Intelligent airway management system and method

Abstract

The invention provides an intelligent airway management system and method, and the method comprises the steps: screening an effective region of an intubation structure in each airway image through the pixel contribution degree in each airway image, and determining the recognition attention of each airway image; determining a spatial topological structure relationship of the intubation structure in the airway of the patient among multiple view angles according to a spatial mapping relationship among key nodes related to the intubation structure in each airway image; performing cross-view cross fusion on each local feature and each global feature in the airway image at different visual angles based on the spatial topological structure relationship and all recognition attention to obtain a cross fusion feature vector of the airway state of the patient; and generating a risk prediction level of the difficult airway of the patient by combining a pre-constructed difficult airway prediction model with the cross fusion feature vector of the airway state of the patient. By adopting the scheme of the invention, airway state recognition based on a multi-view spatial topological relation can be realized, and the accuracy of difficult airway prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of airway management technology, and more specifically, to an intelligent airway management system and method. Background Art

[0002] Airway management is a core operation in the fields of anesthesia, intensive care, and emergency medicine. Its main goal is to ensure that patients receive adequate oxygen supply and carbon dioxide removal to ensure unobstructed breathing, prevent suffocation, and optimize oxygen supply. Therefore, airway management is widely used in anesthesia, emergency department, intensive care, trauma treatment, and pre-hospital emergency care. Effective airway management can reduce the risk of respiratory failure and improve the success rate of rescue. In medical practice, airway management involves multiple links such as endotracheal intubation, mechanical ventilation, ventilation monitoring, and prevention of airway complications.

[0003] Intelligent airway management is a modern medical technology based on artificial intelligence, sensor technology and automated control, which aims to improve the safety, accuracy and individualization of airway management. During clinical anesthesia and emergency treatment, rapid and accurate identification of difficult airways is crucial to ensuring the safety of intubation operations. However, in existing airway management, the prediction of difficult airways still relies mainly on manual experience and single-view image analysis. It is limited by a single observation angle, limited utilization of spatial information and weak local structure recognition capabilities, resulting in insufficient information dimensions, poor spatial structure perception capabilities, and weak model adaptability to individual differences. These problems often lead to low prediction accuracy, making it difficult to detect potential high-risk intubated patients in a timely manner. There is a large risk of misjudgment, resulting in low accuracy in difficult airway prediction and unstable airway risk assessment. Therefore, how to realize airway status recognition based on multi-view spatial topological relationships to improve the accuracy of difficult airway prediction has become a difficult problem faced by the industry. Summary of the Invention

[0004] The present application provides an intelligent airway management system and method, which can realize airway state recognition based on multi-perspective spatial topological relationships to improve the accuracy of difficult airway prediction.

[0005] In a first aspect, the present application provides a difficult airway prediction method for an intelligent airway management system to predict a patient's difficult airway, the method comprising the following steps: Acquire the patient's airway images at different viewing angles; The effective areas of the intubation structure in each airway image at different viewing angles are screened by the pixel contribution when identifying the intubation structure in each airway image. Based on each effective area and the patient's cervical spine mobility, the recognition attention of each airway image when predicting a difficult airway is determined. Multi-scale dilated convolution is performed on each airway image to extract local and global features of each airway image. The spatial topological relationship of the intubation structure in the patient's airway across multiple viewpoints is determined based on the spatial mapping relationship between key nodes related to the intubation structure in each airway image. Based on the spatial topological structure relationship and all recognition attentions, each local feature and each global feature are cross-fused across views to obtain a cross-fused feature vector of the patient's airway state; The risk prediction level of the patient's difficult airway is generated by combining a pre-built difficult airway prediction model with a cross-fusion feature vector of the patient's airway status.

[0006] In some embodiments, screening out effective areas of the cannula structure in the airway image at different viewing angles based on pixel contribution when identifying the cannula structure in each airway image specifically includes: An airway image under a viewing angle is selected as a selected airway image; dividing the selected airway image into a plurality of sub-regions; Determine the local gradient change of pixels in each sub-region, and then determine the pixel contribution of each sub-region to the identification of the cannula structure; Filter out all sub-regions in the selected airway image whose pixel contributions are greater than a preset threshold, and form the valid region of the intubation structure in the airway image under the corresponding viewing angle; Continue to screen out the valid areas of the cannula structure in the airway images under the remaining viewing angles.

[0007] In some embodiments, determining each airway image based on each effective area in combination with the patient's cervical spine mobility to perform difficult airway prediction specifically includes: Determine the patient's cervical spine mobility based on the patient's cervical spine motion imaging; adjusting the attention weights of the airway images at different viewing angles according to the cervical spine mobility; The recognition attention of each airway image when performing difficult airway prediction is determined by adjusting the various attention weights and effective areas.

[0008] In some embodiments, performing multi-scale dilated convolution on each airway image to extract local features and global features of each airway image specifically includes: Perform multi-scale dilation convolution on each airway image according to multiple dilation rates to obtain feature maps of each airway image at different scales; Local and global features of each airway image are extracted based on feature maps at different scales.

