A heart-lung abnormality recognition apparatus and device

Abstract

The application discloses a kind of heart-lung abnormality identification device and equipment.By inhibiting the element of non-cardiopulmonary region in global feature map, the information transmission of non-cardiopulmonary region is inhibited, the interference of information of non-cardiopulmonary region to heart-lung abnormality identification result is avoided, and the accuracy and identification efficiency of heart-lung abnormality identification are improved.By constructing clustering code base based on negative samples in advance, and inputting the dimensionality reduction feature map into the clustering code base for retrieval, the heart-lung abnormality identification result is determined, the problem that positive samples are scarce and effective supervision training cannot be carried out is solved, the labeling cost is reduced, and at the same time, the heart-lung abnormality identification result is determined by retrieving from the clustering code base, and the identification efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of cardiopulmonary abnormality recognition technology, and in particular to a cardiopulmonary abnormality recognition device and equipment. Background Technology

[0002] Chest X-rays are currently the most common tool for screening and diagnosing pleural diseases. In routine clinical practice, chest X-ray screening and diagnosis rely on the experience of specialist physicians, which is time-consuming and prone to subjective assessment errors. In recent years, with the advancement of medical image processing technology, especially deep learning methods, computer-aided screening and diagnosis have been increasingly applied in radiology departments.

[0003] Due to the low specificity of chest X-rays, the inter-class and intra-class differentiation is very low, which limits the performance of anomaly recognition tasks and results in low anomaly recognition accuracy. Summary of the Invention

[0004] This invention provides a cardiopulmonary abnormality identification device and equipment to improve the accuracy and efficiency of cardiopulmonary abnormality identification.

[0005] In a first aspect, embodiments of the present invention provide a cardiopulmonary abnormality identification device, comprising:

[0006] The global feature map extraction module is used to extract global feature maps representing global information from the acquired chest X-ray images;

[0007] The local feature map extraction module is used to suppress elements in the non-cardiopulmonary region of the global feature map to obtain a local feature map that represents information about the cardiopulmonary region.

[0008] The dimension reduction processing module is used to perform dimension reduction processing on the local feature map to obtain a dimension-reduced feature map that represents effective information of the local feature map.

[0009] The identification result determination module is used to input the dimensionality reduction feature map into a clustering code pre-trained based on negative samples for retrieval and determine the cardiopulmonary abnormality identification result, which includes positive and negative results.

[0010] Secondly, embodiments of the present invention also provide a computer device, including the cardiopulmonary abnormality recognition device as provided in the first aspect of the present invention.

[0011] The cardiopulmonary abnormality recognition device provided in this invention comprises a global feature map extraction module that extracts a global feature map representing global information from an acquired chest X-ray image; a local feature map extraction module that suppresses elements in non-cardiopulmonary regions of the global feature map to obtain local feature maps representing information in the cardiopulmonary region; a dimensionality reduction module that performs dimensionality reduction processing on the local feature maps to obtain dimensionality-reduced feature maps representing effective information in the local feature maps; and a recognition result determination module that inputs the dimensionality-reduced feature maps into a pre-trained clustering codebook based on negative samples for retrieval to determine the cardiopulmonary abnormality recognition result, which includes positive and negative results. By suppressing elements in non-cardiopulmonary regions of the global feature map, the transmission of information from non-cardiopulmonary regions is suppressed, avoiding interference from information in non-cardiopulmonary regions on the cardiopulmonary abnormality recognition result, thus improving the accuracy and efficiency of cardiopulmonary abnormality recognition. By pre-constructing a clustering codebook based on negative samples and inputting the dimensionality-reduced feature maps into the clustering codebook for retrieval to determine the cardiopulmonary abnormality recognition result, the problem of scarce positive samples leading to the inability to conduct effective supervised training is solved, reducing annotation costs. Furthermore, the method of determining the cardiopulmonary abnormality recognition result by retrieving from the clustering codebook improves recognition efficiency. Attached Figure Description

[0012] Figure 1 This is a flowchart of a cardiopulmonary abnormality identification method provided in Embodiment 1 of the present invention;

[0013] Figure 2A This is a flowchart of a cardiopulmonary abnormality identification method provided in Embodiment 2 of the present invention;

[0014] Figure 2B This is a data processing flowchart of a cardiopulmonary abnormality identification method provided in an embodiment of the present invention;

[0015] Figure 2C This is a schematic diagram of a dimensionality-reduced feature map provided in an embodiment of the present invention;

[0016] Figure 3 This is a schematic diagram of the structure of a cardiopulmonary abnormality recognition device provided in Embodiment 3 of the present invention;

[0017] Figure 4 This is a schematic diagram of the structure of a computer device provided in Embodiment 4 of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0019] Example 1

[0020] Figure 1 This is a flowchart of a cardiopulmonary abnormality identification method provided in Embodiment 1 of the present invention. This embodiment is applicable to cardiopulmonary abnormality identification. The method can be executed by the cardiopulmonary abnormality identification device provided in this embodiment of the present invention. This device can be implemented by software and / or hardware, and is typically configured in a computer device, such as... Figure 1 As shown, the method specifically includes the following steps:

[0021] S101. Extract a global feature map from the acquired chest X-ray image to represent global information.

