System and method for assessing risk of sleep disordered breathing in children based on cbct dual-channel analysis

Abstract

The application discloses a CBCT-based double-channel analysis child sleep breathing abnormality risk assessment system and method, utilizes CBCT to collect local three-dimensional image data of a maxillofacial part, carries out head position correction and anatomical landmark positioning, extracts upper airway sagittal area, volume and multi-dimensional angle and distance parameters, constructs a logistic regression clinical prediction model combining single-factor and multi-factor logistic regression to realize risk prediction, synchronously standardizes and equally-interval samples image data, generates a multi-channel three-dimensional image data cube, inputs a 3D ResNet network based on an identity shortcut connection to realize deep learning classification, and finally carries out weighted cooperative analysis on the results of two diagnostic channels and outputs child obstructive sleep apnea risk classification. Through the double-channel fusion design of parameter driving and image learning, the accuracy and applicability of early screening and risk assessment of child obstructive sleep apnea are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent medical monitoring and management, in particular to a CBCT-based dual-channel analysis system and method for evaluating the risk of sleep breathing abnormalities in children. BACKGROUND

[0002] Child obstructive sleep apnea (OSA) is a common sleep disorder characterized by partial or complete obstruction of the upper airway during sleep, affecting normal ventilation and sleep patterns. The etiology, clinical manifestations, and diagnostic criteria of children are almost independent standards and screening areas. Children have a high prevalence and significant, persistent harm, and there is a need for technical assistance to detect early, evaluate efficacy, and technical means.

[0003] The current gold standard for diagnosis is polysomnography (PSG), which requires processing and optimization of PSG signals while simplifying the detection equipment. However, there are significant limitations in the child population:

[0004] Difficulty in coordination: multiple sensors need to be connected, which can cause discomfort and resistance;

[0005] Unnatural environment: the hospital laboratory environment is different from daily sleep, which can cause a "first night effect";

[0006] Poor resources and accessibility: sleep centers are concentrated in large cities, equipment is scarce, and waiting time is long;

[0007] High cost and long time-consuming: expensive equipment, professional overnight monitoring, and complex analysis.

[0008] Existing PSG interpretation methods and existing patent diagnostic algorithms require a large amount of preprocessing by humans to be used for OSA evaluation modeling, and the technology is sensitive, focusing on adult OSA or PSG technology optimization, and lacking effective alternatives in the field of child OSA. Child OSA is different from adult OSA, with more common causes such as adenoid / tonsil hypertrophy, craniofacial structure abnormalities, and different diagnostic criteria (child AHI > 1 is abnormal).

[0009] Therefore, the present application proposes a CBCT-based dual-channel analysis system and method for evaluating the risk of sleep breathing abnormalities in children. SUMMARY

[0010] The present application aims to provide a CBCT-based dual-channel analysis system and method for evaluating the risk of sleep breathing abnormalities in children, which solves the problems of complex operation, high cost, and poor compliance in existing child obstructive sleep apnea (OSA) diagnosis relying on polysomnography.

[0011] In a first aspect, the present application provides a method for evaluating the risk of sleep breathing abnormalities in children based on CBCT dual-channel analysis, comprising the following steps:

[0012] S101: Correcting the head position of the CBCT original image data and performing three-dimensional positioning trace processing to obtain a set of jaw and facial image parameters, wherein the set of jaw and facial image parameters includes sagittal area, volume, and multi-dimensional angle parameters and multi-dimensional distance parameters of different sections of the upper airway;

[0013] S102: Based on the set of jaw and facial image parameters, a single-factor regression screening and a multi-factor logistic regression modeling are adopted to construct a logistic regression clinical prediction model for obstructive sleep apnea in children;

[0014] S103: The CBCT original image data is subjected to equal-interval sampling and pixel value standardization processing to generate an equal-dimension three-dimensional image data cube for deep learning;

[0015] S104: Using a three-dimensional convolutional neural network model with the three-dimensional image data cube as input, an image deep learning classification result for obstructive sleep apnea in children is obtained, wherein the three-dimensional convolutional neural network model is a 3D ResNet network with an identity shortcut connection structure;

[0016] S105: The output result of the logistic regression clinical prediction model and the image deep learning classification result of the three-dimensional convolutional neural network model are subjected to collaborative analysis to obtain a risk classification result for obstructive sleep apnea in children.

[0017] As a preferred technical solution of the present application, the step of obtaining the set of jaw and facial image parameters comprises:

[0018] Correcting the head position of the CBCT original image data based on standard anatomical landmark points;

[0019] According to the corrected CBCT original image data, the sagittal area and volume of the nasopharynx region, the palatopharynx region, the glossopharynx region, and the laryngopharynx region are extracted;

[0020] In a three-dimensional coordinate system, the multi-dimensional angle parameters and multi-dimensional distance parameters of the mandible, hyoid bone, and craniofacial skeleton are measured;

[0021] The above-mentioned area, volume, angle, and distance parameters are combined to form the set of jaw and facial image parameters.

