Abnormality identification method and device for pulmonary ventilation function examination, electronic equipment and computer storage medium

Abstract

The invention provides an anomaly identification method and device for pulmonary ventilation function examination, electronic equipment and a computer storage medium. The abnormal recognition method for pulmonary ventilation function examination comprises the following steps: S1, acquiring a pulmonary ventilation function examination curve to be recognized, and processing the pulmonary ventilation function examination curve by adopting an adaptive Fourier (AFD) method; and S2, inputting the lung ventilation function examination curve obtained after processing into a trained YOLO algorithm model, wherein the YOLO algorithm model identifies abnormal conditions of the lung ventilation function examination curve. According to the abnormity identification method, the AFD technology and the deep learning technology are combined, the abnormal condition of pulmonary ventilation function examination can be efficiently and accurately identified, the reliability and consistency of examination data are ensured, medical personnel can be assisted in completing examination quality control, and the examination efficiency is improved. And the problem of insufficient training and practical experience of medical personnel in some regions and some medical institutions is effectively relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of physiological signal analysis, and more specifically, to an abnormality recognition method, device, electronic equipment and computer storage medium for pulmonary ventilation function examination. Background Art

[0002] Chronic respiratory diseases are a major public health challenge facing China, adversely affecting the quality of life and health of millions of people. The "Healthy China 2030" initiative emphasizes the key role of early diagnosis, and in particular, the key role of pulmonary ventilation function testing in this endeavor.

[0003] Pulmonary function test is an important diagnostic tool for evaluating respiratory diseases. It diagnoses and monitors various respiratory diseases such as asthma and chronic obstructive pulmonary disease (COPD) by measuring lung capacity and airflow rate. The flow-volume curve (FV curve for short) and the volume-time curve (VT curve for short) are both important tools for evaluating pulmonary ventilation function. The flow-volume curve mainly reflects the changes in flow and volume, and is suitable for evaluating airway obstruction and lung volume problems; the volume-time curve focuses more on changes in time, and is suitable for evaluating airway obstruction and restrictive lung disease. The analysis of these two curves can provide important reference for clinical diagnosis and treatment.

[0004] The data collected by the pulmonary ventilation function test is affected by many factors, and the obtained test data and the corresponding curve shape may show various abnormal conditions, such as coughing, glottis closure, interface blockage, air leakage, etc. These abnormal conditions may affect the acceptability and usability of the pulmonary ventilation function test. Therefore, professional medical personnel are needed to identify and quality control the abnormal conditions of the pulmonary ventilation function test curve.

[0005] In this regard, effectively trained medical personnel are a necessary condition for completing high-quality examinations. Studies have shown that the quality of pulmonary ventilation function examinations is closely related to the training of medical personnel. In my country, there are about 100 million patients with chronic obstructive pulmonary disease, but only 12% of patients have undergone pulmonary ventilation function tests, which may be related to the lack of effectively trained medical personnel. At present, my country has issued relevant policies requiring 50% of medical institutions to be equipped with spirometers, and the spirometer software can output FV curves and VT curves in addition to indicators; at the same time, medical institutions equipped with spirometers should use spirometers at a rate of 90%. For these policies to be effectively implemented, a large number of effectively trained and practiced medical personnel are indispensable, but there is an imbalance in the development and capabilities of different regions and medical institutions in my country. In order to overcome the problem of insufficient training of medical personnel in some regions and some medical institutions, strong support tools are needed to assist in completing inspection quality control.

[0006] Therefore, it is of great significance to provide a method that can assist medical personnel in efficiently and accurately identifying abnormalities in pulmonary ventilation function tests. Summary of the invention

[0007] The purpose of the present invention is to provide a method for identifying abnormalities in pulmonary ventilation function tests. The method for identifying abnormalities in pulmonary ventilation function tests combines AFD technology and deep learning technology, and can efficiently and accurately identify abnormalities in pulmonary ventilation function tests, ensure the reliability and consistency of test data, and provide strong support for improving the accuracy of clinical diagnosis.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] In a first aspect, the present invention provides a method for identifying abnormalities in a pulmonary ventilation function test, comprising the following steps:

[0010] S1. Obtaining a pulmonary ventilation function test curve to be identified, and processing the pulmonary ventilation function test curve using an adaptive Fourier transform (AFD) method;

[0011] S2. Input the processed pulmonary ventilation function test curve into the trained YOLO algorithm model, and the YOLO algorithm model identifies the abnormality of the pulmonary ventilation function test curve.

[0012] Optionally, the pulmonary ventilation function test curves of the present invention include but are not limited to flow-volume (FV) curves and VT curves. Usually, the pulmonary function test software in the pulmonary function test equipment outputs the corresponding data as well as the FV curve and the VT curve.

[0013] The abnormal conditions of the pulmonary ventilation function test curve include but are not limited to coughing, glottis closure, port blockage, and air leakage. Depending on the different abnormal conditions, the pulmonary ventilation function test curve will have different abnormal shapes.

[0014] As one embodiment, the identification includes marking the abnormal morphological part of the pulmonary ventilation function test curve. Furthermore, the marking result is presented by visual means.

[0015] Optionally, the present invention uses the adaptive Fourier (AFD) method to process the pulmonary ventilation function test curve, including: performing n-order adaptive Fourier decomposition on the pulmonary ventilation function test curve, and obtaining multiple frequency components through step-by-step decomposition, wherein the frequency components include frequency components containing key information and noise frequency components; removing the noise frequency components, and reconstructing the pulmonary ventilation function test curve according to the frequency components containing key information; p≥2. Optionally, p is 2 to 5. Specifically, p is 2, 3, 4 or 5. Only the frequency components obtained by adaptive Fourier decomposition of order 2 or above can effectively present the key features.

[0016] Optionally, the multiple frequency components of the present invention are arranged in order from low to high frequency, and the first x low-frequency frequency components are selected from the multiple frequency components as the frequency components containing key information; x is 3 to 6; specifically, x is 3, 4, 5 or 6. Preferably, x is 4 or 5. The higher the frequency component, the more noise it contains, and the high-frequency frequency component is generally regarded as the noise frequency component.

[0017] The frequency components of the first x low frequencies can effectively retain the key features of the abnormal situation during the examination. Re-superimposing these decomposed frequency components, that is, re-superimposing the frequency components of the first m low frequencies, reconstructing the pulmonary ventilation function test curve can significantly improve the smoothness of the pulmonary ventilation function test curve and reduce noise interference. In addition, not only the key features of the original pulmonary ventilation function test curve are successfully retained, but also the problem of overfitting is effectively avoided.

