module uses a multi-layer perceptron MLP model, and the MLP model includes an input layer, a hidden layer and an output layer; Input layer: used to receive eight facial and neck contour feature values; Hidden layer: used for nonlinear transformation and feature extraction of input feature values; Output layer: used to output the risk level of OSA of the individual to be tested; The risk assessment module includes: a characteristic value assessment submodule, a prediction submodule and an adjustment submodule.

4. The OSA identification method based on face and neck contour image features according to claim 3 is characterized in that: The calculation formula of the eigenvalue evaluation submodule is as follows: in: FFUA refers to the comprehensive evaluation index of characteristic values, Ha refers to the lower 1 / 3 height of the face, Wf refers to the facial width, Dc refers to the chin-neck depth, Hn refers to the neck height, Wj refers to the mandibular soft tissue width, Wn refers to the neck width, HC refers to the chin-neck height, Dn refers to the neck depth, KT refers to the chin-neck angle, A1, A2, A3, A4 and A5 refer to weight coefficient one, weight coefficient two, weight coefficient three, weight coefficient four and weight coefficient five respectively; The processing process of the eigenvalue evaluation submodule is as follows: the lower 1 / 3 height Ha, facial width Wf, chin-neck depth Dc, neck height Hn, mandibular soft tissue width Wj, neck width Wn, chin-neck height HC, neck depth Dn and chin-neck angle KT are input into the eigenvalue evaluation submodule, and the eigenvalue comprehensive evaluation index FFUA is output; The calculation formula of the prediction submodule is as follows: in: RVLU refers to the OSA risk prediction value, RAG refers to age, RGD refers to gender, the output is 1 for males and 0 for females, BMI refers to body mass index, B1, B2, B3, and B4 refer to model parameter one, model parameter two, model parameter three, and model parameter four, respectively; The processing process of the prediction submodule is as follows: based on the characteristic value comprehensive evaluation index FFUA, and inputting age RAG, gender RGD and body mass index BMI into the prediction submodule, outputting the OSA risk prediction value RVLU; The calculation formula of the adjustment submodule is as follows: AHIA=C1×RVLU+C2×(RVLU-RVLUa) 2 +V0; where: AHIA refers to the adjusted predicted value, RVLUa refers to the mean RVLU, which is used to adjust the central position of the square term, V0 refers to the constant term, which is used to adjust the overall offset of the predicted value, C1 and C2 refer to the first model parameter and the second model parameter, respectively; The processing process of the adjustment submodule is as follows: the OSA risk prediction value RVLU is input into the adjustment submodule, and based on the calculation of the OSA risk prediction value RVLU and the RVLU average RVLUa, the adjusted prediction value AHIA is output.

5. The OSA identification method based on face and neck contour image features according to claim 4, characterized in that: The face recognition submodule includes: an image reading unit, an image conversion to grayscale unit, a face detection unit and a detection result drawing unit.

6. The OSA identification method based on face and neck contour image features according to claim 5, characterized in that: The image reading unit is to use OpenCV to read the image file that needs to be detected, and load the image data from the storage medium into the memory for subsequent processing.

7. The OSA identification method based on face and neck contour image features according to claim 6, characterized in that: The image conversion to grayscale unit is to convert the image data into a grayscale image.

8. The OSA identification method based on face and neck contour image features according to claim 7, characterized in that: The face detection unit is to detect the position of the face in the grayscale image using the dlib face detector; the detection result drawing unit is to draw a rectangular frame on the image data and mark the detected face position.

Images