[0022] As a preferred technical solution of the present application, the step of constructing the logistic regression clinical prediction model comprises:

[0023] Randomly dividing the set of jaw and facial image parameters into a training set and a validation set;

[0024] The single-factor regression method is used to screen parameters with P value less than or equal to 0.1 as modeling variables;

[0025] Based on the screened modeling variables, a multi-factor logistic regression modeling is performed to construct a risk prediction model of obstructive sleep apnea;

[0026] According to the model regression coefficient, a Nomogram score chart is drawn to realize individual risk scoring and prediction of children.

[0027] As a preferred technical solution of the present application, the step of generating an equal-dimension three-dimensional image data cube for deep learning includes:

[0028] The CBCT original image data is sampled at equal intervals to generate a three-dimensional data cube of a preset size;

[0029] According to the set window width and window level, the image pixel value is standardized to a preset gray value range in a preset pixel interval;

[0030] The standardized pixel matrix is combined into a three-dimensional image data cube in a multi-channel input form, and each channel of the three-dimensional image data cube corresponds to a standardized two-dimensional slice data in the axial, sagittal and coronal dimensions.

[0031] As a preferred technical solution of the present application, the step of obtaining the image deep learning classification result includes:

[0032] A 3D ResNet network based on an identity shortcut connection structure is built, and the 3D ResNet network includes a residual module and a global average pooling layer;

[0033] The three-dimensional image data cube and its corresponding obstructive sleep apnea label are input, the 3D ResNet network is trained, and the corresponding classification features are output after multi-layer convolution and residual learning;

[0034] The multi-classification probability result output by the 3D ResNet network generates an obstructive sleep apnea grading category, and the classification features and preset grading labels are used for supervised training to optimize the weight parameters of the 3D ResNet network;

[0035] The classification accuracy, precision and recall rate of the 3D ResNet network are evaluated based on the training set and the validation set.

[0036] As a preferred technical solution of the present application, the step of training the 3D ResNet network includes:

[0037] Construct two network architectures of 3D ResNet50 and 3D ResNet101 respectively, and adopt a cross-entropy loss function and an Adam optimizer to jointly control the training process of the 3D ResNet network;

[0038] The 3D ResNet50 and 3D ResNet101 are trained respectively by using the training set;

[0039] The classification performance of the two networks is compared on the validation set, and the network with the highest classification accuracy on the validation set is selected as the final three-dimensional convolutional neural network model.

[0040] As a preferred technical solution of the present application, the generation logic of the risk classification result is:

[0041] The risk prediction result of the logistic regression clinical prediction model is taken as a first risk factor, and the image deep learning classification result is taken as a second risk factor;

[0042] The first risk factor and the second risk factor are fused based on a weighted collaborative decision mechanism to generate the risk classification result of the child obstructive sleep apnea;

[0043] The weights of the first risk factor and the second risk factor are dynamically adjusted according to the classification accuracy of historical clinical samples.

[0044] In a second aspect, the present application provides a child sleep respiratory abnormality risk assessment system based on CBCT double-channel analysis, which is based on the implementation of the first aspect and includes a head position correction and image parameter extraction module, a logistic regression prediction modeling module, an image data cube generation module, an image deep learning analysis module, and a collaborative analysis module. Each module is connected through wired and / or wireless connection.

[0045] The head position correction and image parameter extraction module performs head position correction and three-dimensional positioning trace processing on the CBCT original image data to obtain a set of maxillofacial image parameters, which includes sagittal area, volume, and maxillofacial multi-dimensional angle parameters and multi-dimensional distance parameters for different segments of the upper airway.

[0046] The logistic regression prediction modeling module constructs a logistic regression clinical prediction model for child obstructive sleep apnea according to the set of maxillofacial image parameters, which is generated based on single-factor regression screening and multi-factor logistic regression modeling and is used to output the risk prediction result of child obstructive sleep apnea.

[0047] The image data cube generation module is used for equal-interval sampling and pixel value standardization processing of the CBCT original image data to generate an equal-dimension multi-channel three-dimensional image data cube.

[0048] An image deep learning analysis module uses a three-dimensional convolutional neural network model taking the three-dimensional image data cube as input to obtain an image deep learning classification result of obstructive sleep apnea in children, and the three-dimensional convolutional neural network model is a 3D ResNet network with an identity shortcut connection structure.

[0049] A synergistic analysis module synergistically analyzes the output result of the logistic regression clinical prediction model and the image deep learning classification result of the three-dimensional convolutional neural network model to obtain a risk classification result of obstructive sleep apnea in children.

[0050] In the above technical solution, the present application provides technical effects and advantages:

[0051] The present application realizes the complementary fusion of parameter driving and data driving through the dual-channel synergistic design of logistic regression clinical prediction and three-dimensional convolutional neural network image analysis, effectively improving the accuracy and reliability of the diagnosis of obstructive sleep apnea in children.

[0052] Based on the modeling of child-specific jaw and facial anatomical parameters, the physiological characteristics and individual differences of different age groups are taken into account to enhance the adaptability and generalizability to different child samples. The CBCT original image data is used in parallel in parameter extraction and image data cube generation, taking into account structural parameters and original image information, to improve data utilization efficiency and diagnostic value.