[0018] The frequency component of the present invention can be represented by the following function:

[0019]

[0020] in,

[0021] p represents the order of decomposition; q represents the order of frequency components; B n (e jt ) represents the orthogonal basis function of the nth order in the AFD decomposition; k represents the number of frequency components generated in the pth order decomposition; n p represents the total number of decompositions performed at order p. Optionally, k is 4 to 8, specifically 4, 5, 6, 7, 8; n p =k×p.

[0022] Optionally, the YOLO algorithm model described in the present invention is an improved YOLOv8 algorithm model; the improved YOLOv8 algorithm model is mainly composed of a backbone network, a neck network and a head network, the backbone network uses C2f-EMSCP to replace the C2f module of the original YOLOv8 algorithm model; the aggregation and distribution mechanism (Gather-Distribute Mechanism) is used to replace the neck network of the original YOLOv8 algorithm model; the dynamic target detection head (Dynamic Head) is used to replace the head network of the original YOLOv8 algorithm model; the input end of the backbone network is used to introduce the pulmonary ventilation function test curve obtained after AFD processing.

[0023] In step S1, before performing AFD processing, the pulmonary ventilation function test curve to be identified is also preprocessed; the preprocessing includes processing the data of the pulmonary ventilation function test curve using threshold setting and interquartile range (IQR) method, and then processing it using signal segmentation and cubic spline interpolation method. The threshold setting and interquartile range method are used to eliminate noise, and the signal segmentation and cubic spline interpolation method are used to optimize the data continuity and smoothness of the pulmonary ventilation function test curve.

[0024] Furthermore, the method also includes processing the data of the pulmonary ventilation function test curve using a median filtering technique.

[0025] The training method of the YOLO algorithm model of the present invention can refer to the prior art. As one embodiment, the training method of the YOLO algorithm model of the present invention includes the following steps: performing the preprocessing and AFD processing on the original pulmonary ventilation function test curve, and then annotating the abnormal conditions of the pulmonary ventilation function test curve before and after the processing to obtain an image set; dividing the image set into a training set and a validation set in proportion, and using the training set and the validation set to train and evaluate the YOLO algorithm model.

[0026] Optionally, the method of obtaining the pulmonary ventilation function test curve to be identified in the present invention includes: directly using pulmonary ventilation function test numerical data to generate the pulmonary ventilation function test curve, or extracting it from an image recording the pulmonary ventilation function test curve.

[0027] Whether directly using the pulmonary ventilation function test numerical data to generate the pulmonary ventilation function test curve, or extracting it from an image that records the pulmonary ventilation function test curve, the purpose is to allow the algorithm model to obtain a recognizable pulmonary ventilation function test curve.

[0028] By adopting the method of extracting from the image, it is possible to get rid of the limitation that the pulmonary function meter and the device for identifying abnormal pulmonary ventilation function tests are fixedly connected or integrated together. Various terminals can be used to upload the image of the pulmonary ventilation function test curve to the device or electronic device for identifying abnormal pulmonary ventilation function tests (referred to as the identification device), thereby realizing the separate deployment of the pulmonary function meter and the identification device, and the cooperation between primary medical institutions and medical institutions capable of deploying the identification device, thereby reducing the cost and maintenance pressure of the primary medical institutions.

[0029] Optionally, the method for extracting from an image recording a pulmonary ventilation function test curve is mainly implemented based on the OpenCV library, including template matching, Canny edge detection and Hough line transformation methods, which can ensure the accuracy and reliability of the data.

[0030] In a second aspect, another object of the present invention is to provide a device for identifying abnormal conditions in a pulmonary ventilation function test. The device for identifying abnormal conditions in a pulmonary ventilation function test comprises:

[0031] An input unit, used for obtaining a pulmonary ventilation function test curve to be identified;

[0032] A processing unit, used for processing the pulmonary ventilation function test curve to be identified, wherein the processing is AFD processing, or preprocessing and AFD processing;

[0033] The detection unit is used to identify abnormal conditions of the processed pulmonary ventilation function test curve according to the YOLO algorithm model.

[0034] In a third aspect, the present invention further provides an electronic device. The electronic device comprises at least one memory and at least one processor; the at least one memory is coupled to the at least one processor, the at least one memory is used to store a computer program, the at least one processor is used to call the computer program, the computer program comprises instructions, and when the instructions are executed by the at least one processor, the electronic device executes the abnormality identification method for pulmonary ventilation function test described in the first aspect.

[0035] In a fourth aspect, the present invention further provides a computer storage medium, wherein the computer storage medium includes computer instructions, and when the computer instructions are executed on an electronic device, the electronic device executes the abnormality identification method for pulmonary ventilation function test described in the first aspect.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The abnormality identification method for pulmonary ventilation function test described in the present invention combines AFD technology and deep learning technology, can efficiently and accurately identify abnormalities in pulmonary ventilation function test, ensure the reliability and consistency of test data, and provide strong support for improving the accuracy of the next step of clinical diagnosis; at the same time, it can assist medical personnel in completing examination quality control and effectively alleviate the problem of insufficient training and practical experience of medical personnel in some regions and some medical institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the abnormality identification method for pulmonary ventilation function test in Example 1.

[0039] Figure 2 Schematic diagram of the improved YOLOv8 algorithm model.

[0040] Figure 3 Schematic diagram of the EMSCP network architecture and multi-scale feature fusion.

[0041] Figure 4 This is a processing diagram of the low-stage aggregation and distribution mechanism.

[0042] Figure 5 It is a processing diagram of the high-level aggregation and distribution mechanism.

[0043] Figure 6 This is a processing diagram of the information distribution (Inject) module.

[0044] Figure 7 Schematic diagram and application of the DyHead module: (a) Schematic diagram of the DyHead module; (b) Detailed view of the Dyhead module; (c) Application in a primary detector.

[0045] Figure 8 This is a flow chart of the abnormality identification method for pulmonary ventilation function test in Example 2.

[0046] Fig. 9 is a template image, where (a) locates the horizontal axis and (b) locates the vertical axis.

[0047] Fig.10 This is a schematic diagram of the data distribution after being processed by the interquartile range method.

[0048] Fig.11 This is a schematic diagram of data distribution after signal segmentation and cubic spline interpolation processing.