[0022] A chest X-ray, also known as a chest film, is created by projecting X-rays through the chest onto film. A normal chest X-ray image is a composite projection of various tissues and organs inside and outside the thoracic cavity, including the soft tissues of the chest wall, bones, heart and major blood vessels, lungs, pleura, and diaphragm. Normally inflated lungs appear as uniformly transparent areas on chest X-rays. On a frontal chest X-ray (taken when the subject is standing upright facing the X-ray source), the transparency of both lung fields is basically the same, and its transparency is directly proportional to the amount of gas contained in the lungs. The lung fields on both sides are divided into three longitudinal zones (inner, middle, and outer) on the left and three transverse zones (upper, middle, and lower) on the right. The lungs also include the hilum, lung markings, lobes and segments, trachea, and bronchi. The mediastinum is located behind the sternum and in front of the thoracic vertebrae, between the two lungs, and includes the heart, great vessels, trachea, main bronchus, esophagus, lymphatic tissue, thymus, nerves, and fat. On chest X-rays, apart from the trachea and main bronchus being distinguishable, the other mediastinal structures lack contrast.

[0023] By processing the acquired chest X-ray image, for example by inputting it into a pre-trained global network model, global feature maps are extracted to represent global information. A global feature map represents the chest X-ray image using all its features; it represents the overall attributes of the image. Common global feature maps include color features, texture features, and shape features. Because global feature maps are low-level visual features at the pixel level, they possess good invariance, are computationally simple, and are intuitively represented.

[0024] It should be noted that the embodiments of the present invention do not limit the network model used for extracting the global feature map, as long as it can obtain a global feature map that represents global information.

[0025] S102. Suppress the elements in the non-cardiopulmonary region of the global feature map to obtain a local feature map that represents the information of the cardiopulmonary region.

[0026] In this embodiment of the invention, the cardiopulmonary region refers to the area containing the heart, left lung, and right lung in the X-ray image, while the non-cardiopulmonary region refers to the area outside the central lung region in the X-ray image. Since cardiopulmonary abnormalities are mainly concentrated in the cardiopulmonary region, in this embodiment, elements in the non-cardiopulmonary region are suppressed in the global feature map to obtain a local feature map representing information about the cardiopulmonary region. By suppressing elements in the non-cardiopulmonary region of the global feature map, information transmission from the non-cardiopulmonary region is inhibited, avoiding interference from information in the non-cardiopulmonary region on the cardiopulmonary abnormality identification results, thus improving the accuracy and efficiency of cardiopulmonary abnormality identification.

[0027] It should be noted that, in this embodiment of the invention, the specific suppression method for elements in non-cardiopulmonary regions of the global feature map is not limited. For example, elements in non-cardiopulmonary regions can be given a smaller weight than elements in cardiopulmonary regions, or elements in non-cardiopulmonary regions can be set to zero directly, as long as the suppression of elements in non-cardiopulmonary regions of the global feature map can be achieved.

[0028] S103. Perform dimensionality reduction processing on the local feature map to obtain a dimensionality-reduced feature map that can effectively represent the local feature map.

[0029] After suppressing elements in non-cardiopulmonary regions of the global feature map to obtain local feature maps representing information about the cardiopulmonary region, the resulting local feature maps have very high dimensionality. This leads to a large workload and low efficiency when searching within the clustering encoding book. In this embodiment of the invention, dimensionality reduction processing is performed on the high-dimensional local feature maps to obtain dimensionality-reduced feature maps that effectively represent the local feature maps. This reduces the workload during subsequent searching within the clustering encoding book while preserving the effective information of the local feature maps, saving computational resources and improving search efficiency.

[0030] S104. Input the dimensionality-reduced feature map into the clustering codebook pre-trained based on negative samples for retrieval to determine the cardiopulmonary abnormality identification results, which include positive and negative results.

[0031] In this embodiment of the invention, a large number of negative samples (i.e., chest X-ray images with normal cardiopulmonary function) are pre-clustered and encoded to obtain a clustered encoding template based on negative samples. In real-world clinical scenarios, positive samples are scarce, and some diseases are even lacking. Furthermore, the annotation of positive samples requires detailed annotation by professionals on the specific location of lesions, abnormality categories, etc., which often takes a significant amount of time, making effective supervised training impossible. This embodiment of the invention pre-clusters and encodes a large number of negative samples to obtain a clustered encoding template based on negative samples. Then, the dimensionality-reduced feature map is input into the pre-trained clustered encoding template based on negative samples for retrieval to determine the cardiopulmonary abnormality identification result. This solves the problem of the scarcity of positive samples leading to the inability to conduct effective supervised training and reduces annotation costs.