[0053] Relying on CBCT as the image input source, combining deep learning technology and regression modeling analysis, a non-invasive and standardized evaluation method is provided for early screening of obstructive sleep apnea in children. The synergistic analysis module integrates the classification results of the dual-channel model through a risk factor weighting mechanism to generate a unified risk classification output, reducing subjective errors and inconsistencies in manual interpretation. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0055] Figure 1 A step flowchart of the child sleep respiratory abnormality risk assessment method of the present application;

[0056] Figure 2 A clinical feature extraction point diagram of the child sleep respiratory abnormality risk assessment method of the present application;

[0057] Figure 3 An adenoid size evaluation diagram of the child sleep respiratory abnormality risk assessment method of the present application;

[0058] Figure 4 The upper airway obstruction degree quantitative diagram for the child sleep respiratory abnormality risk assessment method of the present application;

[0059] Figure 5 The nomogram diagram of the clinical prediction model of the present application;

[0060] Figure 6 The ROC curve diagram of the clinical prediction model of the present application;

[0061] Figure 7 The 3D Resnet architecture diagram of the present application;

[0062] Figure 8 The data set distribution diagram of the present application;

[0063] Figure 9 The ROC curve diagram of the optimal 3D Resnet prediction model of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described in more detail below in combination with the drawings in the embodiments of the present application.

[0065] In the drawings, the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The described embodiments are part of the embodiments of the present application, rather than all the embodiments. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application. The embodiments of the present application will be described in detail below in combination with the drawings.

[0066] Embodiment 1

[0067] Please refer to Figure 1 The present embodiment provides a child sleep respiratory abnormality risk assessment method based on CBCT double-channel analysis, which comprises the following steps:

[0068] S101: Correcting the head position of the CBCT original image data and performing three-dimensional positioning tracing processing to obtain a jaw and facial image parameter set, wherein the jaw and facial image parameter set comprises sagittal area, volume and jaw and facial multidimensional angle parameters and multidimensional distance parameters for different segments of the upper airway;

[0069] The CBCT original image data is CBCT original image data obtained by cone beam computer tomography (CBCT). The CBCT technology adopts a pulsed X-ray source and a detector array to obtain a tomographic image with a craniofacial region as a scanning range. The CBCT image can clearly show the structure boundary of the craniofacial hard tissue and the upper airway gap region, and generate equidistant three-dimensional data without structure overlap, which is beneficial to subsequent accurate upper airway parameter extraction and structure analysis. The CBCT overcomes the problem of traditional two-dimensional cephalometric measurement, and can clearly and non-magnifyingly display the craniofacial tissue without structure overlap, and the spatial relationship of the structure is not affected by the head position.

[0070] The CBCT uses pulsed exposure, and the single scanning dose is 30-120 μSv. The scanning range is locally focused (craniofacial region), which is different from the whole body / large range 200-2000 μSv dose of traditional multi-row CT, and conforms to the ALARA principle. Moreover, the exposure time only needs 5-20 seconds, almost all children can successfully cooperate and are non-invasive, such as Figure 2 The CBCT provides a key three-dimensional parameter acquisition basis for the logistic regression clinical model. The extracted clinical features such as angles, distances, areas, and volumes have a standardized basis, and solve the problem of high difficulty of standardization identification and parameter extraction of craniofacial and upper airway structures in CBCT images. Through a unified anatomical landmark positioning method, the subsequent parameter extraction has repeatability and clinical consistency.

[0071] Specifically, the step of acquiring the craniofacial image parameter set comprises:

[0072] Correcting the head position of the CBCT original image data based on the standard anatomical landmarks;

[0073] According to the corrected CBCT original image data, the sagittal area and volume of the nasopharynx region, the palatopharynx region, the glossopharynx region, and the laryngopharynx region are extracted;

[0074] In a three-dimensional coordinate system, multi-dimensional angle parameters and multi-dimensional distance parameters of the mandible, hyoid bone, and craniofacial skeleton are measured;

[0075] The above area, volume, angle, and distance parameters are combined to form the craniofacial image parameter set.

[0076] S102: Based on the craniofacial image parameter set, a single factor regression screening and a multi-factor logistic regression modeling are adopted to construct a logistic regression clinical prediction model for children with obstructive sleep apnea;

[0077] Specifically, as shown in Figure 3The schematic diagram of adenoid size evaluation supports the collection of parameters such as volume and area of different regions of the upper airway (especially the nasopharynx region) in the logistic regression model, strengthens the specific recognition ability of the model for adenoid hypertrophy, and solves the problems of difficulty in quantifying adenoid volume and large subjective judgment error in children with OSA. By visualizing and quantifying the size of the adenoid, its performance in different obstruction levels is determined. The steps for constructing the logistic regression clinical prediction model include:

[0078] Randomly divide the maxillofacial image parameter set into a training set and a validation set;

[0079] Single factor regression method is used to screen parameters with P value less than or equal to 0.1 as modeling variables;

[0080] Based on the screened modeling variables, a multi-factor logistic regression model is constructed to build a prediction model for obstructive sleep apnea risk;

[0081] According to the model regression coefficient, a Nomogram score chart is drawn to realize individual risk scoring and prediction of children.