[0049] Fig.12 The present invention is a functional module block diagram of a device for identifying abnormal conditions in a pulmonary ventilation function test.

[0050] Fig.13 This is a specific structural block diagram of an electronic device.

[0051] Fig.14 Schematic diagram of image data with anomalies marked in the dataset.

[0052] Fig.15 The data diagrams are for 1st to 3rd order AFD processing, where from left to right they are (a) 1st order, (b) 2nd order, and (c) 3rd order.

[0053] Fig.16 The figure shows the comparison between AFD and the other two methods (EMD and VMD), where from left to right they are (a) EMD, (b) VMD, and (c) third-order AFD.

[0054] Fig.17 2D plots for t-SNE visualization, (a) original YOLO; (b) YOLO-EMD; (c) YOLO-VMD; (d) YOLO-AFD.

[0055] Fig.18Grad-CAM visualization heatmap for FV curve anomaly detection. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] Example 1

[0058] like Figure 1 As shown, a method for identifying abnormalities in a pulmonary ventilation function test comprises the following steps:

[0059] S1. Acquire a pulmonary ventilation function test curve to be identified; process the pulmonary ventilation function test curve using an adaptive Fourier transform (AFD) method;

[0060] S2. Input the processed pulmonary ventilation function test curve into the trained YOLO algorithm model, and the YOLO algorithm model identifies the abnormality of the pulmonary ventilation function test curve.

[0061] About AFD Processing

[0062] Mathematical basis of AFD

[0063] AFD is an adaptive decomposition algorithm. AFD uses the matching pursuit algorithm to construct an adaptive basis function system. Each basis function B n (e jt ) is composed of modified Blaschke products. The modified Blaschke product definition formula of each basis function is as follows:

[0064]

[0065] Each shape parameter a n Constrained to the unit disk This constraint ensures that the adaptive process remains stable and the function remains bounded in the analytical domain. n The unit disk The complex coefficients in are adaptive parameters chosen to maximize energy convergence at each decomposition level. For any sequence of a in the unit disk D n are orthogonal. The processed signal G(e jt ) can be expressed as a series expansion form of adaptive basis functions.

[0066] SignalG(e jt ) through these basis functions B n (e jt ) is broken down as follows:

[0067]

[0068] Each coefficient A n Through the signal G and basis function B n The inner product of <G,B n > OK. Subsequently, each level of reduction of residual G n Calculated as:

[0069]

[0070] A systematic approach to decompose the signal into organized and integrated components is demonstrated. The successive reduction of the residuals highlights the effectiveness of the method in extracting important signal features, reducing noise, and enhancing signal interpretability.

[0071] n-order AFD decomposition

[0072] Using the above AFD principle, in order to ensure that the signal is completely decomposed, the p-order AFD decomposition method adopted in the present invention establishes multiple levels of decomposition results, and can significantly reduce the errors caused by redundant data by strategically integrating some decomposition results. The hierarchical decomposition process is formalized by p-order AFD decomposition, and the aggregation process of any given order p is summarized. This method effectively extracts the hierarchical information of the signal from the macroscopic to the microscopic level. The decomposition and aggregation process in the p-order AFD is expressed as follows:

[0073]

[0074] where m(q,p) and M(q,p) are functions that determine the original decomposition range of the contribution to the q-th level combination component at order p. Depending on the specific aggregation scheme chosen, these functions are defined as follows:

[0075]

[0076] The parameters p and q determine the depth of the iterative selection process and the number of basis functions combined at each layer to ensure the best signal approximation and reduce the reconstruction error. p , and n p The order p increases. Adjusting the parameters p and q allows transformations to be performed at different scales and locations, accurately capturing the details and local features of the signal. This approach selects and optimizes the basis function at each stage to maximize the capture of key information while removing noise and irrelevant components.

[0077] In the above formula, B n (e jt) represents the orthogonal basis functions of the nth level in the AFD decomposition, usually expressed in the form of modified Blaschke products.

[0078] a n Determine the basis function B n (e jt ) characteristic parameters, constrained on the unit disk Inside.

[0079] represents the set of all complex numbers z with an absolute value less than 1, as a n The allowed parameter space.

[0080] G(t) represents the 2 Continuous analytic signal in space, used for decomposition.

[0081] G n (e jt ) represents the residue after the nth level decomposition, which is obtained by jt ) minus the contribution of the previous basis functions.

[0082] e{a n}(e jt ) means in a n The evaluation function at G n Projection to B n It is the key to calculate the coefficient.

[0083] A n Represents the coefficient of the nth level, given by the inner product <G,B n >Get, quantify B n Contribution when representing G.

[0084] <G,B n > indicates the coefficient A used to calculate n The inner product of n The representation intensity in .

[0085] p represents the order of decomposition and defines the depth of hierarchical analysis in the AFD process.

[0086] q represents the qth frequency component under this order AFD decomposition, that is, the order of the frequency component.

[0087] n p Indicates the total number of decompositions performed at order p, affecting the granularity of the analysis. That is, the fineness of the decomposition. Higher n p Higher values ​​usually mean a finer resolution, thus capturing more details of the signal.

[0088] k represents the number of output components or output graphs generated at each decomposition order p, that is, the number of signals or modes generated in the p-order decomposition. It determines how many different signal components the decomposition result will be divided into at this order (that is, how many frequency components described in this application the decomposition result will be divided into).

[0089] Represents the output component indexed by q at the decomposition order p, that is, the corresponding frequency component.

[0090] Applying the above method and principle, the present invention uses the adaptive Fourier (AFD) method to process the FV curve, including: performing p-order adaptive Fourier decomposition on the FV curve, and obtaining multiple frequency components through step-by-step decomposition, wherein the frequency components include frequency components containing key information and noise frequency components; removing the noise frequency components, and reconstructing the FV curve according to the frequency components containing key information; p≥2. Optionally, p is 2 to 5. Specifically, p is 2, 3, 4 or 5.

[0091] The multiple frequency components are arranged in order from low to high frequency, and the first x low-frequency frequency components are selected from the multiple frequency components as the frequency components containing key information; x is 3 to 6; specifically, x is 3, 4, 5 or 6. Preferably, x is 4 or 5.

