A difficult airway assessment system based on double position cervical x-ray

Abstract

The present application relates to a kind of difficult airway evaluation system based on double body position cervical X-ray, the system includes X-ray data acquisition unit, feature fusion extraction unit and prediction unit;The X-ray data acquisition unit is used to obtain the cervical X-ray data of two different body positions of the object to be evaluated;The feature fusion extraction unit is used to carry out multi-level feature extraction and fusion based on the cervical X-ray data corresponding to different body positions, obtain double body position fusion feature vector;The prediction unit is used to carry out classification prediction based on the double body position fusion feature vector, obtain difficult airway evaluation result.The present application solves the problem that single body position data is used when image data is used for difficult airway evaluation in the prior art, which cannot accurately and effectively evaluate difficult airway, and is easy to cause airway misjudgment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of clinical medicine, in particular to a difficult airway evaluation system based on double-position cervical X-ray. BACKGROUND

[0002] Tracheal intubation is a key technology in anesthesiology, emergency medicine and critical care medicine. Poor airway management can lead to body injury, brain hypoxia, and even death. Difficult airway refers to the unexpected or unexpected airway management difficulty or failure encountered by clinicians with clinical anesthesia training in clinical practice. Preoperative airway evaluation is crucial, and screening difficult airway patients can help reduce the risk of difficult ventilation and even asphyxiation death. At present, the evaluation of difficult airway mainly relies on the experience of anesthesiologists or the combination of X-ray, ultrasound, magnetic resonance imaging and other single imaging data, which is highly subjective and has low accuracy, and is prone to cause airway state misjudgment. SUMMARY

[0003] In view of the above analysis, the present application aims to provide a difficult airway evaluation system based on double-position cervical X-ray; solve the limitation of single position image in the prior art when using image data to evaluate difficult airway, thereby further improving the accuracy of difficult airway evaluation.

[0004] The purpose of the present application is mainly realized by the following technical solutions:

[0005] On the one hand, the present application discloses a difficult airway evaluation system based on double-position cervical X-ray, which comprises an X-ray data acquisition unit, a feature fusion extraction unit and a prediction unit;

[0006] The X-ray data acquisition unit is used to acquire cervical X-ray data of a subject in two different positions;

[0007] The feature fusion extraction unit is used to perform multi-level feature extraction and fusion based on the cervical X-ray data corresponding to different positions to obtain a double-position fusion feature vector;

[0008] The prediction unit is used to perform classification prediction based on the double-position fusion feature vector to obtain a difficult airway evaluation result.

[0009] Further, the cervical X-ray data in the two different positions includes neutral position cervical X-ray data and hyperextension position cervical X-ray data.

[0010] Further, the feature fusion extraction unit comprises n feature extraction and fusion modules, one pooling module and a vector fusion module arranged in sequence;

[0011] The feature extraction and fusion module includes two parallelly arranged feature extraction layers and a feature map fusion layer corresponding to each feature extraction layer; the feature map fusion layer is used for performing feature map fusion on the feature map output by the corresponding feature extraction layer and the feature map output by the other feature extraction layer, and inputting the corresponding feature extraction layer of the next layer for size compression and feature extraction;

[0012] The pooling module includes two parallelly arranged pooling layers, which are used for performing pooling operation on the outputs of the two feature map fusion layers in the last feature extraction and fusion module;

[0013] The vector fusion module is used for splicing the features output by the two pooling layers, and performing feature mapping through a fully connected layer to obtain the double-body-position fusion feature vector.

[0014] Further, the feature extraction layers in the n feature extraction and fusion modules are constructed based on n stages of a ResNet18 model;

[0015] The first feature extraction and fusion module performs size compression and feature extraction on the cervical X-ray data of two different body positions through two parallelly arranged feature extraction layers, and inputs two feature map fusion layers for feature map fusion.

[0016] The second to nth feature extraction and fusion modules receive the outputs of the corresponding feature map fusion layers of the previous layer through two parallelly arranged feature extraction layers, and perform size compression and feature extraction and feature map fusion.

[0017] Further, the feature extraction and fusion module includes a depth separable convolution layer, a non-local neural network layer, a multi-resolution pooling layer and a convolution layer arranged in sequence;

[0018] The depth separable convolution layer takes the two received feature maps as the input of the corresponding channels, and fuses the spatial information of the two feature maps based on different channels;

[0019] The non-local neural network is used for capturing long-distance dependence based on the fused feature data, and inputting the multi-resolution pooling layer to extract feature representations of different scales;

[0020] The convolution layer fuses the feature representations of different scales into a single-dimensional vector to obtain the feature representation after feature map fusion.