[0082] S103: Equal-interval sampling and pixel value standardization processing are performed on the CBCT original image data to generate an equal-dimension three-dimensional image data cube for deep learning;

[0083] Specifically, the step of generating an equal-dimension three-dimensional image data cube for deep learning includes:

[0084] The CBCT original image data is sampled at equal intervals to generate a three-dimensional data cube of a predetermined size;

[0085] According to the set window width and window level, the image pixel value is standardized to the [0, 255] gray value range in the [-1000HU, 1000HU] interval;

[0086] The standardized pixel matrix is combined into a three-dimensional image data cube in the form of multi-channel input.

[0087] S104: Using a three-dimensional convolutional neural network model with the three-dimensional image data cube as input, an image deep learning classification result of children's obstructive sleep apnea is obtained, and the three-dimensional convolutional neural network model is a 3D ResNet network with an identity shortcut connection structure. 3D-ResNet is a deep learning architecture that combines three-dimensional convolutional neural network (3D CNN) and residual learning network (ResNet), designed specifically for processing time-space data such as medical image sequences.

[0088] Specifically, the step of obtaining the image deep learning classification result includes:

[0089] A 3D ResNet network based on an identity shortcut connection structure is built, and the 3D ResNet network comprises a residual module and a global average pooling layer.

[0090] The 3D ResNet network is trained by inputting the three-dimensional image data cube and the corresponding obstructive sleep apnea label, and outputs corresponding classification features after multi-layer convolution and residual learning.

[0091] The 3D ResNet network outputs a multi-class probability result to generate an obstructive sleep apnea classification, and the classification features and a preset classification label are used for supervised training to optimize the weight parameters of the 3D ResNet network.

[0092] The classification accuracy, precision and recall of the 3D ResNet network are evaluated based on a training set and a validation set.

[0093] The step of training the 3D ResNet network comprises:

[0094] Two network architectures, 3D ResNet50 and 3D ResNet101, are respectively constructed, and a cross-entropy loss function and an Adam optimizer are used to jointly control the training process of the 3D ResNet network.

[0095] The 3D ResNet50 and 3D ResNet101 are respectively trained using a training set.

[0096] The classification performance of the two networks is compared on a validation set, and the network with the highest classification accuracy on the validation set is selected as the final three-dimensional convolutional neural network model.

[0097] S105: The output result of the logistic regression clinical prediction model and the image deep learning classification result of the three-dimensional convolutional neural network model are analyzed cooperatively to obtain a risk classification result of child obstructive sleep apnea.

[0098] Specifically, as shown in the quantitative diagram of upper airway obstruction degree, Figure 4 As a reference for 3D data cube generation and anatomical structure space understanding, the model enhances the ability to identify different obstruction degrees and is used to determine the severity of OSA, solving the problem that traditional 2D images cannot accurately express the three-dimensional obstruction degree of the upper airway, and improving the intuitiveness and accuracy of the measurement of key indicators such as the minimum cross-sectional area and volume of the upper airway. The generation logic of the risk classification result is as follows:

[0099] The risk prediction result of the logistic regression clinical prediction model is used as a first risk factor, and the image deep learning classification result is used as a second risk factor.

[0100] fusing the first risk factor and the second risk factor based on a weighted collaborative decision mechanism to generate a risk classification result of the child obstructive sleep apnea;

[0101] wherein: the weights of the first risk factor and the second risk factor are dynamically adjusted according to the classification accuracy of historical clinical samples.

[0102] Exemplarily, the collaborative analysis step comprises:

[0103] when the logistic regression clinical prediction model and the three-dimensional convolutional neural network model both output negative results, the child is determined to be in a low risk category;

[0104] when the logistic regression clinical prediction model and the three-dimensional convolutional neural network model both output positive results, the child is determined to be in a high risk category;

[0105] when the output results of the two models are inconsistent, the risk classification result is generated by weighted fusion judgment according to the model confidence and the output weight.

[0106] The machine learning model centered on children shows good diagnostic performance for OSA, is optimized for children aged 6-12 years, and locks specific parameters for children OSA, which benefits children a lot from early diagnosis and treatment. Deep learning can quantify high-dimensional radiology phenotypes beyond human perception. The present embodiment constructs a non-invasive prediction model for the OSA clinical scenario, processes and runs high-dimensional data, and provides an AI tool for OSA diagnosis and prognosis evaluation, which is fast, convenient, comfortable, and highly relevant. The existing data provides new clinical value. Channel one clinical model is better at capturing static anatomical indicators, while image model can comprehensively find disease characteristics. The present embodiment first realizes the establishment of a clinical parameter model and an image deep learning model for children OSA, retains the anatomical parameter system familiar to doctors, integrates the high-dimensional feature extraction capability of deep learning, and guides the clinical decision path of the prediction result.

[0107] Embodiment 2

[0108] Based on embodiment 1, the present embodiment selects children aged 6-12 years as the target population, and after night polysomnography, OSA patients and non-OSA patients meeting the inclusion and exclusion criteria of the present embodiment include basic demographic data and CBCT original image data of the target population.