[0092] About YOLO algorithm model recognition

[0093] The YOLO algorithm model is an improved YOLOv8 algorithm model; Figure 2 As shown, the improved YOLOv8 algorithm model is mainly composed of a backbone network, a neck network and a head network. The backbone network uses C2f-EMSCP to replace the C2f module of the original YOLOv8 algorithm model; the aggregation and distribution mechanism (Gather-Distribute Mechanism) is used to replace the neck network of the original YOLOv8 algorithm model; the dynamic target detection head (Dynamic Head, DyHead) is used to replace the head network of the original YOLOv8 algorithm model; the input end of the backbone network is used to introduce the pulmonary ventilation function test curve obtained after AFD processing.

[0094] The main structure of the original YOLOv8 algorithm model consists of a backbone network, a neck network, and a head network. The backbone network is responsible for feature extraction and uses the C2f module as the main building block. The C2f module has fewer parameters and higher feature extraction capabilities. The backbone network also uses depthwise separable convolution and dilated convolution to further enhance the effect of feature extraction.

[0095] The neck network of the original YOLOv8 algorithm model includes the SPPF module (fast spatial pyramid pooling) for multi-scale pooling operations; the PAA module (probabilistic anchor point allocation) for optimizing the allocation of positive and negative samples; and the PAN module (path aggregation network) for path aggregation of features at different levels to improve the richness of feature expression.

[0096] The head network of the original YOLOv8 algorithm model is responsible for target detection and classification, including the detection head and the classification head. The detection head uses a series of convolutional layers and deconvolutional layers to generate bounding box and confidence predictions; the classification head uses global average pooling to classify feature maps and output the probability distribution of each category, thereby supporting multi-category classification tasks.

[0097] In the improved YOLOv8 algorithm model of the present invention, the Efficient Multi-Scale Convolution with Position-aware Module (EMSCP) is a multi-scale convolution lightweight network architecture with a position-aware module. Figure 3 As shown in the figure, EMSCP uses multi-scale convolution operations to capture different features in the image, and merges feature maps at different scales through cascade operations to obtain richer feature representations. In addition, EMSCP divides the input features into four groups by channel through a grouping strategy, and each group is processed using convolution kernels of different sizes (1x1, 3x3, 5x5, and 7x7). Then, the inter-channel features are fused through point-to-point convolution, which improves the perception and expression of the model while maintaining computational efficiency and lightweight model.

[0098] The design of EMSCP significantly reduces the number of model parameters and computational overhead, and enhances the generalization ability of the model, making it suitable for efficient real-time target detection on resource-constrained devices, especially for embedded devices and mobile applications. The feature map segmentation strategy and the method of selectively processing part of the feature map significantly reduce the parameter scale of the network and optimize the model structure. In addition, by replacing the backbone network in the network and the C2f module in the neck network, EMSCP shows excellent performance in processing complex pulmonary function test data, which not only reflects the innovation of this study in improving the model performance, but also provides a new technical approach for the efficient processing and analysis of pulmonary function test data.

[0099] The aggregation and distribution mechanism described in the present invention is mainly composed of a feature alignment module (Feature Alignment Module, FAM), an information fusion module (Information Fusion Module, IFM) and an information distribution module (Information Injection Module, Inject).

[0100] The aggregation and distribution mechanism of the present invention solves the problem that the original YOLOv8's neck adopts the path aggregation network (PAN) in processing multi-level feature fusion, which has inefficiencies in efficiency and integrity. For example, the feature fusion from the first level (level-1) to the third level (level-3) must be transferred through the second level (level-2). This indirect fusion method may cause partial loss of information during the transmission process, thereby limiting the maximization of the model fusion effect.

[0101] FAM obtains feature maps of different scales from the backbone network and adjusts the size through upsampling or downsampling techniques to ensure the consistency of the feature maps. Subsequently, IFM performs further feature fusion based on feature alignment, generates global feature maps and splits them according to the needs of different scales. Finally, the Inject module enhances the feature expression capabilities of each layer through the self-attention mechanism and effectively distributes global features to each layer.

[0102] The aggregation and distribution mechanisms operate in the high and low stages respectively. The low stage focuses on small targets, while the high stage focuses on improving the detection accuracy of large targets. Figure 4 As shown in Figure 1, with B4 as the benchmark, the feature alignment module (Low-FAM) uses average pooling to align the features. The larger feature maps B2 and B3 are downsampled, while B5 is upsampled by bilinear interpolation to ensure that the size of the feature maps is uniform. These feature maps are merged into F concat , and processed by the low-level information fusion module (Low-IFM) using RepConv to generate F fusion . RepConv optimizes BatchNormalization parameters to enhance the expressiveness of fusion features. fusion Divided into F inj-P3 and F in-P4 :

[0103] F coneat =Low_FAM([B2,B3,B4,B5])

[0104] F fusion =RepConv(F concat )

[0105] F inj_P3 ,Finj_P4 =Split(F fusion ) (6)

[0106] At high levels, such as Figure 5 As shown in Figure 2, the feature alignment module (High-FAM) uses average pooling to reduce the feature size of P3 and P4, and these features are combined through the Concat operation. The merged feature F concat After being processed by a high-level information fusion module (High-IFM) with a Transformer module, F fusion . F fusion Divided into F inj-N4 and F in-N5 :

[0107] F concat =High_FAM([P3,P4,P5])

[0108] F fision = Transformer(F concat )

[0109] F iηj_N4 ,F iηj_N5 =Split(Conv(F fusion )) (7)

[0110] Information distribution modules (Inject) are used in both high and low stages, such as Figure 6 As shown, the global features generated by IFM are effectively fused with the local features of the current scale. The convolution process involves global features activated by the Sigmoid function as well as local features, and then the global feature residual is added. Information is further extracted and integrated through RepConv to ensure that features of different scales and stages are effectively merged and utilized. This significantly improves the detection accuracy and robustness of the model. The formulas involved are as follows:

[0111] F global_1_Pi = resize(Sigmoid(Conv(F inj_Pi )))

[0112] F global_2_Pi = resize(Conv(F inj_Pi ))

[0113] Pi = Rep.Conv(Conv local_Bi *F global_1_Pi +F global_2_Pi ) (8)

[0114] Among them, _local represents local features, and _global represents global features obtained after processing by the low-stage IFM module. These steps ensure that features between different scales and stages are effectively integrated and utilized, thereby improving the detection accuracy and robustness of the model.