[0021] Further, the feature fusion extraction unit is trained based on a pre-constructed training sample set; the training sample set includes double-body-position cervical X-ray data of a plurality of difficult airway patients and healthy subjects, and difficult airway or non-difficult airway labels corresponding to each X-ray data;

[0022] In the training process, iterative optimization is performed based on the pre-constructed composite loss and compactness loss, denoted as:

[0023] ;

[0024] wherein, and are balance hyperparameters, obtained through training, is a composite loss function; is a compactness loss function.

[0025] Further, the composite loss is constructed from a center loss and a discrete loss, denoted as:

[0026] ;

[0027] wherein, is a center loss, used to force the corresponding dual position fusion feature vectors of healthy subjects to gather to a shared center point in the feature space, is a discrete loss, used to make the corresponding dual position fusion feature vectors of difficult airway samples diverge in the non-center point region of the feature space, is a dual position fusion feature vector corresponding to the i-th sample data, is a label corresponding to the i-th sample data, denotes a non-difficult airway label, denotes a difficult airway label, c is a shared center point, and m is a preset margin, and are respectively and the number of samples.

[0028] Further, the compactness loss is used to minimize the standard deviation of the distance of the dual position fusion feature vector to the shared center point, so as to promote the compactness of the intra-class features;

[0029] The compactness loss is denoted as:

[0030] ;

[0031] wherein, is a compactness loss, is a standard deviation of the distance of the dual position fusion feature vector to the center point, k is 0 or 1, is an average distance of the dual position fusion feature vector to the center point. Further, in constructing the training sample set, sample augmentation is also performed on the difficult airway sample data by the following method:

[0032] Learning a latent distribution of difficult airway samples in a feature space based on a variational autoencoder, sampling new latent vectors based on the latent distribution, and decoding into pseudo-difficult airway sample data;

[0033] Constructing a discriminative network to screen the pseudo-difficult airway sample data, retaining pseudo-difficult airway sample data with a confidence greater than a preset threshold as difficult airway sample data, labeling difficult airway labels, and adding the difficult airway sample data to the training sample set for model training.

[0034] Further, in constructing the training sample set, data augmentation is further included by random rotation, horizontal and vertical flipping, random cropping, and random adjustment of brightness, contrast, and saturation.

[0035] The present application can achieve at least the following beneficial effects:

[0036] 1. The present application uses neutral position cervical X-ray data and hyperextension position cervical X-ray data as two different body position cervical X-ray data, and solves the problem of missed detection when single body position image data is used for difficult airway evaluation, effectively reduces the incidence of unexpected difficult airway, and further ensures the safety of patients during the perioperative period.

[0037] 2. In the present application, feature fusion is used in the middle layer of the model when double body position cervical X-ray data is fused, and different channel spatial information is combined for fusion. In the output layer of the model, vector fusion is used, and the extracted feature vectors are spliced and transmitted to the full connection layer for classification. Through this multi-layer hybrid fusion method, the spatial information is retained, and the expression ability of the features is enhanced through splicing in the final stage. By using different fusion strategies at different levels, the advantages of both are fully utilized, and the accuracy of feature extraction and difficult airway evaluation is improved.

[0038] 3. In the present application, the samples of non-difficult airway subjects are forced to gather to a certain center point in the feature space during model training, and feature consistency is achieved by minimizing the distance between them and the center point; at the same time, the difficult samples are punished for their behavior of approaching the center point, so that they naturally disperse in other areas of the feature space. Through this design, the model is encouraged to automatically learn the typical normal discriminative features of non-difficult airway samples, and make more robust judgments when encountering difficult airway samples with deviated features, thereby improving the recognition ability of the model in the identification and evaluation of difficult airway samples with difficult to define abnormal morphology.