[0109] OSA diagnostic criteria: children undergo overnight sleep monitoring in the hospital, and the detection includes night continuous respiration, arterial oxygen saturation, electroencephalogram, electrocardiogram, heart rate and other indicators. According to the polysomnography results, sleep apnea hypopnea syndrome and its severity are diagnosed.

[0110] Children interpretation index: 1≤AHI<5 is mild OSA; 5≤AHI<10 is moderate, ≥10 is severe.

[0111] S102: Construct a logistic regression clinical prediction model for children with obstructive sleep apnea based on the jaw-facial image parameter set, and obtain it by single factor regression screening and multi-factor logistic regression modeling of CBCT original image data;

[0112] Channel one: logistic regression clinical prediction model based on high correlation features

[0113] Unique selection of clinical parameters: based on clinical highly correlated medical knowledge, a child-specific measurement system is established, image data of OSA patients are collected, head position correction is performed on CBCT data, and key feature points are extracted.

[0114] Model variables include: positioning trace, angle measurement, distance measurement on three-dimensional data; and area calculation, volume calculation for different sections of upper airway anatomical structure.

[0115] Upper airway feature points: total sagittal area of nasopharynx, horizontal sagittal area of hard palate, total volume of nasopharynx, total sagittal area of palatopharynx, horizontal sagittal area of soft palate tip, volume of palatopharynx, total sagittal area of glossopharynx, horizontal sagittal area of epiglottis, volume of glossopharynx, total sagittal area of laryngopharynx, horizontal sagittal area of hyoid bone, volume of laryngopharynx, total sagittal area of upper airway, minimum cross-sectional area of upper airway, total volume of upper airway.

[0116] Three-dimensional jaw-facial distance index (mm): SN-Anterior Cranial Base, S-Ba-Posterior Cranial Base, Pog-Na Perp R, Pog-Na Perp L, A-Na Perp R, A-Na Perp L, ANS-Me, S-Go R, S-Go L, Hy-MP R, Hy-MP L, Hy-PP, Hy-FH R, Hy-FH L, PNS-Sp, Tn-EpPp R, Tn-EpPp L, EpR-EpL, PpR-PpL, UR4-UL4, UR6-UL6.

[0117] Three-dimensional maxillofacial angle index (°): FMA R, FMA L, SN-GoGn R, SN-GoGn, Y-axis (SGn-FH)R, Y-axis (SGn-FH) L, NS-Co R, NS-Co L, Cranial Flexure (NS-Ba), Gonial Angle(Co-Go-Me) R, Gonial Angle (Co-Go-Me) L, Facial Angle (FH-NPog) R, Facial Angle (FH-NPog) L, AB to Facial Plane, Convexity (NA-APog).

[0118] Screening and construction of model variables All data were included in the dataset, and the dataset was randomly divided into training set and validation set according to the ratio of 7:3.

[0119] Based on R language, univariate regression was used to screen predictive variables, and predictive variables with P < 0.1 were screened for multivariate logistic regression to construct a prediction model for OSA in children. The clinical prediction model for OSA in children was fitted based on multivariate logistic regression, and the Nomogram of the prediction model was drawn, as shown in the figure below. Figure 5 As shown in the figure. Based on the contribution of each influencing factor to the outcome variable (the size of the regression coefficient), each value level of each influencing factor is assigned a score. The individual scores are then summed to obtain a total score. Finally, through the functional conversion relationship between the total score and the probability of the outcome event, the predicted value of the individual outcome event is calculated. This makes the prediction model results more readable and facilitates patient evaluation.

[0120] Model validation: draw the ROC curve, such as Figure 6 As shown in the figure, the area under the ROC curve (AUC) is calculated to obtain the C-statistic. When the C-statistic is greater than 0.5, it proves that the prediction model has predictive ability, and the closer the statistic is to 1, the better the discrimination of the model.

[0121] Channel 2: A three-dimensional convolutional neural network model for diagnosing and grading OSA in children based on deep learning. The three-dimensional convolutional neural network model includes three parts: algorithm selection, data processing, and model performance evaluation:

[0122] Analysis Algorithm Selection: CBCT data is a dense 3D voxel dataset. 3D Convolutional Neural Networks (CNNs) are an effective tool for processing this type of 3D data. 3D CNNs can automatically learn features from 3D medical images, eliminating the need for tedious manual feature design. They are also capable of effectively modeling complex 3D structures, making them widely used in tasks such as distinguishing normal from diseased tissue and performing regional segmentation.

[0123] In view of the above characteristics, the present embodiment selects a 3D ResNet network based on 3D CNN to classify and identify the risk severity of sleep breathing abnormalities in children with OSA (obstructive sleep apnea) symptoms from CBCT images. As shown in Figure 7 A simplified architecture diagram of the 3D ResNet network is shown.

[0124] The core idea of the 3D ResNet network compared to other CNN architectures is the introduction of an identity shortcut connection structure. This structure allows the network to directly skip one or more layers, successfully solving the problem of increasing training error rate due to the increasing depth of the network.