[0115] The dynamic target detection head (DyHead) combines the three key perception capabilities of scale perception, spatial perception and task perception, such as Figure 7 As shown in the figure. In the DyHead architecture, the output of the backbone is regarded as a three-dimensional tensor with the dimensions of level×space×channel, where level represents the feature level, space represents the product of the width and height of the feature map, and channel is the number of channels. By introducing a self-attention mechanism between the feature layer, spatial position, and output channel, DyHead effectively improves the detection performance.

[0116] These self-attention mechanisms include inter-layer attention mechanism, spatial position attention mechanism and output channel attention mechanism. The inter-layer attention mechanism is used to achieve scale perception, the spatial position attention mechanism enhances spatial perception, and the output channel attention mechanism strengthens task processing capabilities. The combined application of these mechanisms significantly optimizes the expressiveness of the DyHead detection head without the need for additional computing resources. The implementation formula is as follows:

[0117] W(F)=π C (π S (π L (F)·F)·F)·F (9)

[0118] Among them, π represents the attention mechanism; L, S, and C represent the application of the attention mechanism in the inter-layer dimension, spatial position dimension, and channel dimension, respectively. L represents scale-aware attention; π S represents spatial perception attention; π C Represents task-aware attention.

[0119] By applying the above framework, it is possible to strengthen the expression of target detection features at the appropriate level and optimize the adaptability to different scales and spatial positions, thereby significantly enhancing the representation ability of the detection head and improving the detection performance of small targets. The formulas for calculating scale-aware attention, space-aware attention, and task-aware attention are as follows:

[0120]

[0121] in, It represents the average pooling operation on the feature map F in the dimensions of height H and width W to capture the global information of the image; f is the fully connected layer operation; σ is the hard-sigmoid function.

[0122] Example 2

[0123] like Figure 8 As shown, the abnormality identification method of the pulmonary ventilation function test of this embodiment, before performing AFD processing, also includes preprocessing the pulmonary ventilation function test curve to be identified, specifically, including the following steps:

[0124] S1. Obtaining a pulmonary ventilation function test curve to be identified;

[0125] S2. Preprocessing the pulmonary ventilation function test curve;

[0126] S3. Processing the pulmonary ventilation function test curve using an adaptive Fourier transform (AFD) method;

[0127] S4. Input the processed pulmonary ventilation function test curve into the trained YOLO algorithm model, and the YOLO algorithm model identifies the abnormality of the pulmonary ventilation function test curve.

[0128] In step S1 of the abnormality identification method,

[0129] Get the pulmonary ventilation function test curve to be identified, taking the FV curve as an example:

[0130] This embodiment extracts and preprocesses the image recording the FV curve, thereby obtaining the pulmonary ventilation function test curve to be identified.

[0131] The method for extracting from images with recorded FV curves uses a series of image processing techniques, mainly based on the OpenCV library, including template matching, Canny edge detection and Hough line transform methods.

[0132] Specifically, the template matching technique is used to locate and identify predefined template images, which represent the characteristic parts of the coordinate axis, such as corner points, scale lines or special marks, such as Fig. 9 As shown. After successfully identifying the coordinate axis position, further image processing and data analysis are performed. This includes cropping the image to only include the area near the coordinate axis and extracting data points in the chart or analyzing the graphical trend in this area. First, the image edges are highlighted through the Canny algorithm, which can maintain clear edge identification even in noisy images. Second, these edge points are transformed into Hough space through Hough line transform to identify straight lines, which represent key structural elements in the image, such as chart axes or boundaries.

[0133] In step S2 of the abnormality identification method,

[0134] Preprocess the FV curve:

[0135] In order to improve the accuracy and credibility of the extracted FV curve data, a comprehensive data preprocessing method was adopted.

[0136] First, the different colors in the FV curve are extracted by adjusting the image color. Then, the curves of different colors are accurately identified and labeled using the visual review method, successfully identifying abnormal data points caused by noise, which appear as irregular spikes and mutations. Based on the professional knowledge of pulmonary function testing, a threshold setting method is used to set a specific threshold T and evaluate each data point xi to identify and eliminate noise data points, as shown in the following formula:

[0137] ∣x i -μ∣>T·σ (11)

[0138] Among them, xi represents each data point; μ represents the mean value of the data set; σ represents the standard deviation of the data set.

[0139] Secondly, the interquartile range (IQR) is used as a statistical tool to define the normal fluctuation range of the data for lung function data without abnormal data points, effectively excluding data points that deviate from this range. IQR provides a robust outlier detection mechanism, significantly reducing the impact of extreme values ​​on the integrity of data analysis. The specific operation steps include: first sorting the FV curve data, then calculating the first quartile (Q1) and the third quartile (Q3) of the data set, and determining the IQR based on this, as shown in formula (2). This value reveals the data distribution between Q1 and Q3, that is, the middle 50% of the data range. Using IQR, those abnormal data points (xi) that may interfere with the overall analysis of the FV curve are effectively identified and excluded, as shown in formula (3).

[0140] IQR=Q3-Q1 (12)

[0141] Outlier definition:

[0142]

[0143] Among them, Q1 is the first quartile of the data, indicating that 25% of all the data are less than or equal to this value; Q3 is the third quartile of the data, indicating that 75% of all the data are less than or equal to this value.

[0144] like Fig.10 The figure shows a curve diagram after being processed by the interquartile range method. The interquartile range method can accurately eliminate abnormal data points caused by test errors, atypical breathing patterns of patients, or device measurement deviations, thereby ensuring the high accuracy and scientificity of FV curve analysis.

[0145] Again, signal segmentation and cubic spline interpolation methods were used to improve the reliability and accuracy of the data. The implementation formula is as follows.

[0146] S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3 (14)

[0147] Where S i (x) is the spline function of the interval containing x, x i is a node or control point, and a i 、b i 、c i、 d i are the coefficients of the spline function.

[0148] By using signal segmentation technology, the data can be accurately distinguished according to the physiological stage characteristics revealed by the FV curve, and each stage of the breathing process can be analyzed specifically to effectively identify and correct the data deviation caused by the patient's irregular breathing pattern. Furthermore, by applying cubic spline interpolation technology, a smooth transition is achieved at the missing data point, outliers are corrected, and the continuity and physiological relevance of the data are significantly improved, making the processed FV curve smoother while maintaining its physiological significance, such as Fig.11 shown.

[0149] It should be noted that the method of extracting curves from images and the preprocessing method listed in this embodiment are not necessary means to obtain the pulmonary ventilation function test curve to be identified, but are only one of the implementation methods.