[0039] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and obtained by the structure particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0041] Figure 1 Schematic diagram of the structure of a difficult airway assessment system based on dual-position cervical spine X-rays according to an embodiment of the present invention;

[0042] Figure 2 Schematic diagram of clustering on the feature layer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0044] A difficult airway assessment system based on dual-position cervical spine X-rays according to an embodiment of the present invention is provided. Figure 1 As shown, the system includes an X-ray data acquisition unit, a feature fusion extraction unit and a prediction unit;

[0045] The X-ray data acquisition unit is used to acquire cervical spine X-ray data of two different body positions of the subject to be evaluated;

[0046] The feature fusion extraction unit is used to perform multi-level feature extraction and fusion based on the cervical spine X-ray data corresponding to different body positions to obtain a dual-position fusion feature vector;

[0047] The prediction unit is used to perform classification prediction based on the dual-position fusion feature vector to obtain a difficult airway assessment result.

[0048] Specifically, in actual clinical practice, it is found that the causes of difficult airway in patients are quite diverse, involving multiple different structural abnormalities, such as limited cervical spine mobility, oropharyngeal structural abnormalities, etc., and their characteristics show strong discreteness and heterogeneity in X-ray data, and the performance is also different in cervical spine X-ray data in different body positions; therefore, using imaging data of a single body position for difficult airway assessment is prone to missed detection; in order to solve this problem, the X-ray data acquisition unit of this embodiment uses commonly used clinical X-ray equipment to collect neutral cervical spine X-ray data and hyperextended cervical spine X-ray data as cervical spine X-ray data in two different body positions, and uses the data for difficult airway assessment after data fusion.

[0049] After obtaining the X-ray data, image preprocessing such as size adjustment, noise reduction, grayscale conversion, feature binarization and normalization is first performed so that the data can be input into the feature fusion extraction unit for feature extraction and difficult airway assessment.

[0050] Preferably, the feature extraction fusion unit constructed by the present application comprises n feature extraction fusion modules, 1 pooling module and a vector fusion module arranged in sequence;

[0051] The feature extraction fusion module comprises two feature extraction layers arranged in parallel and two feature map fusion layers corresponding to the two feature extraction layers respectively; the feature map fusion layer is used for performing feature map fusion on the feature map output by the corresponding feature extraction layer and the feature map output by the other feature extraction layer, and inputting the size compression and feature extraction of the corresponding feature extraction layer of the next layer;

[0052] The pooling module comprises two pooling layers arranged in parallel, which are used for performing pooling operation on the outputs of the two feature map fusion layers in the last feature extraction fusion module respectively;

[0053] The vector fusion module is used for splicing the features output by the two pooling layers, and performing feature mapping through a fully connected layer to obtain the double-body-position fusion feature vector.

[0054] In particular, the feature extraction layers in the n feature extraction fusion modules are constructed based on n stages of ResNet18 model;

[0055] The first feature extraction fusion module performs size compression and feature extraction on the cervical X-ray data of two different body positions through two feature extraction layers arranged in parallel, and inputs two feature map fusion layers for feature map fusion;

[0056] The 2nd to n-th feature extraction fusion modules receive the outputs of the corresponding feature map fusion layers of the previous layer through two feature extraction layers arranged in parallel, and perform size compression, feature extraction and feature map fusion.

[0057] The feature extraction fusion module comprises a depth separable convolution layer, a non-local neural network layer, a multi-resolution pooling layer and a convolution layer arranged in sequence;

[0058] The depth separable convolution layer takes the two received feature maps as the input of the corresponding channels, and fuses the spatial information of the two feature maps based on different channels; that is, the two channels of the depth separable convolution layer correspond to two feature maps respectively, and the received feature maps are convolved through different convolution kernels.

[0059] The non-local neural network is used for capturing long-distance dependence based on the fused feature data, and inputting the multi-resolution pooling layer to extract feature representations of different scales;

[0060] The convolution layer fuses the feature representations of different scales into a single-dimensional vector to obtain the feature representation after feature map fusion.

[0061] That is, the feature extraction and fusion module of the present embodiment integrates features from two inputs using a dual position model to take advantage of the complementary information from dual position cervical X-ray data.

[0062] In terms of feature fusion strategy, the present embodiment combines a feature map fusion layer and a vector fusion layer for feature fusion, i.e., a HybridFusion method that combines the advantages of MapFusion and VecFusion, wherein:

[0063] In the MapFusion method, the images corresponding to the two body positions are subjected to feature extraction through independent ResNet18 networks, and the output of each network is regarded as the input of the corresponding two channels in the CNN. Depthwise Separable Convolutions are used to separate spatial and channel information, and Non-local Neural Network is used to capture long-range dependencies, and then Multi-Resolution Pooling is used to extract feature representations of different scales, and finally a convolution layer is used to fuse the single-dimensional vectors to generate the final feature representation. Unlike VecFusion, MapFusion retains spatial information, so it is more suitable for tasks that require spatial details. The advantage is that it processes and fuses multi-channel features through convolution layers to combine spatial context information.