[0125] In order to meet the requirements of data privacy and the algorithm itself, the data set is processed, mainly in 3 steps:

[0126] Data cleaning and preprocessing: convert the original DICOM data to a 3D array that does not contain patient and hospital and equipment information.

[0127] Equal interval sampling: due to the high resolution of CBCT, in order to adapt the input resolution of the 3D ResNet network and reduce the GPU memory occupation during training, the original DICOM resolution is uniformly scaled and padded to a 96x96x96 cube through equal interval sampling.

[0128] It should be noted that: based on the difference of clinical parameters, the focus of the disease can be focused on the distance, angle, proportion between bone structures, and the size, proportion of soft tissue, air space, etc. observation points.

[0129] The CT value of human tissue ranges from -1000 to +1000, with a total of 2000 divisions, of which the lowest air CT value is about -1000 HU. Window width is the CT value range selected when displaying the image. Within this range, the tissue structure is divided into multiple levels (gray scale) according to its density from white to black. Window level is the anchor point of the window technique, which is the center of the window width and also the center of the gray scale display.

[0130] In order to improve the detail contrast of the head and face region, set the window width and window level to [-1000, 1000] to transform the CBCT pixel value to the range of [0, 255], and all pixels below -1000 are taken as 0, and all pixel values above 1000 are taken as 255.

[0131] Soft tissue window: window width 200~350, window level 30~50; bone window: window width 1000~1500, window level 250~500.

[0132] Dataset division: At present, 113 valid data for demonstration are used, and the AHI value is converted into three categories of asymptomatic, mild, and moderate according to the interpretation international standard of children's PSA, such as Figure 8 The distribution of the demonstration data set is shown, in which asymptomatic accounts for 30.1%, mild accounts for 56.6%, and moderate accounts for 13.3%. The leave-one-out method 4:1 is used for the allocation of the training set and the test set.

[0133] Model performance selection and evaluation: The basic residual block of the 3D ResNet network is composed of two 3x3 convolution layers, and the input is directly connected to the output through an identity mapping in the middle. This design idea enables the network to still retain the input information when it is deep, avoiding the problem of gradient disappearance.

[0134] ResNet of different depths: According to the task requirements and the limitation of computing resources, ResNets of different depths can be constructed. Considering the high complexity of 3D data, this embodiment compares the training effects of Resnet 50 and Resnet 101 models with deeper levels, and finds that the accuracy of the trained 3D Resnet50 network architecture is always between 80% and 85%, while the training effect of the 3D Resnet101 network architecture is better, and the accuracy can reach more than 90%. The best training result, the accuracy is 91.3%, the precision is 93.5%, the recall is 93.5%, and the F1 value is 93.5%. The AUC of the three classifications of asymptomatic, mild, and moderate is calculated respectively, which is 0.942, 0.944, and 1.0, respectively, such as Figure 9 The ROC curve graph of the optimal 3D Resnet prediction model is shown.

[0135] Decision system: Based on the clinical-image dual-channel verification system of children's OSA, the diagnostic prediction results of the disease are obtained respectively. According to the results of the model, the need for PSG examination can be reduced or delayed to a certain extent, and in the context of fast and practicality, the disease can be screened and compared before and after treatment.

[0136] When the prediction results of the dual-channel model are both no, and the image data is accurate and qualified, it is recommended to combine clinical signs, and PSG can be considered to be temporarily suspended;

[0137] When the prediction results of the dual-channel model are both yes, it is a high-risk object, and if deformity intervention measures are needed, portable sleep monitoring instruments can be selected for further judgment or PSG diagnosis;

[0138] When the clinical medical prediction model result is no and the image prediction model result is yes, it is recommended to check the image data quality and measurement positioning index;

[0139] When the clinical medical prediction model result is yes and the image prediction model is no, under the condition of ensuring the accuracy and quality of the image data, it is a low-risk population, and it is recommended to select further screening means as needed after combining clinical signs.

[0140] Embodiment 3

[0141] This embodiment does not detail part as shown in embodiment 1, this embodiment provides a child sleep breathing abnormality risk assessment system based on CBCT double channel analysis, including head position correction and image parameter extraction module, logistic regression prediction modeling module, image data cube generation module, image deep learning analysis module and collaborative analysis module, each module is connected through wired and / or wireless.

[0142] The head position correction and image parameter extraction module corrects the head position of the CBCT original image data and performs three-dimensional positioning tracking processing to obtain a set of maxillofacial image parameters, which includes sagittal area, volume for different segments of upper airway, and multi-dimensional angle parameters and multi-dimensional distance parameters of maxillofacial region;

[0143] The logistic regression prediction modeling module constructs a logistic regression clinical prediction model for child obstructive sleep apnea according to the set of maxillofacial image parameters, which is generated based on single factor regression screening and multi-factor logistic regression modeling, and is used to output the risk prediction result of child obstructive sleep apnea;

[0144] The image data cube generation module is used for equal-interval sampling and pixel value standardization processing of the CBCT original image data to generate an equal-dimension multi-channel three-dimensional image data cube;

[0145] The image deep learning analysis module uses a three-dimensional convolutional neural network model with the three-dimensional image data cube as input to obtain an image deep learning classification result of child obstructive sleep apnea, and the three-dimensional convolutional neural network model is a 3D ResNet network with an identical shortcut connection structure;

[0146] The collaborative analysis module performs collaborative analysis on the output result of the logistic regression clinical prediction model and the image deep learning classification result of the three-dimensional convolutional neural network model to obtain a risk classification result of child obstructive sleep apnea.