[0009] In some embodiments, determining the spatial topological structure relationship of the cannula structure in the patient's airway between multiple perspectives based on the spatial mapping relationship between key nodes related to the cannula structure in each airway image specifically includes: Demarcate key nodes related to the intubation structure in each airway image; Determine the spatial mapping relationship between key nodes related to the intubation structure in each airway image based on the calibration results; Determine multi-view correlation information of the cannula structure in the patient's airway through the spatial mapping relationship; The spatial topological structure relationship of the intubation structure in the patient's airway between the multiple perspectives is determined based on the multi-perspective association information.

[0010] In some embodiments, cross-view cross-fusion of local features and global features is performed based on the spatial topological structure relationship and all recognition attentions to obtain a cross-fusion feature vector of the patient's airway state, specifically including: Perform weighted processing on each local feature and each global feature based on all recognition attention; Based on the spatial topological structure relationship, the same intubation structure features in the airway images from different perspectives are aligned and matched, and then a long-range dependency relationship is established between the intubation structure features obtained by the alignment and matching; The weighted local features and global features are interactively fused according to the long-distance dependency relationship to obtain a cross-fusion feature vector of the patient's airway state.

[0011] In some embodiments, generating a patient's difficult airway risk prediction level by combining a pre-built difficult airway prediction model with a cross-fusion feature vector of the patient's airway status specifically includes: Pre-build a difficult airway prediction model; Using the cross-fused feature vector of the patient's airway status as input to the difficult airway prediction model; The difficult airway prediction model outputs a risk prediction level of the patient's difficult airway.

[0012] In a second aspect, the present application provides an intelligent airway management system, including a difficult airway prediction unit, wherein the difficult airway prediction unit includes: An acquisition module, used to acquire airway images of the patient at different viewing angles; A processing module is used to screen out effective areas of the intubation structure in the airway image at different viewing angles based on the pixel contribution when identifying the intubation structure in each airway image, and determine the recognition attention of each airway image when performing difficult airway prediction based on each effective area and the patient's cervical spine mobility; The processing module is further configured to perform multi-scale dilated convolution on each airway image to extract local and global features of each airway image, and determine the spatial topological relationship of the intubation structure in the patient's airway between multiple perspectives based on the spatial mapping relationship between key nodes related to the intubation structure in each airway image; The processing module is further configured to perform cross-view cross-fusion of each local feature and each global feature based on the spatial topological structure relationship and all recognition attentions to obtain a cross-fusion feature vector of the patient's airway state; An execution module is used to generate a risk prediction level of the patient's difficult airway by combining a pre-built difficult airway prediction model with a cross-fusion feature vector of the patient's airway status.

[0013] In a third aspect, the present application provides a computer device, comprising a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the above-mentioned difficult airway prediction method.

[0014] In a fourth aspect, the present application provides a computer-readable storage medium, in which instructions or codes are stored. When the instructions or codes are run on a computer, the computer implements the above-mentioned difficult airway prediction method when executing the computer.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: In the present application, airway images of the patient at different perspectives are obtained; the effective areas of the intubation structure in the airway images at different perspectives are screened out based on the pixel contribution when identifying the intubation structure in each airway image, and the recognition attention of each airway image for difficult airway prediction is determined based on each effective area combined with the patient's cervical spine mobility; multi-scale dilated convolution is performed on each airway image to extract the local features and global features of each airway image, and the spatial topological structure relationship of the intubation structure in the patient's airway between multiple perspectives is determined based on the spatial mapping relationship between the key nodes related to the intubation structure in each airway image; based on the spatial topological structure relationship and all the recognition attention, each local feature and each global feature are cross-fused across views to obtain a cross-fused feature vector of the patient's airway status; the risk prediction level of the patient's difficult airway is generated by combining a pre-constructed difficult airway prediction model with the cross-fused feature vector of the patient's airway status.

[0016] It can be seen that in this application, firstly, the recognition attention of each airway image when predicting a difficult airway is determined based on each effective area combined with the patient's cervical spine mobility, which can reflect the changes in the patient's airway visibility under different perspectives, make up for the information loss caused by the observation angle deviation of a single perspective, and improve the expression weight of the key intubation area in the airway image, providing a more valuable information basis for subsequent feature extraction and multi-perspective fusion; secondly, the spatial topological structure relationship of the intubation structure in the patient's airway between multiple perspectives is determined based on the spatial mapping relationship between the key nodes related to the intubation structure in each airway image, which can effectively make up for the defects of the single-perspective structure projection information, so that the corresponding logic of the intubation-related structure can be established between different perspectives, so that the subsequent airway image features no longer exist in isolation, but have associated meaning under the spatial logic, thereby improving the subsequent understanding of the variation of complex intubation structures. It is significantly helpful for the identification of difficult airways with obvious position offset or abnormal morphology of the intubation structure, thereby enhancing the comprehensiveness and reliability of difficult airway prediction; then, based on the spatial topological structure relationship and all the recognition attention, each local feature and each global feature is recognized. The cross-view features are cross-fused to obtain a cross-fused feature vector of the patient's airway status. By aligning and weighting key intubation structural features between views, long-distance information linkage and deep feature interaction are achieved across views, fusing local and global semantic information in each perspective. This not only preserves the integrity of structural details but also enhances global perception, making the recognition of intubation structure more spatially consistent. Finally, a pre-built difficult airway prediction model is combined with the cross-fused feature vector of the patient's airway status to generate a risk prediction level for the patient's difficult airway. This significantly improves the difficult airway prediction model's ability to recognize complex and variable airway states, allowing high-risk intubation patients to be accurately identified earlier. Compared with traditional classification models that rely on single-view static images, this approach has stronger information integration capabilities and risk identification efficiency, fundamentally enhancing the accuracy and stability of difficult airway recognition in the intelligent airway management system. Furthermore, the risk prediction level of the patient's difficult airway can assist in formulating airway intubation strategies. In summary, this solution can realize airway status recognition based on multi-view spatial topological relationships and improve the accuracy of difficult airway prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0018] Figure 1is an exemplary flow chart of a difficult airway prediction method according to some embodiments of the present application; Figure 2 is an exemplary flow chart for screening effective areas according to some embodiments of the present application; Figure 3 is an exemplary flow chart of determining spatial topological structure relationships according to some embodiments of the present application; Figure 4 is a schematic structural diagram of a difficult airway prediction unit according to some embodiments of the present application; Figure 5 3 is a schematic diagram of the structure of a computer device for implementing a difficult airway prediction method according to some embodiments of the present application. DETAILED DESCRIPTION