[0064] In the VecFusion method, the images corresponding to the two body positions are subjected to feature extraction through independent ResNet18 networks to obtain feature representations of each channel. Then, after the pooling layer in the last layer, the features extracted from the two channels are spliced. This spliced feature vector is passed to a Fully Connected Layer for further transformation to obtain the final feature representation. This method is simple and direct, and can effectively combine the features of the two channels, but does not retain spatial context information and mainly relies on the splicing of feature vectors to express information.

[0065] In this embodiment, the HybridFusion method is used to combine VecFusion and MapFusion for feature fusion. In this method, MapFusion is used in the middle layer of ResNet18 to combine the spatial information of different channels for fusion. In the output layer of ResNet, VecFusion is used to splice the extracted feature vectors and then pass them to the fully connected layer for classification. Through this multi-layer fusion method, HybridFusion not only retains the spatial information, but also enhances the expression ability of features through splicing in the final stage. By using different fusion strategies at different levels, the advantages of both are fully utilized.

[0066] Further, the feature fusion extraction unit is trained based on a pre-constructed training sample set; the training sample set includes double-position cervical X-ray data of a plurality of difficult airway patients and healthy subjects, and a difficult airway or non-difficult airway label corresponding to each X-ray data;

[0067] When constructing the training sample set, the following method is also used to perform sample augmentation on the difficult airway sample data:

[0068] Based on the variational autoencoder, the latent distribution of the difficult airway sample in the feature space is learned, a new latent vector is sampled based on the latent distribution, and the pseudo-difficult airway sample data is decoded;

[0069] A discriminative network is constructed to filter the pseudo-difficult airway sample data, and the pseudo-difficult airway sample data with a confidence greater than a preset threshold is retained as the difficult airway sample data, labeled with a difficult airway label and added to the training sample set for model training.

[0070] In this embodiment, after obtaining the sample data, all original cervical neutral position and hyperextension position X-ray data (1024x1024 pixels) are subjected to standardization preprocessing. To meet the general data input dimension requirements of the convolutional neural network (such as ResNet18), the X-ray data is uniformly scaled to 224x224 pixels. This standardization processing has double advantages: first, the unified image size facilitates the integration of pre-trained models and fully utilizes the transfer learning efficiency; second, it ensures the consistency of data dimensions during model training and evaluation, improving the stability of the algorithm and the comparability of the results.

[0071] In addition, in the clinical data set, the number of patients with non-difficult airways is much larger than that of patients with difficult airways, and directly training a deep learning model is susceptible to class imbalance, which can cause bias in model training, resulting in limited recognition ability of the model for the minority class (i.e., difficult cases), and the network tends to predict the majority class, thereby weakening its ability to accurately identify the minority class. To alleviate this problem, the embodiment introduces a generative model for data augmentation to expand the feature diversity of difficult samples in the training set, thereby improving the generalization ability and sensitivity to abnormal patterns of the model.

[0072] Specifically, the embodiment models the high-dimensional feature representation of difficult airway samples based on the Variational Autoencoder (VAE) framework. By learning the latent distribution of difficult airway samples in the feature space, new latent vectors are sampled from the distribution and decoded into images or high-level features of "pseudo-difficult airway samples". Although these synthetic samples do not directly correspond to a real patient, they have typical difficult airway features in a statistical sense, effectively supplementing the distribution range of real difficult airway samples. Generated samples can participate in training together with real samples, and can also be used to optimize the "repulsion mechanism" of the difficult group in the feature clustering loss.

[0073] In addition, to ensure the medical reliability of the generated samples, the embodiment also introduces a discriminator network to perform quality screening on the generated samples, and only keeps the difficult sample features with high confidence judged by the discriminator for subsequent training. This strategy combining generation and discrimination not only enhances the diversity of training data, but also improves the recognition robustness of the model for edge cases. The data augmentation strategy based on the generative model shows significant advantages in tasks involving extremely imbalanced samples, diverse but rare abnormal patterns of medical images, and provides more abundant and reliable training support for automatic recognition of difficult airways.