[0147] The step of obtaining the set of maxillofacial image parameters includes:

[0148] The CBCT original image data is corrected based on standard anatomical landmark points;

[0149] The sagittal area and volume of the nasopharynx region, the palatopharynx region, the glossopharynx region and the laryngopharynx region are extracted according to the corrected CBCT original image data.

[0150] Measuring multi-dimensional angle parameters and multi-dimensional distance parameters of the mandible, hyoid bone and craniofacial skeleton in a three-dimensional coordinate system;

[0151] Combining the above area, volume, angle and distance parameters to form a set of maxillofacial image parameters.

[0152] The step of constructing the logistic regression clinical prediction model comprises:

[0153] Randomly dividing the maxillofacial image parameter set into a training set and a validation set;

[0154] Using a single-factor regression method to screen parameters with a P value less than or equal to 0.1 as modeling variables;

[0155] Based on the screened modeling variables, a multi-factor logistic regression modeling is performed to construct a risk prediction model for obstructive sleep apnea;

[0156] According to the model regression coefficient, a Nomogram score chart is drawn to realize individual risk scoring and prediction of children.

[0157] The step of generating an equal-dimension three-dimensional image data cube for deep learning comprises:

[0158] The CBCT original image data is sampled at equal intervals to generate a three-dimensional data cube of a preset size;

[0159] According to the set window width and window level, the image pixel value is standardized to a preset gray value range in the preset pixel interval;

[0160] The standardized pixel matrix is combined into a three-dimensional image data cube in a multi-channel input form, and each channel of the three-dimensional image data cube corresponds to a standardized two-dimensional slice data in the axial, sagittal and coronal dimensions.

[0161] The step of obtaining the image deep learning classification result comprises:

[0162] A 3D ResNet network based on an identity shortcut connection structure is built, which includes a residual module and a global average pooling layer;

[0163] The three-dimensional image data cube and its corresponding obstructive sleep apnea label are inputted, the 3D ResNet network is trained, and the corresponding classification features are outputted after multi-layer convolution and residual learning;

[0164] The multi-class probability result outputted by the 3D ResNet network generates an obstructive sleep apnea classification category, and the classification features and the preset classification label are supervised and trained to optimize the weight parameters of the 3D ResNet network;

[0165] Based on the training set and the validation set, the classification accuracy, precision and recall of the 3D ResNet network are evaluated.

[0166] The step of training the 3D ResNet network comprises:

[0167] Two network architectures of 3D ResNet50 and 3D ResNet101 are respectively constructed, and the cross-entropy loss function and the Adam optimizer are used to jointly control the training process of the 3D ResNet network.

[0168] The 3D ResNet50 and 3D ResNet101 are trained respectively using the training set;

[0169] The classification performance of the two networks is compared on the validation set, and the network with the highest classification accuracy on the validation set is selected as the final three-dimensional convolutional neural network model.

[0170] The generation logic of the risk classification result is:

[0171] The risk prediction result of the logistic regression clinical prediction model is taken as the first risk factor, and the image deep learning classification result is taken as the second risk factor.

[0172] Based on the weighted collaborative decision mechanism, the first risk factor and the second risk factor are fused to generate the risk classification result of the child obstructive sleep apnea.

[0173] Wherein: the weights of the first risk factor and the second risk factor are dynamically adjusted according to the classification accuracy of the historical clinical samples.

[0174] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for assessing the risk of sleep disordered breathing in children based on CBCT dual-channel analysis, characterized in that, The method comprises the following steps: S101: Correcting the head position of CBCT original image data and performing three-dimensional positioning tracking processing to obtain a maxillofacial image parameter set, wherein the maxillofacial image parameter set comprises sagittal area, volume of different sections of upper airway, and multi-dimensional angle parameters and multi-dimensional distance parameters of the maxillofacial region; S102: Based on the maxillofacial image parameter set, a single factor regression is used to screen and a multi-factor logistic regression modeling is used to construct a logistic regression clinical prediction model for children with obstructive sleep apnea; The step of constructing the logistic regression clinical prediction model comprises: randomly dividing the maxillofacial image parameter set into a training set and a validation set; using a single factor regression method to screen parameters with a P value less than or equal to 0.1 as modeling variables; based on the screened modeling variables, performing multi-factor logistic regression modeling to construct an obstructive sleep apnea risk prediction model; drawing a Nomogram score chart according to the model regression coefficients to realize individual risk scoring and prediction of children; S103: performing equal-interval sampling and pixel value standardization processing on the CBCT original image data to generate an equal-dimension three-dimensional image data cube for deep learning; The step of generating an equal-dimension three-dimensional image data cube for deep learning comprises: performing equal-interval sampling on the CBCT original image data to generate a three-dimensional data cube with a preset size; standardizing the image pixel values to a preset gray value range in a preset pixel interval according to the set window width and window level; combining the standardized pixel matrix into a three-dimensional image data cube in a multi-channel input form, wherein each channel of the three-dimensional image data cube corresponds to a standardized two-dimensional slice data in the axial, sagittal and coronal dimensions; S104: using a three-dimensional convolutional neural network model with the three-dimensional image data cube as input to obtain an image deep learning classification result of children with obstructive sleep apnea, wherein the three-dimensional convolutional neural network model is a 3D ResNet network with an identity shortcut connection structure; S105: performing collaborative analysis on the output result of the logistic regression clinical prediction model and the image deep learning classification result of the three-dimensional convolutional neural network model to obtain a risk classification result of children with obstructive sleep apnea.