[0150] The steps S3 and S4 are the same as those in Example 1 and will not be described again.

[0151] Example 3

[0152] The following describes an apparatus for executing the above-mentioned abnormality identification method for pulmonary ventilation function test.

[0153] Fig.12 The specific structural block diagram of the device provided in this embodiment is shown, and the device for identifying abnormal conditions of pulmonary ventilation function examination includes:

[0154] An input unit 10, used to obtain a pulmonary ventilation function test curve to be identified;

[0155] A processing unit 20, used for processing the pulmonary ventilation function test curve to be identified, wherein the processing is AFD processing, or preprocessing and AFD processing;

[0156] The detection unit 30 is used to identify abnormal conditions of the processed pulmonary ventilation function test curve according to the YOLO algorithm model.

[0157] The device provided in this embodiment and the abnormality identification method for pulmonary ventilation function test provided in this application belong to the same concept. The specific implementation process is detailed in the full text of the specification and will not be repeated here.

[0158] Example 4

[0159] The following describes an electronic device that can execute the above-mentioned abnormality identification method for pulmonary ventilation function test.

[0160] Fig.13 A specific structural block diagram of an electronic device 100 provided in this embodiment is shown, and the electronic device 100 includes: at least one memory 101 and at least one processor 102; the at least one memory 101 is coupled to the at least one processor 102, the at least one memory 101 is used to store a computer program, and the at least one processor 102 is used to call the computer program, and the computer program includes instructions, and when the instructions are executed by the at least one processor, the electronic device executes the above-mentioned abnormality identification method for pulmonary ventilation function examination.

[0161] Example 5

[0162] The following describes a computer storage medium containing the above-mentioned abnormality identification method for pulmonary ventilation function test.

[0163] A computer storage medium includes computer instructions. When the computer instructions are executed on an electronic device, the electronic device executes the above-mentioned abnormality identification method for pulmonary ventilation function examination.

[0164] Application Cases

[0165] Origin and processing of data

[0166] The data set involved in this invention was obtained from the pulmonary function test conducted by the Pulmonary Function Testing Center of the First Affiliated Hospital of Guangzhou Medical University and was used after privacy anesthesia. The data quality control process was strictly carried out in accordance with the 2019 updated version of the ATS / ERS pulmonary function test standardization guidelines, and two assessors who had received systematic pulmonary function knowledge training were responsible for independently performing the quality assessment. In the case of differences in the evaluation results, a senior expert with more than ten years of experience in pulmonary function testing was further adjudicated to ensure the accuracy and fairness of the data processing process.

[0167] To ensure the high quality of the dataset, we strictly graded the data. The dataset was divided into qualified and unqualified categories. The unqualified data was further subdivided into two categories: image data and numerical data, with eight subcategories. Image data included cough, initial leak, glottic closure, obstructed mouthpiece or spirometer, etc. Fig.14 As shown; the numerical data are FIVC-FVC<0.100L or 5% of FVC, Poor effort, Qualified, no expiratory plateau, and Hesitation. The image part was annotated by using the labelImg tool. It was divided into training set, validation set, and test set in a ratio of 7:2:1 for model training, performance verification, and testing. The training set was selected from selected spirometry reports, each report containing at least three FV curves, a total of 1000 reports were annotated, and 3000 curves were generated.

[0168] Training method of algorithm model

[0169] The training method for the improved YOLO algorithm model includes the following steps: performing the preprocessing and AFD processing on the original pulmonary ventilation function test curve, and then annotating the abnormal conditions of the pulmonary ventilation function test curve before and after the processing to obtain an image set; dividing the image set into a training set and a validation set in proportion, and using the training set and the validation set to train and evaluate the improved YOLO algorithm model.

[0170] Experimental equipment parameters and settings

[0171] The test experiment was carried out with Core TM i9-13900H CPU and The model training process was accelerated on a Windows 11 system with an RTX4090 GPU, using CUDA 11.7 and PyTorch 1.13.1. The specific parameter configuration is shown in Table 1.

[0172] In order to ensure the authenticity and reliability of the comparison with other algorithm models or recognition methods, other algorithm models or recognition methods are trained and tested under the same conditions as the algorithm model described in the present invention.

[0173] Table 1

[0174]

[0175]

[0176] A series of key performance indicators are used to comprehensively evaluate the comprehensive performance of the algorithm model and the abnormality recognition method of pulmonary ventilation function test. These indicators include floating-point operations per second (FLOPs), frames per second (FPS), mean average precision (mAP), precision (P), recall rate (R) and F1 score. FLOPs is an indicator for evaluating the computational complexity of the algorithm model. A lower value indicates that the algorithm model can maintain high efficiency in an environment with limited computing resources. The FPS indicator measures the speed at which the algorithm model processes data. A higher FPS indicates that the algorithm model can respond quickly in real-time analysis scenarios. mAP evaluates the generalization ability and accuracy of the algorithm model in multiple pulmonary function abnormality detection tasks. The F1 score evaluates the algorithm model's ability to identify real abnormal situations with high accuracy by combining precision and recall rate. These comprehensive indicators ensure the accuracy and practicality of the algorithm model. The specific calculation formula is as follows:

[0177]

[0178] Among them, TP represents the number of correctly predicted positive samples, FP represents the number of negative samples incorrectly predicted as positive, FN represents the number of positive samples incorrectly predicted as negative, N represents the total number of categories, and T pre Represents the image preprocessing time, T inf represents the model inference time, T NMS Indicates the time for non-maximum suppression of each image.

[0179] Experimental results and analysis

[0180] Comparison of AFD performance evaluation with other methods

[0181] AFD technology is used to perform multi-scale time-frequency decomposition on the pulmonary ventilation function test curve. By increasing the decomposition level, more accurate signal features can be extracted. Fig.15 As shown in FIG. 1 , the step-by-step decomposition process of the pulmonary ventilation function test curve data using the 1st, 2nd, and 3rd order AFD methods is shown in detail. As one implementation method, the data of the pulmonary ventilation function test curve are decomposed into 6 frequency components. Each column in the figure is marked as the first to sixth frequency components (Comp1-Comp6) from top to bottom, each representing a different frequency range, and the frequency of the frequency components is from low to high. From left to right, column (a) is the 1st order AFD, column (b) is the 2nd order AFD, and column (c) is the 3rd order AFD.