[0074] Further, when constructing the training sample set, data augmentation is also needed to prevent overfitting, including data augmentation through random rotation, horizontal and vertical flipping, random cropping, and random adjustment of brightness, contrast, and saturation. Data augmentation enables the model to resist changes in image orientation, lighting conditions, and other disturbances, thereby reducing the likelihood of overfitting and improving the model's performance on real-world data. Specifically, the embodiment applies random scaling cropping with a scale range of (0.8, 1.0). Random horizontal and vertical flipping is performed, with the image rotation angle being up to 15 degrees. Color adjustment is performed, with the changes in brightness, contrast, saturation, and hue being set to 0.2. In addition, random affine transformation is also used, with the maximum rotation angle being 15 degrees and the shear angle being 10 degrees. By introducing data transformed through the above methods, the model can be exposed to a variety of transformed images, thereby enhancing its generalization ability for unseen data.

[0075] Further, in the training process, the embodiment optimizes the model based on the feature layer clustering classification method. In view of the significant imbalance of the category distribution in the application, especially the number of non-difficult airway cases is much larger than that of difficult airway cases, the embodiment not only solves this problem through the sample expansion strategy, but also designs a feature space clustering-based classification mechanism to replace or enhance the traditional cross-entropy loss function. Through the feature space clustering-based classification mechanism, the problem of diversity of difficult airway features can also be solved, that is, in actual clinical practice, patients with non-difficult airway often show highly consistent anatomical features, so their extracted image features in the high-dimensional feature space are usually naturally clustered in a stable area. The causes of difficult airway patients are diverse, which may involve multiple structural abnormalities, such as limited cervical spine mobility and oropharyngeal structural abnormalities, and therefore their features in the space show strong discreteness and heterogeneity.

[0076] Based on this observation, the embodiment does not attempt to cluster normal and difficult samples into two tight groups during the training process, but uses a feature space clustering-based classification mechanism to iteratively optimize the center loss to force the normal samples of healthy subjects to be clustered around a shared center point in the feature space, achieving feature consistency by minimizing the distance between the center point and the center point. At the same time, the difficult samples are not forced to cluster, but are punished by the dispersion loss for their behavior of approaching the center point, so that they naturally disperse in other areas of the feature space. The design of the embodiment encourages the model to automatically learn the typical discriminative features of non-difficult airway in X-ray data, and makes more robust judgments when encountering difficult airway samples with deviating features. A schematic diagram of clustering at the feature layer is shown in FIG. 1. Figure 2

[0077] This feature space-based asymmetric clustering mechanism fully considers the heterogeneity of difficult laryngoscopy in real clinical environment, and can also reduce the training bias caused by the imbalance of sample quantity. On multiple randomly divided validation sets, this method stably improves the recognition ability of the model, indicating that it has broad potential for application in medical image classification, especially in tasks where abnormal morphology is difficult to define.

[0078] In particular, the embodiment uses a pre-constructed composite loss function combined with a compact loss to implement the feature layer clustering-based classification method. The total loss function during model training is represented as:

[0079] ;

[0080] where L is the total loss function, and are balance hyperparameters, which are obtained through training, ​a composite loss function; a compactness loss function.

[0081] The composite loss is constructed by a center loss and a dispersion loss, denoted as:

[0082] ;

[0083] wherein, is a center loss, for forcing the double-view fusion feature vectors corresponding to healthy subjects to be gathered to a shared center point in the feature space, is a dispersion loss, for making the double-view fusion feature vectors corresponding to difficult airway samples to be diverged in the non-center point region of the feature space, is the double-view fusion feature vector corresponding to the i-th sample data, is the label corresponding to the i-th sample data, denotes a non-difficult airway label, denotes a difficult airway label, c is the shared center point, the shared center point is updated based on gradient optimization after random initialization as the network learns, and m is a preset margin, and are respectively and the number of samples.

[0084] The compactness loss is used to minimize the standard deviation of the distance of the double-view fusion feature vector to the shared center point, so as to promote the compactness of the intra-class features;

[0085] The compactness loss is denoted as:

[0086] ;

[0087] wherein, is a compactness loss, is the standard deviation of the distance of the double-view fusion feature vector to the shared center point, k takes 0 or 1, is the average distance of the double-view fusion feature vector to the shared center point.