2. The CBCT-based dual-channel analysis method for assessing the risk of sleep disordered breathing in children according to claim 1, wherein, The step of obtaining the maxillofacial image parameter set comprises: correcting the head position of the CBCT original image data based on standard anatomical landmark points; extracting the sagittal area and volume of different sections of the upper airway according to the corrected CBCT original image data, wherein the sections of the upper airway include the nasopharynx region, the palatopharynx region, the glossopharynx region and the laryngopharynx region; measuring multi-dimensional angle parameters and multi-dimensional distance parameters of the maxillofacial region in a three-dimensional coordinate system, wherein the maxillofacial region includes the mandible, the hyoid bone and the craniofacial skeleton; combining the sagittal area, volume of different sections of the upper airway, and multi-dimensional angle parameters and multi-dimensional distance parameters of the maxillofacial region to form the maxillofacial image parameter set.

3. The CBCT-based dual-channel analysis method for assessing the risk of sleep disordered breathing in children according to claim 1, wherein, The step of obtaining the image deep learning classification result comprises: building a 3D ResNet network based on an identity shortcut connection structure, wherein the 3D ResNet network comprises a residual module and a global average pooling layer; Input the three-dimensional image data cube and its corresponding obstructive sleep apnea label, train the 3DResNet network, and output corresponding classification features after multi-layer convolution and residual learning; Generate obstructive sleep apnea classification categories through the multi-class probability results output by the 3D ResNet network, and supervise the training based on the classification features and preset classification labels to optimize the weight parameters of the 3D ResNet network; Evaluate the classification accuracy, precision, and recall of the 3D ResNet network based on the training set and validation set.

4. The CBCT-based dual-channel analysis method for assessing the risk of sleep disordered breathing in children according to claim 3, characterized in that, The steps of training the 3D ResNet network include: Respectively construct 3D ResNet50 and 3D ResNet101 network architectures, and jointly control the training process of the 3D ResNet network using cross-entropy loss function and Adam optimizer; Train the 3D ResNet50 and 3D ResNet101 using the training set; Compare the classification performance of the two networks on the validation set, and select the network with the highest classification accuracy on the validation set as the final three-dimensional convolutional neural network model.

5. The CBCT-based dual-channel analysis method for assessing the risk of sleep disordered breathing in children according to claim 1, wherein, The generation logic of the risk classification result is as follows: Take the risk prediction result of the logistic regression clinical prediction model as the first risk factor, and take the image deep learning classification result as the second risk factor; Fuse the first risk factor and the second risk factor based on a weighted collaborative decision mechanism to generate the risk classification result of child obstructive sleep apnea; Wherein: the weights of the first risk factor and the second risk factor are dynamically adjusted according to the classification accuracy of historical clinical samples.

6. A CBCT-based dual-channel analysis system for assessing the risk of sleep disordered breathing in children, implementation of the CBCT-based dual-channel analysis method for assessing the risk of sleep disordered breathing in children according to any one of claims 1 to 5, characterized in that, It includes a head correction and image parameter extraction module, a logistic regression prediction modeling module, an image data cube generation module, an image deep learning analysis module, and a collaborative analysis module, and each module is connected through wired and / or wireless connection; The head correction and image parameter extraction module performs head correction and three-dimensional positioning trace processing on the CBCT original image data to obtain a set of maxillofacial image parameters, which includes sagittal area, volume, and maxillofacial multi-dimensional angle parameters and multi-dimensional distance parameters for different segments of the upper airway; The logistic regression prediction modeling module constructs a logistic regression clinical prediction model for child obstructive sleep apnea based on the set of maxillofacial image parameters, which is generated based on single-factor regression screening and multi-factor logistic regression modeling, and is used to output child obstructive sleep apnea risk prediction results; The image data cube generation module is used for equal-interval sampling and pixel value standardization processing of the CBCT original image data to generate an equal-dimension multi-channel three-dimensional image data cube; The image deep learning analysis module uses a three-dimensional convolutional neural network model with the three-dimensional image data cube as input to obtain image deep learning classification results of child obstructive sleep apnea, and the three-dimensional convolutional neural network model is a 3D ResNet network with an identity shortcut connection structure. A synergistic analysis module is configured to synergistically analyze an output result of the logistic regression clinical prediction model and an image deep learning classification result of the three-dimensional convolutional neural network model to obtain a risk classification result of the child obstructive sleep apnea.

Images