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

[0020] refer to Figure 1 , which is an exemplary flow chart of a difficult airway prediction method according to some embodiments of the present application. The difficult airway prediction method 100 mainly includes the following steps: In step 101 , airway images of a patient at different viewing angles are acquired.

[0021] In specific implementation, multiple camera modules or three-dimensional structured light devices can be arranged in the patient's head and neck area to capture airway images from multiple angles (front, side, upward, etc.), thereby obtaining airway images of the patient at different perspectives. Among them, it is necessary to simultaneously capture airway images from multiple perspectives to ensure image temporal consistency and avoid image dislocation caused by slight movement of the patient to be intubated. Other methods can also be used to implement this in other embodiments, which are not specifically limited here.

[0022] It should be noted that the airway images in this application are used to reflect the anatomical information of the patient's airway at different perspectives.

[0023] In step 102, the effective areas of the intubation structure in the airway images under different viewing angles are screened out by the pixel contribution when identifying the intubation structure in each airway image, and the recognition attention of each airway image when predicting a difficult airway is determined based on each effective area and the patient's cervical spine mobility.

[0024] In some embodiments, screening out valid areas of the cannula structure in the airway images at different viewing angles by using pixel contribution when identifying the cannula structure in each airway image can be achieved by the following steps: An airway image under a viewing angle is selected as a selected airway image; dividing the selected airway image into a plurality of sub-regions; Determine the local gradient change of pixels in each sub-region, and then determine the pixel contribution of each sub-region to the identification of the cannula structure; Filter out all sub-regions in the selected airway image whose pixel contributions are greater than a preset threshold, and form the valid region of the intubation structure in the airway image under the corresponding viewing angle; Continue to screen out the valid areas of the cannula structure in the airway images under the remaining viewing angles.

[0025] In a specific implementation, the selected airway image can be divided into multiple sub-regions in the following manner, namely: the selected airway image can be divided into multiple sub-regions by an image grid division algorithm (such as a sliding window or a fixed-size slicing method), for example, the size of the sub-region is set to 16×16 or 32×32 pixels. In other embodiments, the size of the sub-region can also be adjusted according to the image resolution or the typical size of the intubation structure to facilitate local feature analysis of different spatial regions in the selected airway image; determining the local gradient change of the pixels in each sub-region, and then determining the pixel contribution of each sub-region to the identification of the intubation structure can be achieved in the following manner That is, a gradient operator (such as Sobel, Scharr, or Canny) can be used to calculate the local gradient change of pixels in each sub-region. For example, an edge operator (Sobel, etc.) is used to extract a gradient amplitude map, and then the mean value of the gradient amplitude in each sub-region is counted as the pixel contribution of the corresponding sub-region when identifying the intubation structure, so as to explore the structural significance of the local region in the selected airway image and provide a numerical basis for subsequent region screening. Among them, the higher the pixel contribution, the more obvious the edge structure information of the corresponding sub-region, which can be used for the model to identify the relevant parts of the intubation. In other embodiments, other methods can also be used for determination, which is not limited here.

[0026] For specific implementation, refer to Figure 2As shown, this figure is an exemplary flow chart for screening effective regions in some embodiments of the present application. Screening out subregions in a selected airway image where all pixel contributions are greater than a preset threshold to form the effective region of the intubation structure in the airway image at the corresponding viewing angle can be achieved in the following manner: first, an adaptive threshold method (e.g., based on a histogram or Otsu algorithm) can be used to set a threshold based on the above-mentioned pixel contribution results and historical experimental experience. The threshold is used to screen out key subregions in the selected airway image. Then, all subregions in the selected airway image where pixel contributions are greater than the threshold are screened and, combined with connected domain analysis techniques, adjacent regions are merged. Morphological operations (e.g., dilation and erosion) are then used to optimize region boundaries, ultimately forming an effective region mask containing continuous high-contribution pixels. This effective region mask is then used as the effective region of the intubation structure in the airway image at the corresponding viewing angle. This mask is the most valuable structure recognition region at the current viewing angle, laying the spatial perception foundation for subsequent multi-view structure topology construction and feature fusion. Other screening methods may also be used in other embodiments, which are not limited here.