[0088] That is, the classification method of the present application is different from the traditional classification method, which usually uses a fully connected layer to map the feature map to a two-dimensional vector for classification. The present embodiment is to cluster directly on the feature layer. Through the composite loss, the features of the cases that are easy to perform laryngoscopy (non-difficult airway samples, label = 0) are caused to converge to a center point, while the features of the cases that are difficult to perform laryngoscopy (difficult airway samples, label = 1) are forced to separate from it. In addition, through the compactness loss, it is ensured that within each class, the standard deviation of the distance of the feature vector to the center point remains small, thereby promoting the compactness of the intra-class features. Through the balance of the two loss functions, the requirement for clear separation between classes and the requirement for close clustering of intra-class features are balanced, thereby improving the classification ability of the model.

[0089] Further, in order to verify the prediction ability of the evaluation system of the present application, experiments were carried out on an NVIDIA A100 GPU with 24GB of memory, the main programming language was Python, and PyTorch was used for data processing and deep learning model construction.

[0090] Although the X-ray image is usually a grayscale image, in order to be consistent with the images used in the pre-training stage of the feature extraction module, all images are read in RGB three-channel format. The RGB values of each image are normalized so that their mean and variance match those of the pre-training dataset.

[0091] In order to evaluate the robustness and generalization ability of the model, a strict five-fold cross-validation strategy was adopted. In each fold, the dataset was randomly divided into a training set and a test set in a 9:1 ratio, and this process was repeated five times to generate five different training-test divisions. Each division is used independently to train and evaluate the network, and the final performance index is obtained by averaging the results of all folds. This method reduces the influence of a single division and ensures the reliable performance of the model on diverse and unseen datasets.

[0092] During training, the model ran for 150 cycles, and the training dataset was randomly sampled. The Adam optimizer was used with an initial learning rate of 0.001, and the OneCycleLR learning rate scheduler was used to adjust the learning rate during training. Experiments show that the difficult airway evaluation system constructed by the present application has good accuracy and generalization ability when evaluating difficult airways based on X-ray data, and realizes the function of automatic identification of difficult airways.

[0093] In summary, the difficult airway evaluation system based on double-position cervical X-ray of the application uses neutral position cervical X-ray data and hyperextension position cervical X-ray data as two different position cervical X-ray data, and the data fusion is used for difficult airway evaluation, solves the problem of missed detection when single position image data is used for difficult airway evaluation, effectively reduces the incidence of unexpected difficult airway, and further guarantees the safety of patients during the perioperative period. And when the double-position cervical X-ray data is fused, the feature map fusion is used in the middle layer of the model, and the spatial information of different channels is combined for fusion. And in the output layer of the model, the vector fusion is used, the extracted feature vectors are spliced and transmitted to the full connection layer for classification. Through this multi-layer hybrid fusion method, the spatial information is retained, and the expression ability of the features can be enhanced through splicing in the final stage. By using different fusion strategies at different levels, the advantages of both are fully utilized, and the accuracy of feature extraction and difficult airway evaluation is improved. During model training, normal samples of healthy subjects are forced to gather to a certain center point in the feature space, and the feature consistency is realized by minimizing the distance between the center point; at the same time, the difficult samples are punished for being close to the center point, so that they naturally disperse in other areas of the feature space. Through this design, the model is encouraged to automatically learn the typical normal discriminant features of non-difficult airway samples, and make more stable judgments when encountering samples with deviated features, and improve the recognition ability of the model in the difficult airway recognition evaluation with difficult to define abnormal morphology.

[0094] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer readable storage medium. Among them, the computer readable storage medium is a disk, an optical disk, a read-only memory or a random access memory, etc.

[0095] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A difficult airway assessment system based on dual-position cervical spine X-rays, characterized by: The system includes an X-ray data acquisition unit, a feature fusion extraction unit and a prediction unit; The X-ray data acquisition unit is used to acquire cervical spine X-ray data of two different body positions of the subject to be evaluated; the cervical spine X-ray data of the two different body positions include neutral position cervical spine X-ray data and hyperextension position cervical spine X-ray data; The feature fusion extraction unit is used to perform multi-level feature extraction and fusion based on the cervical spine X-ray data corresponding to different body positions to obtain a dual-position fusion feature vector; The feature fusion extraction unit is constructed based on the ResNet18 model. When performing feature fusion, feature map fusion is used in the middle layer of the ResNet18 model to combine the spatial information of different channels for fusion; In the output layer of ResNet18, vector fusion is used to concatenate the extracted feature vectors; The feature fusion extraction unit is trained based on a pre-constructed training sample set; the training sample set includes dual-position cervical spine X-ray data of multiple difficult airway patients and healthy subjects, as well as difficult airway or non-difficult airway labels corresponding to each X-ray data; During training, iterative optimization is performed based on the pre-built composite loss and compactness loss, which is expressed as: ； in, and To balance the hyperparameters, we obtain is the composite loss function; is the compactness loss function; The composite loss is constructed by the central loss and the discrete loss, and is expressed as: ； in, is the center loss, which is used to force the dual-position fusion feature vectors corresponding to healthy subjects to converge to the shared center point in the feature space. is a discrete loss, which is used to make the dual-position fusion feature vector corresponding to the difficult airway sample diverge in the non-center point area of ​​the feature space. is the dual-position fusion feature vector corresponding to the i-th sample data, is the label corresponding to the i-th sample data, Indicates non-difficult airway label, Indicates the difficult airway label, c is the shared center point, m is the preset margin, and They are and The number of samples; The prediction unit is used to perform classification prediction based on the dual-position fusion feature vector to obtain a difficult airway assessment result.