[0182] The decomposition and aggregation process in p-order AFD is expressed as follows:

[0183] According to the formulas (4) and (5) shown above

[0184] The pulmonary ventilation function test curve data is expressed as follows after the decomposition process in the p-order AFD:

[0185] Among them, Comp1 in the 2nd-order AFD:

[0186] At this time, it is the second order, and the value of p is 2; at this time, it is the first frequency component, and the value of q is 1; the number of frequency components is 6, and the corresponding value of k is 6; n p The value is 12.

[0187] After calculation, m(q,p)=1; M(q,p)=2;

[0188] So we get

[0189] Comp2 in 2nd-order AFD:

[0190] At this time, it is the second order, and the value of p is 2; at this time, it is the second frequency component, and the value of q is 2; the number of frequency components is 6, and the corresponding value of k is 6; n p The value is 12.

[0191] After calculation, m(q,p)=3; M(q,p)=4;

[0192] So we get

[0193] Similar to the above calculation, Comp3~6 in the 2nd-order AFD are obtained respectively:

[0194] Comp3:

[0195] Comp4:

[0196] Comp5:

[0197] Comp6:

[0198] For Comp1 in 3rd order AFD:

[0199] At this time, it is the third order, and the value of p is 3; at this time, it is the first frequency component, and the value of q is 1; the number of frequency components is 6, and the corresponding value of k is 6; n p The value is 18.

[0200] After calculation, m(q,p)=1; M(q,p)=3;

[0201] So we get

[0202] Comp2 in 3rd-order AFD:

[0203] At this time, it is the third order, and the value of p is 3; at this time, it is the first frequency component, and the value of q is 2; the number of frequency components is 6, and the corresponding value of k is 6; n p The value is 18.

[0204] After calculation, m(q,p)=4; M(q,p)=6;

[0205] So we get

[0206] Similar to the above calculation, Comp3~6 in the 3rd-order AFD are obtained respectively:

[0207] Comp3:

[0208] Comp4:

[0209] Comp5:

[0210] Comp6:

[0211] Analysis of Comp1 shows that through the second and third stage decomposition, the frequency components retain key features, such as coughing. By superimposing these frequency components containing key features, the original signal can be reconstructed to obtain a smoother pulmonary ventilation function test curve and reduce noise.

[0212] As one of the implementation methods, superimposing the first four frequency components in the same column can retain key features and prevent model overfitting. For example, for a 2nd-order AFD, frequency components Comp1 to Comp4 are superimposed. Similarly, for a 3rd-order AFD, frequency components Comp1 to Comp4 are superimposed. After superposition, a reconstructed pulmonary ventilation function test curve is obtained for the next step of algorithm model training or evaluation and identification.

[0213] To further evaluate the effectiveness of the p-order AFD, the 3rd-order AFD is compared with two widely used time-frequency analysis methods, EMD and VMD, as Fig.16 As shown in the figure, while both VMD and 3rd-order AFD retain important features such as cough, 3rd-order AFD provides a clearer and more explicit representation. In contrast, EMD suffers from pattern mixing, resulting in the loss of key signal characteristics. To further verify the effectiveness of these decomposition methods, the first 4 frequency components of each method are further integrated into the algorithm model for comparison.

[0214] To this end, the YOLOv8 algorithm model was used as a benchmark, and a series of comparative experiments were conducted to evaluate different mathematical signal decomposition methods. At the same time, the effects of replacing the C2f module in the backbone network with C2f-EMSCP, replacing the PAN structure of the neck network with the Gather-Distribute mechanism, and the DyHead of the head network on performance were also evaluated. The results are shown in Table 2. After processing, the mAP index has improved, especially the third-order AFD, which has increased significantly compared with EMD and VMD. This shows that the mathematical decomposition method effectively reduces the interference of high-frequency signals in the pulmonary ventilation function test curve and reduces the error rate.

[0215] Table 2

[0216]

[0217] Experimental results show that after improving the backbone network of the model through EMSCP, the mAP increased by 0.6%, 2.6% and 3.3% under the conditions of EMD, VMD and 3rd-order AFD, respectively, while the number of parameters and the amount of calculation decreased. This shows that the EM-SCP module reduces the redundant information in the C2f module channel and effectively reduces the convolution operation, thereby improving the performance while reducing the complexity of the algorithm model.

[0218] After the neck network was replaced with the Gather-Distribute mechanism, the mAP increased by 0.9%, 3.6%, and 4.1% under the EMD, VMD, and AFD conditions, respectively, with a slight increase in the number of parameters and computation. This shows that the Gather-Distribute mechanism effectively solves the problem of information loss during downsampling and improves accuracy. Despite the increased computational complexity, the FPS of the algorithm model shows that real-time performance is maintained.

[0219] Replacing the head network with DyHead improves mAP by 0.8%, 3.4%, and 3.8% under EMD, VMD, and AFD conditions, respectively. DyHead expands the receptive field, focuses on target features, and reduces background interference, thereby enhancing the recognition of small targets and key features, thereby improving detection performance.

[0220] In the improved YOLOv8 algorithm model, EMSCP, Gather-Distribute mechanism and DyHead are combined to improve mAP by 1.7%, 4.1% and 5.0% under EMD, VMD and AFD conditions respectively. Despite a slight increase in parameters under AFD conditions, the model remains lightweight and real-time reasoning is not affected. The detection speed is increased by 53FPS in the GPU environment and 103FPS in the CPU environment. EMSCP compensates for the computational complexity introduced by the Gather-Distribute mechanism, enabling the YOLOv8 algorithm model to maintain high efficiency while significantly improving the speed and accuracy of FV curve detection.

[0221] Comparison of different algorithm models

[0222] In order to further verify the effectiveness of the algorithm model adopted by the present invention, that is, the combination of p-order AFD and YOLOv8 (denoted as: YOLO-AFD, where p is 3), it is compared with other target detection algorithm models and several lightweight models. All algorithm models are trained and tested under the same conditions. The results are shown in Table 3.

[0223] Table 3

[0224]

[0225] The data in the table above show that YOLO-AFD has significant advantages over other algorithm models in terms of detection accuracy, number of parameters, and computational load. Although Faster-RCNN has a higher mAP than the original YOLOv8, its huge amount of computation and model size result in a long inference time, making it unsuitable for real-time detection tasks, that is, it is not suitable for detecting abnormalities in the pulmonary ventilation function curve.