[0027] It should be noted that the sub-region in this application represents a small image area in the airway image; the local gradient change in this application represents the degree of change in the spatial variation of the brightness value in the sub-region, which is used to reflect the strength of the change in the internal edge structure of the corresponding sub-region in the airway image; the pixel contribution in this application represents the degree of influence of the pixel in the airway image on the recognizability of the intubation structure, which is an important basis for determining the structural validity; the effective area in this application represents the image part in the airway image that plays a key role in determining the intubation path, which has representative structural features and is the key area for subsequent feature extraction.

[0028] In some embodiments, determining the recognition attention of each airway image for difficult airway prediction based on each effective area combined with the patient's cervical spine mobility can be achieved by the following steps: Determine the patient's cervical spine mobility based on the patient's cervical spine motion imaging; adjusting the attention weights of the airway images at different viewing angles according to the cervical spine mobility; The recognition attention of each airway image when performing difficult airway prediction is determined by adjusting the various attention weights and effective areas.

[0029] In specific implementation, determining the patient's cervical spine range of motion based on the patient's cervical spine motion image can be achieved in the following manner, namely: the changes in the patient's cervical spine posture during the movement process (head flexion and extension, left and right rotation, etc.) can be recorded by a visual ultrasound or three-dimensional structured light camera system to form a continuous motion image sequence. Then, medical image registration and skeleton tracking technology (such as a bone point detection algorithm based on posture estimation) can be used to automatically calculate the angle change range of key cervical spine segments (C1-C7, etc.) based on the cervical spine motion image, thereby quantifying the patient's cervical spine range of motion, which is usually expressed as the total value of the flexion and extension angle, rotation angle and range of motion, and normalized. Finally, the result obtained after normalization is used as the patient's cervical spine range of motion; adjusting the attention weight of the airway image at different perspectives according to the cervical spine range of motion can be achieved in the following manner, namely: assigning corresponding attention weights to the airway images at different perspectives according to the cervical spine range of motion. For example, when the patient's cervical spine movement is limited, the anatomical structure of the front-view airway image is more clearly exposed, and the angle of rotation can be automatically increased. The attention distribution (i.e., attention weight) of the airway image under this perspective, and for the airway image obtained under excessive torsion angle, the attention weight of the airway image under this perspective is reduced accordingly. This process can be implemented by a weighting function or an attention allocation model (such as Softmax weighting or adaptive attention module). In other embodiments, other methods can also be used to implement it, which is not limited here; by adjusting the various attention weights and each effective area, the recognition attention of each airway image for difficult airway prediction can be determined in the following way, namely: the effective area extracted from each airway image is fused with the corresponding attention weight region by region to form a spatial weight map reflecting the importance of airway image recognition, and each spatial weight map is used as the recognition attention of the airway image under the corresponding perspective in the difficult airway prediction process, so as to guide the subsequent difficult airway prediction model to focus on high-value areas, thereby realizing unified regulation of the recognition attention of multi-perspective airway images; in other embodiments, other methods can also be used for determination, which is not limited here.

[0030] It should be noted that the cervical spine mobility in this application represents the range of influence of the patient's head and neck movement on the opening and closing of the airway structure; the weight force in this application represents the intensity of attention given to the airway image under the corresponding perspective, which is used to adjust the resource allocation of the difficult airway model in multi-perspective signal processing; the recognition attention in this application represents the spatial attention distribution of the effective area of ​​the intubation structure in the airway image under the corresponding perspective during the difficult airway prediction process, which can be used to guide the difficult airway prediction model to focus on the image area that plays a key role in the final recognition.

[0031] In step 103, multi-scale dilated convolution is performed on each airway image to extract local and global features of each airway image. The spatial topological relationship of the intubation structure in the patient's airway between multiple perspectives is determined based on the spatial mapping relationship between key nodes related to the intubation structure in each airway image.

[0032] In some embodiments, performing multi-scale dilated convolution on each airway image to extract local features and global features of each airway image can be achieved by the following steps: Perform multi-scale dilation convolution on each airway image according to multiple dilation rates to obtain feature maps of each airway image at different scales; Local and global features of each airway image are extracted based on feature maps at different scales.

[0033] In specific implementation, multi-scale dilated convolution is performed on each airway image according to multiple dilation rates, and the feature maps of each airway image at different scales can be obtained in the following way, namely: dilated convolution operations with multiple dilation rates (such as 1, 2, 4, and 8) are introduced to each airway image, and each dilation rate corresponds to a receptive field size to form a multi-scale dilated convolution module. The multi-scale dilated convolution module inserts holes inside the convolution kernel (that is, skips several pixels in the convolution operation) without increasing the size of the convolution kernel, thereby modeling the association between local and distant pixels in the airway image, thereby outputting feature maps at different scales; extracting local and global features of each airway image based on feature maps at different scales can be achieved in the following way, namely: the YOLO system can be used to extract local and global features of each airway image. The feature extraction methods in the column (such as YOLOv3, YOLOv4, YOLOv5, etc.) perform feature response analysis on the feature maps at each scale respectively. For example, the feature maps generated by small dilation rates (such as 1 and 2) reflect the detailed information (such as texture and edges) of the local area in the image. Such feature extraction is used as the local feature of the corresponding airway image, which is mainly used to identify the boundary and detailed configuration of the intubation structure in the airway image; the feature maps generated by large dilation rates (such as 4 and 8) have a wider coverage range and can perceive the morphology and spatial layout of the entire airway in the airway image. Such feature extraction is used as the global feature of the corresponding airway image to capture the overall accessibility and structural patency of the intubation path in the airway image; other methods can also be used to implement this in other embodiments, which are not limited here.