2. The difficult airway assessment system based on dual-position cervical spine X-ray according to claim 1, characterized in that: The feature fusion extraction unit includes n feature extraction fusion modules, a pooling module and a vector fusion module arranged in sequence; Each feature extraction and fusion module includes two parallel feature extraction layers and two feature map fusion layers corresponding to the two feature extraction layers; the feature map fusion layer is used to fuse the feature map output by the corresponding feature extraction layer and the feature map output by another feature extraction layer, and input the feature map into the feature extraction layer corresponding to the next layer for size compression and feature extraction; The pooling module includes two pooling layers set in parallel, which are used to perform pooling operations on the outputs of the two feature map fusion layers in the last feature extraction and fusion module respectively; The vector fusion module is used to splice the features output by the two pooling layers and perform feature mapping through a fully connected layer to obtain the dual-position fusion feature vector.

3. The difficult airway assessment system based on dual-position cervical spine X-ray according to claim 2, characterized in that: The feature extraction layers in the n feature extraction and fusion modules are constructed based on the n stages of the ResNet18 model; The first feature extraction and fusion module uses two parallel feature extraction layers to perform size compression and feature extraction on the cervical spine X-ray data of two different positions, and then inputs them into two feature map fusion layers for feature map fusion. The 2nd to nth feature extraction and fusion modules receive the output of the corresponding feature map fusion layer of the previous layer through two parallel feature extraction layers, and perform size compression, feature extraction and feature map fusion.

4. The difficult airway assessment system based on dual-position cervical spine X-ray according to claim 2, characterized in that: The feature extraction and fusion module includes a depth-separable convolution layer, a non-local neural network layer, a multi-resolution pooling layer and a convolution layer arranged in sequence; The depthwise separable convolutional layer takes the two received feature maps as inputs of the corresponding channels and fuses the two feature maps based on spatial information of different channels; The non-local neural network layer is used to capture long-range dependencies based on the fused feature data, and input into the multi-resolution pooling layer to extract feature representations of different scales; The convolution layer fuses the feature representations of different scales into a single-dimensional vector to obtain a feature representation after feature map fusion.

5. The difficult airway assessment system based on dual-position cervical spine X-ray according to claim 1, characterized in that: The compactness loss is used to minimize the standard deviation of the distance between the dual-position fusion feature vector and the shared center point to promote the compactness of intra-class features; The compactness loss is expressed as: ； in, is the compactness loss, is the standard deviation of the distance from the dual-position fusion feature vector to the shared center point, k is 0 or 1, is the average distance from the dual-position fusion feature vector to the shared center point.

6. The difficult airway assessment system based on dual-position cervical spine X-ray according to claim 1, characterized in that: When constructing the training sample set, the difficult airway sample data is also augmented by the following method: Learning the potential distribution of difficult airway samples in the feature space based on a variational autoencoder, sampling a new potential vector based on the potential distribution, and decoding the new potential vector into pseudo difficult airway sample data; A discriminant network is constructed to screen the pseudo difficult airway sample data, retaining the pseudo difficult airway sample data with a confidence level greater than a preset threshold as difficult airway sample data, marking the difficult airway label and adding it to the training sample set for model training.

7. The difficult airway assessment system based on dual-position cervical spine X-ray according to claim 6, characterized in that: When constructing the training sample set, data augmentation is also included through random rotation, horizontal and vertical flipping, random cropping, and random adjustment of brightness, contrast, and saturation.

Images