[0226] In contrast, YOLO-AFD improves mAP by 5% and surpasses Faster-RCNN in detection accuracy. YOLO-AFD performs well in detection speed and accuracy, and is particularly suitable for detecting abnormal pulmonary ventilation function.

[0227] To further test the quality control of YOLO-AFD for pulmonary ventilation function curve inspection, the results are shown in Table 4 compared with other methods. YOLO-AFD achieved higher performance in terms of mAP, highlighting its effectiveness in identifying abnormal pulmonary ventilation function.

[0228] Table 4

[0229]

[0230] The t-distributed stochastic neighbor embedding (t-SNE) algorithm is used to reduce the dimension of the semantic features extracted from the final fully connected layer of the tested algorithm model, such as Fig.17 As shown, the results are displayed on a two-dimensional plane. The YOLO-AFD algorithm model outperforms EMD and VMD in terms of accuracy and data discrimination. Compared with other algorithm models, the original YOLOv8 algorithm model and the YOLO-EMD algorithm model show chaotic data points without obvious clustering structure. The YOLO-VMD algorithm model forms some clusters, but there is significant overlap and the feature discrimination is not ideal. In contrast, the YOLO-AFD algorithm model shows excellent feature spacing and clear classification effects, verifying the reliability of the algorithm model.

[0231] In order to address the opacity of the algorithm model's decision logic, Gradient-weighted ClassActivation Mapping (Grad-CAM) is used to enhance visual interpretability. Grad-CAM generates heat maps that highlight the key areas that the algorithm model focuses on during the recognition process, such as spikes and steep waveform abnormalities that appear in the pulmonary ventilation function curve. Grad-CAM generates heat maps by weighted summation of activations on different feature layers, thereby enhancing interpretability without modifying the network architecture. Critical areas are displayed in red or orange, while non-critical areas are displayed in blue or purple, such as Fig.18 shown.

[0232] When using the YOLO-AFD algorithm model to analyze the FV curve, the model showed obvious advantages and had a high accuracy rate in identifying abnormalities such as spikes during coughing and curve drops during air leakage.

[0233] The original YOLOv8 and the improved YOLO-EMD often mistake general waveform fluctuations as significant features, resulting in excessive attention to the entire waveform. Although the improved YOLO-VMD algorithm model increases the focus on excessive fluctuation areas, such as spikes and spikes, it still lacks in precise positioning. In contrast, the YOLO-AFD algorithm model can more accurately extract and highlight these key features, representing highly focused areas in dark red, and the color depth intuitively indicates the importance of the feature.

[0234] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for identifying abnormalities in a pulmonary ventilation function test, characterized in that: The following steps are involved: S1. Obtaining a pulmonary ventilation function test curve to be identified, and processing the pulmonary ventilation function test curve using an adaptive Fourier transform (AFD) method; S2. Input the processed pulmonary ventilation function test curve into the trained YOLO algorithm model, and the YOLO algorithm model identifies the abnormality of the pulmonary ventilation function test curve.

2. The abnormality identification method for pulmonary ventilation function test according to claim 1, characterized in that: The processing of the pulmonary ventilation function test curve using the adaptive Fourier method includes: performing p-order adaptive Fourier decomposition on the pulmonary ventilation function test curve, and obtaining multiple frequency components through step-by-step decomposition, wherein the frequency components include frequency components containing key information and noise frequency components; removing the noise frequency components, and reconstructing the pulmonary ventilation function test curve according to the frequency components containing key information; p≥2; preferably, p is 2, 3, 4 or 5.

3. The abnormality identification method for pulmonary ventilation function test according to claim 2, characterized in that: The multiple frequency components are arranged in order from low to high frequency, and the first x low-frequency frequency components are selected from the multiple frequency components as the frequency components containing key information; x is 3 to 6; preferably, x is 3, 4, 5 or 6.

4. The abnormality identification method for pulmonary ventilation function test according to claim 1, characterized in that: The YOLO algorithm model is an improved YOLOv8 algorithm model; the improved YOLOv8 algorithm model is mainly composed of a backbone network, a neck network and a head network, the backbone network uses C2f-EMSCP to replace the C2f module of the original YOLOv8 algorithm model; the aggregation and distribution mechanism is used to replace the neck network of the original YOLOv8 algorithm model; the dynamic target detection head is used to replace the head network of the original YOLOv8 algorithm model; The input end of the backbone network is used to introduce the pulmonary ventilation function test curve obtained after AFD processing.

5. The abnormality identification method for pulmonary ventilation function test according to claim 1, characterized in that: In step S1, before performing AFD processing, the pulmonary ventilation function test curve to be identified is also preprocessed; the preprocessing includes processing the data of the pulmonary ventilation function test curve using threshold setting and interquartile range method, and then processing it using signal segmentation and cubic spline interpolation method.

6. The abnormality identification method for pulmonary ventilation function test according to claim 1 or 5, characterized in that: The training method of the YOLO algorithm model includes the following steps: performing the preprocessing and AFD processing on the original pulmonary ventilation function test curve, and then annotating the abnormal conditions of the processed pulmonary ventilation function test curve to obtain an image set; dividing the image set into a training set and a validation set in proportion, and using the training set and the validation set to train and evaluate the YOLO algorithm model.

7. The abnormality identification method for pulmonary ventilation function test according to claim 1, characterized in that: The method of obtaining the pulmonary ventilation function test curve to be identified includes: directly using the pulmonary ventilation function test numerical data to generate the pulmonary ventilation function test curve, or extracting it from an image recording the pulmonary ventilation function test curve.

8. A device for identifying abnormal conditions in a pulmonary ventilation function test, characterized in that: include: An input unit, used for obtaining a pulmonary ventilation function test curve to be identified; A processing unit, used for processing the pulmonary ventilation function test curve to be identified, wherein the processing is AFD processing, or preprocessing and AFD processing; The detection unit is used to identify abnormal conditions of the processed pulmonary ventilation function test curve according to the YOLO algorithm model.

9. An electronic device, characterized in that: comprising at least one memory and at least one processor; The at least one memory is coupled to the at least one processor, the at least one memory is used to store a computer program, the at least one processor is used to call the computer program, the computer program includes instructions, when the instructions are executed by the at least one processor, the electronic device executes the abnormality identification method for pulmonary ventilation function test as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: It comprises computer instructions, and when the computer instructions are executed on an electronic device, the electronic device executes the abnormality identification method for pulmonary ventilation function test as claimed in any one of claims 1 to 7.

Images