[0034] It should be noted that the dilation rate in this application refers to the spacing distance of the sampling points in the convolution kernel in the dilated convolution, which determines the size of the receptive field and is used to control the scale range of the convolution operation; the feature maps at different scales in this application represent the feature maps obtained under different dilation rates, which contain image structure information from details to the whole; the local features in this application represent the features that describe the microstructural details of the intubation-related areas in the airway image; the global features in this application represent the features that reflect the spatial contours and anatomical distribution of the airway area in the airway image.

[0035] In some embodiments, reference Figure 3 As shown in FIG. 1 , this figure is an exemplary flow chart for determining spatial topological structure relationships in some embodiments of the present application. In this embodiment, determining the spatial topological structure relationship of the cannula structure in the patient's airway between multiple perspectives based on the spatial mapping relationship between key nodes related to the cannula structure in each airway image can be achieved by the following steps: First, in step 1031 , key nodes related to the intubation structure in each airway image are calibrated; Next, in step 1032 , the spatial mapping relationship between key nodes related to the intubation structure in each airway image is determined based on the calibration results; Then, in step 1033, multi-view correlation information of the intubation structure in the patient's airway is determined through the spatial mapping relationship; Finally, in step 1034, the spatial topological structure relationship of the intubation structure in the patient's airway between multiple perspectives is determined based on the multi-perspective association information.

[0036] In specific implementation, the key nodes related to the intubation structure in each airway image can be calibrated in the following way: based on the key point detection network (such as HRNet (High-Resolution Network or OpenPose structure), feature points are extracted from each airway image at different perspectives to calibrate key nodes related to the intubation structure. These key nodes typically include important geometric anchor positions that reflect the intubation path, such as laryngeal edge points, upper and lower glottal edge points, and tracheal ring distribution points. Determining the spatial mapping relationship between the key nodes related to the intubation structure in each airway image based on the calibration results can be achieved in the following manner: by combining the view parameters recorded during image acquisition (such as camera angle, camera pose, and projection matrix) with a multi-view geometric reconstruction method (such as a structure-from-motion based 3D reconstruction technique), a one-to-one spatial mapping relationship is established between the key nodes of the same intubation structure in the airway images at each perspective. That is, the relative positions of the key nodes related to the intubation structure in each airway image in the 3D airway structure are determined, and the spatial mapping relationship between the key nodes related to the intubation structure in each airway image can be obtained. In other embodiments, other methods can also be used for determination, which is not limited here.

[0037] In specific implementation, the multi-perspective association information of the intubation structure in the patient's airway can be determined by the spatial mapping relationship. This can be achieved in the following manner, namely: based on the spatial mapping relationship, the projection relationship of all key nodes related to the intubation structure at different perspectives can be mapped to the same three-dimensional reference coordinate system, and by matching the projection coincidence with the Euclidean distance between the corresponding points in space, the multi-perspective association information between the intubation structure points in the patient's airway is calculated and established to ensure that the same intubation structure is accurately identified at different image perspectives; the spatial topological structure relationship between the intubation structure in the patient's airway between multiple perspectives determined by the multi-perspective association information can be achieved in the following manner, namely: a graph neural network (such as a convolutional neural network) can be used. According to the multi-view association information, a spatial topological map of the intubation structure is constructed for all associated key nodes in three-dimensional coordinates. The map reflects the spatial connection mode, geometric distribution and structural continuity between the intubation structure points, thereby forming a spatial topological structure relationship between the intubation structure in the patient's airway between multiple perspectives, so as to provide a geometric structure prior for subsequent feature fusion (i.e., cross-view cross-fusion); other methods can also be used to implement it in other embodiments, which are not limited here.

[0038] It should be noted that the key nodes in this application represent the pixel positions of important anatomical reference points in the airway image that are closely related to the intubation path and airway patency, which can be used to anchor the local position of the internal structure of the patient's airway; the spatial mapping relationship in this application represents the corresponding position of the same intubation structure in the three-dimensional space in the airway images under different perspectives, which can be used to reflect the corresponding transformation relationship between the same key node in the airway image planes of each perspective; the multi-perspective association information in this application represents the spatial consistency relationship of the calibrated key points under different image perspectives, which is used to match and identify the projection of the intubation structure in multiple perspective images; the spatial topological structure relationship in this application represents the spatial connection order and structural organization of the key nodes in the patient's three-dimensional airway structure, which is an important basis for describing the geometric morphology of the patient's intubation channel.

[0039] In step 104 , each local feature and each global feature are cross-fused across views based on the spatial topological structure relationship and all recognition attentions to obtain a cross-fused feature vector of the patient's airway state.

[0040] In some embodiments, cross-view cross-fusion of local features and global features based on the spatial topological structure relationship and all recognition attentions to obtain a cross-fused feature vector of the patient's airway state can be achieved by the following steps: Perform weighted processing on each local feature and each global feature based on all recognition attention; Based on the spatial topological structure relationship, the same intubation structure features in the airway images from different perspectives are aligned and matched, and then a long-range dependency relationship is established between the intubation structure features obtained by the alignment and matching; The weighted local features and global features are interactively fused according to the long-distance dependency relationship to obtain a cross-fusion feature vector of the patient's airway state.

[0041] In specific implementation, weighted processing of each local feature and each global feature based on all recognition attention can be achieved in the following manner, namely: for the airway image at each perspective, weighted processing of the corresponding local features and global features is performed according to the corresponding recognition attention, that is, the feature vector in the airway image at the corresponding perspective is scaled element by element using recognition attention guidance, thereby enhancing the feature response of the high-attention area, suppressing the noise in the irrelevant area, and improving the discriminability of the feature; based on the spatial topological structure relationship, the same intubation structure features in the airway images at different perspectives are aligned and matched, and then a long-distance dependency relationship is established between the intubation structure features obtained by alignment and matching. This can be achieved in the following manner, namely: based on the spatial topological structure relationship, the feature points representing the same intubation structure part at different perspectives are aligned and matched. Specifically, the same intubation structure features in the airway images at different perspectives can be projected to the same unified coordinate system using a three-dimensional spatial mapping transformation, and geometric registration technology (such as ICP (Iterative Closest Point) algorithm is used to perform point cloud-level registration to ensure that the corresponding cannula structural features under different perspectives are accurately aligned. Then, a global contextual relationship of the same cannula structural features between all perspectives is modeled through a non-local attention mechanism (such as the self-attention module in the Transformer), and this global contextual relationship is used as a long-distance dependency relationship. This allows not only the perception of the local neighborhood in the subsequent interactive fusion process, but also the use of cross-image information to collaboratively enhance the structural expression. In other embodiments, other methods may also be used for determination, which is not limited here.

[0042] In specific implementation, the local features and global features after weighted processing are interactively fused according to the long-distance dependency relationship to obtain the cross-fusion feature vector of the patient's airway state, which can be achieved in the following manner, namely: through the multi-head attention mechanism, multiple feature sources (that is, the local features and global features after weighted processing) are information-coupled and selectively aggregated according to the long-distance dependency relationship. The fusion result is a set of cross-fusion feature vectors that comprehensively consider multi-scale features, structural topology information and attention weights, which are used to accurately characterize the patient's overall airway state, that is, the cross-fusion feature vector of the patient's airway state. In other embodiments, other methods can also be used for interactive fusion, which is not limited here.

[0043] It should be noted that the cross-view cross-fusion in this application refers to the process of jointly analyzing the same structural information in airway images from different perspectives. This process can improve the tolerance and correction capabilities of airway image recognition errors; the weighted processing in this application refers to the process of weighted scaling of each position or channel in the feature map according to recognition attention to highlight important structures.

[0044] In addition, it should be noted that the long-distance dependency in this application represents the information connection between intubation structures that are spatially non-adjacent but semantically corresponding under different perspectives. Therefore, it can be used to explore the complementary characteristics between different perspectives to help model the consistency of intubation structures under different perspectives; the cross-fusion feature vector in this application is a comprehensive expression of the airway obtained by fusion of multi-perspective structural information after weighting, alignment, and interaction. It is used to support high-accuracy difficult airway prediction, and the cross-fusion feature vector can provide a solid foundation for the final airway risk prediction.

[0045] In step 105, a risk prediction level of the patient's difficult airway is generated by combining a pre-built difficult airway prediction model with a cross-fusion feature vector of the patient's airway status.

[0046] In some embodiments, generating a patient's difficult airway risk prediction level by combining a pre-built difficult airway prediction model with the cross-fusion feature vector of the patient's airway status can be achieved by the following steps: Pre-build a difficult airway prediction model; Using the cross-fused feature vector of the patient's airway status as input to the difficult airway prediction model; The difficult airway prediction model outputs a risk prediction level of the patient's difficult airway.

[0047] It should be noted that the difficult airway prediction model in this application represents a deep learning model trained based on a large-scale sample of known cases, which is used to determine whether the patient's airway structure has a risk of difficult intubation; preferably, a large-scale labeled clinical data set (including multi-view airway images and risk labels of patients with difficult airway and non-difficult airway patients) can be used to pre-train a difficult airway prediction model through supervised learning. The difficult airway prediction model usually adopts a deep neural network structure (such as a multi-layer perceptron (MLP), a transformer or a graph neural network (GNN)), and its input is a cross-fused feature vector extracted from intubated patients at different perspectives, that is, an airway state expression that has integrated multi-scale features, recognition attention and spatial topological structure.

[0048] In a specific implementation, the output of the patient's difficult airway risk prediction level through the difficult airway prediction model can be achieved in the following manner, namely: the cross-fusion feature vector of the patient's airway state is input into the difficult airway prediction model, and the difficult airway prediction model first performs feature normalization and dimension mapping processing on the input (such as dimension compression through a fully connected layer), and then establishes a nonlinear relationship between features through the intermediate layer of the model, and extracts the feature combination representation most relevant to the difficult airway risk layer by layer. In some embodiments, this process may embed an attention mechanism, residual connection or dropout operation to improve the model's ability to distinguish different feature weights and generalization performance; After the features are processed by the deep network, the output layer outputs a risk prediction level through a softmax or sigmoid function, which is usually expressed as a probability value or a multi-classification label to indicate that the patient belongs to the risk prediction level classification of difficult airway, potential risk airway or normal airway; wherein, the risk prediction level can be set as a binary classification (difficult / non-difficult) or a multi-level classification (such as low risk, medium risk, high risk), or a continuous value risk score can be provided in the form of a probability score to facilitate doctor evaluation for subsequent intervention plan formulation and clinical record keeping. Other methods can also be used in other embodiments for implementation, which is not specifically limited here.

[0049] It should be noted that the risk prediction level in this application represents the predicted result of the difficult airway prediction model in assessing the difficulty of intubation of patients. It can be used to guide clinical pretreatment plans, provide clinical staff with clear and actionable risk warnings, and assist in formulating airway intubation strategies.

[0050] In some embodiments, after the intelligent airway management system obtains the patient's risk prediction level of difficult airway, the patient's risk prediction level of difficult airway can be linked with the patient's basic information to prompt individualized suggestions and convert the risk prediction results into executable medical behaviors. For example, it can combine the patient's age, BMI (Body Mass Index), previous intubation history and other information to give targeted suggestions (such as bronchoscopy first) to achieve truly personalized and intelligent airway management.

[0051] In addition, in another aspect of the present application, in some embodiments, the present application provides an intelligent airway management system, the system further comprising a difficult airway prediction unit, referring to Figure 4 , which is a schematic structural diagram of a difficult airway prediction unit according to some embodiments of the present application. The difficult airway prediction unit 400 includes: an acquisition module 401, a processing module 402, and an execution module 403, which are described as follows: Acquisition module 401, in this application, acquisition module 401 is mainly used to acquire airway images of the patient at different viewing angles; Processing module 402, in this application, is mainly used to screen out effective areas of the intubation structure in the airway image at different viewing angles based on the pixel contribution when identifying the intubation structure in each airway image, and determine the recognition attention of each airway image when performing difficult airway prediction based on each effective area and the patient's cervical spine mobility; The processing module 402 of the present application is further configured to perform multi-scale dilated convolution on each airway image to extract local features and global features of each airway image, and determine the spatial topological structure relationship of the intubation structure in the patient's airway between multiple perspectives based on the spatial mapping relationship between key nodes related to the intubation structure in each airway image; The processing module 402 in the present application is further configured to perform cross-view cross-fusion of each local feature and each global feature based on the spatial topological structure relationship and all recognition attentions to obtain a cross-fusion feature vector of the patient's airway state; The execution module 403 in this application is mainly used to generate a risk prediction level of the patient's difficult airway by combining a pre-built difficult airway prediction model with the cross-fusion feature vector of the patient's airway status.

[0052] The above describes in detail the examples of the intelligent airway management system and method provided by the embodiments of the present application. It can be understood that in order to realize the above functions, the corresponding device includes hardware structures and / or software modules corresponding to the execution of each function. It should be easily appreciated by those skilled in the art that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0053] In some embodiments, the present application also provides a computer device, which includes a memory and a processor, the memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the above-mentioned difficult airway prediction method.

[0054] In some embodiments, reference Figure 5 The dotted line in the figure indicates that the unit or module is optional. The figure is a schematic diagram of the structure of the computer device that implements the difficult airway prediction method of the present application. The difficult airway prediction method in the above embodiment can be Figure 5The computer device 500 is implemented as shown, and the computer device 500 includes at least one processor 501, a memory 502 and at least one communication unit 505. The computer device 500 can be a terminal device, a server or a chip.

[0055] The processor 501 may be a general-purpose processor or a special-purpose processor. For example, the processor 501 may be a central processing unit (CPU), which may be used to control the computer device 500, execute software programs, and process data from the software programs. The computer device 500 may also include a communication unit 505 for inputting (receiving) and outputting (transmitting) signals.

[0056] For example, the computer device 500 may be a chip, the communication unit 505 may be an input and / or output circuit of the chip, or the communication unit 505 may be a communication interface of the chip, and the chip may be a component of a terminal device, a network device, or other device.

[0057] For another example, the computer device 500 may be a terminal device or a server, and the communication unit 505 may be a transceiver of the terminal device or the server, or the communication unit 505 may be a transceiver circuit of the terminal device or the server.

[0058] The computer device 500 may include one or more memories 502, on which a program 504 is stored. The program 504 can be executed by the processor 501 to generate instructions 503, so that the processor 501 executes the method described in the above method embodiment according to the instructions 503. Optionally, data (such as a target audit model) can also be stored in the memory 502. Optionally, the processor 501 can also read data stored in the memory 502. The data can be stored at the same storage address as the program 504, or at a different storage address from the program 504.

[0059] The processor 501 and the memory 502 may be provided separately or integrated together, for example, integrated on a system on chip (SOC) of a terminal device.

[0060] It should be understood that each step of the above method embodiment can be completed by a hardware-based logic circuit or software-based instructions in the processor 501. The processor 501 can be a CPU, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, such as discrete gates, transistor logic devices, or discrete hardware components.

[0061] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] For example, in some embodiments, the present application also provides a computer-readable storage medium, in which instructions or codes are stored. When the instructions or codes are run on a computer, the computer implements the above-mentioned difficult airway prediction method when executing.

[0063] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0064] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A difficult airway prediction method for an intelligent airway management system to predict a patient's difficult airway, characterized in that: The method comprises the following steps: Acquire the patient's airway images at different viewing angles; The effective areas of the intubation structure in each airway image at different viewing angles are screened by the pixel contribution when identifying the intubation structure in each airway image. Based on each effective area and the patient's cervical spine mobility, the recognition attention of each airway image when predicting a difficult airway is determined. Multi-scale dilated convolution is performed on each airway image to extract local and global features of each airway image. The spatial topological relationship of the intubation structure in the patient's airway across multiple viewpoints is determined based on the spatial mapping relationship between key nodes related to the intubation structure in each airway image. Based on the spatial topological structure relationship and all recognition attentions, each local feature and each global feature are cross-fused across views to obtain a cross-fused feature vector of the patient's airway state; The risk prediction level of the patient's difficult airway is generated by combining a pre-built difficult airway prediction model with a cross-fusion feature vector of the patient's airway status.

2. The method according to claim 1, wherein The effective areas of the cannula structure in the airway image at different viewing angles are screened out by the pixel contribution when identifying the cannula structure in each airway image. Specifically, they include: An airway image under a viewing angle is selected as a selected airway image; dividing the selected airway image into a plurality of sub-regions; Determine the local gradient change of pixels in each sub-region, and then determine the pixel contribution of each sub-region to the identification of the cannula structure; Filter out all sub-regions in the selected airway image whose pixel contributions are greater than a preset threshold, and form the valid region of the intubation structure in the airway image under the corresponding viewing angle; Continue to screen out the valid areas of the cannula structure in the airway images under the remaining viewing angles.

3. The method according to claim 1, wherein The specific recognition attention when determining each airway image for difficult airway prediction based on each effective area and the patient's cervical spine mobility includes: Determine the patient's cervical spine mobility based on the patient's cervical spine motion imaging; adjusting the attention weights of the airway images at different viewing angles according to the cervical spine mobility; The recognition attention of each airway image when performing difficult airway prediction is determined by adjusting the various attention weights and effective areas.

4. The method according to claim 1, wherein Perform multi-scale dilated convolution on each airway image to extract the local and global features of each airway image, including: Perform multi-scale dilation convolution on each airway image according to multiple dilation rates to obtain feature maps of each airway image at different scales; Local and global features of each airway image are extracted based on feature maps at different scales.

5. The method according to claim 1, wherein The spatial topological structure relationship of the cannula structure in the patient's airway between multiple perspectives is determined based on the spatial mapping relationship between key nodes related to the cannula structure in each airway image, specifically including: Demarcate key nodes related to the intubation structure in each airway image; Determine the spatial mapping relationship between key nodes related to the intubation structure in each airway image based on the calibration results; Determine multi-view correlation information of the cannula structure in the patient's airway through the spatial mapping relationship; The spatial topological structure relationship of the intubation structure in the patient's airway between the multiple perspectives is determined based on the multi-perspective association information.

6. The method according to claim 1, wherein Based on the spatial topological structure relationship and all recognition attention, each local feature and each global feature is cross-fused across views to obtain a cross-fused feature vector of the patient's airway state, specifically including: Perform weighted processing on each local feature and each global feature based on all recognition attention; Based on the spatial topological structure relationship, the same intubation structure features in the airway images from different perspectives are aligned and matched, and then a long-range dependency relationship is established between the intubation structure features obtained by the alignment and matching; The weighted local features and global features are interactively fused according to the long-distance dependency relationship to obtain a cross-fusion feature vector of the patient's airway state.

7. The method according to claim 1, wherein Generating a risk prediction level of difficult airway for a patient by combining a pre-built difficult airway prediction model with a cross-fusion feature vector of the patient's airway status specifically includes: Pre-build a difficult airway prediction model; Using the cross-fused feature vector of the patient's airway status as input to the difficult airway prediction model; The difficult airway prediction model outputs a risk prediction level of the patient's difficult airway.

8. An intelligent airway management system, comprising a difficult airway prediction unit, characterized in that: The difficult airway prediction unit comprises: An acquisition module, used to acquire airway images of the patient at different viewing angles; A processing module is used to screen out effective areas of the intubation structure in the airway image at different viewing angles based on the pixel contribution when identifying the intubation structure in each airway image, and determine the recognition attention of each airway image when performing difficult airway prediction based on each effective area and the patient's cervical spine mobility; The processing module is further configured to perform multi-scale dilated convolution on each airway image to extract local and global features of each airway image, and determine the spatial topological relationship of the intubation structure in the patient's airway between multiple perspectives based on the spatial mapping relationship between key nodes related to the intubation structure in each airway image; The processing module is further configured to perform cross-view cross-fusion of each local feature and each global feature based on the spatial topological structure relationship and all recognition attentions to obtain a cross-fusion feature vector of the patient's airway state; An execution module is used to generate a risk prediction level of the patient's difficult airway by combining a pre-built difficult airway prediction model with a cross-fusion feature vector of the patient's airway status.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the difficult airway prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions or codes, and when the instructions or codes are executed on a computer, the computer implements the difficult airway prediction method according to any one of claims 1 to 7.

Images