[0032] Specifically, firstly, a large number of negative samples are obtained, and a global feature map representing global information is extracted from each negative sample. Next, elements in non-cardiopulmonary regions of the global feature map of the negative samples are suppressed to obtain a dimensionality-reduced feature map representing effective information of the local feature map of the negative samples. Then, clustering encoding is performed on the dimensionality-reduced feature maps of a large number of negative samples to obtain a clustered encoding version based on the negative samples. In this embodiment of the invention, the clustering encoding algorithm can be the K-MEANS clustering algorithm or other clustering algorithms; this embodiment of the invention is not limited thereto.

[0033] Clustering is used to cluster a large number of negative samples, resulting in a clustered coding book containing multiple categories. Negative samples in each category share similar characteristics and are represented by clusters. The resulting dimensionality-reduced feature map is then input into the clustered coding book for retrieval. The similarity between the dimensionality-reduced feature map and each cluster center is calculated, and the cardiopulmonary abnormality identification result is determined based on the similarity. The cardiopulmonary abnormality identification result includes positive and negative. A higher similarity indicates a greater probability of a negative cardiopulmonary abnormality identification result on the chest X-ray image. For example, if the similarity between the dimensionality-reduced feature map and each cluster center is greater than or equal to a preset value, the cardiopulmonary abnormality identification result is considered negative. Alternatively, if the similarity between the dimensionality-reduced feature map and any cluster center is less than a preset value, the cardiopulmonary abnormality identification result is considered positive.

[0034] The cardiopulmonary anomaly recognition method provided in this invention extracts a global feature map representing global information from an acquired chest X-ray image. Elements in non-cardiopulmonary regions of the global feature map are suppressed to obtain a local feature map representing information about the cardiopulmonary region. This local feature map is then dimensionality-reduced to obtain a dimensionality-reduced feature map representing effective information. The dimensionality-reduced feature map is input into a pre-trained clustering encoding book based on negative samples for retrieval to determine the cardiopulmonary anomaly recognition result, which includes both positive and negative results. By suppressing elements in non-cardiopulmonary regions of the global feature map, information transmission from these regions is inhibited, avoiding interference from non-cardiopulmonary region information on the cardiopulmonary anomaly recognition result, thus improving the accuracy and efficiency of cardiopulmonary anomaly recognition. By pre-constructing a clustering encoding book based on negative samples and inputting the dimensionality-reduced feature map into this book for retrieval to determine the cardiopulmonary anomaly recognition result, the problem of scarce positive samples leading to the inability to conduct effective supervised training is solved, reducing annotation costs. Furthermore, using a retrieval method from the clustering encoding book to determine the cardiopulmonary anomaly recognition result improves recognition efficiency.

[0035] Example 2

[0036] Figure 2A This is a flowchart of a cardiopulmonary abnormality identification method provided in Embodiment 2 of the present invention. Figure 2B This is a data processing flowchart of a cardiopulmonary abnormality identification method provided in an embodiment of the present invention. Based on the aforementioned Embodiment 1, this embodiment describes in detail the process of each step in the above embodiments, such as... Figure 2A , 2B As shown, the method includes:

[0037] S201. Input the chest X-ray image into a pre-trained autoencoder for processing to obtain a global feature map used to represent global information.

[0038] In this embodiment of the invention, a large number of negative samples are used to pre-train the autoencoder. After the autoencoder is trained, the chest X-ray image is input into the autoencoder for processing to obtain a global feature map for representing global information.

[0039] An autoencoder is a self-supervised learning model. Based on backpropagation and optimization methods (such as gradient descent), it uses the input data itself as supervision, ensuring the output is approximately equal to the input. Through self-supervised learning, the autoencoder learns the feature distribution of a large number of negative samples, reducing annotation costs.

[0040] For example, such as Figure 2BAs shown, an autoencoder consists of two parts: an encoder and a decoder. The encoder extracts the most representative information or features from the input negative samples, reducing the amount of information. The decoder learns the reduced features and outputs a global feature map representing global information. This saves a significant amount of time compared to directly learning the original negative samples. During training, the input to the autoencoder is the negative sample, and the output is the reconstructed negative sample. By comparing the error between the input and output of the autoencoder, backpropagation is performed to gradually improve the accuracy of the autoencoder.

[0041] For example, in this embodiment of the invention, the encoder uses a DenseNet-121 network as the backbone network. The feature map size of the last convolutional network layer of the encoder is 1024×7×7, and the decoder uses 5 layers of deconvolutional upsampling. In application, the input of the encoder is a chest X-ray image with an image size of C×H×W. Its pixel values ​​are normalized in advance, and then downsampled by the DenseNet-121 network and upsampled by deconvolution to output a global feature map of the same size as the original image (i.e., the input chest X-ray image).

[0042] S202. Perform semantic segmentation on the chest X-ray image to obtain a mask map including the cardiopulmonary region. In the mask map, the value of the element in the cardiopulmonary region is 1, and the value of the element in the non-cardiopulmonary region is 0.

[0043] Semantic segmentation refers to classifying each pixel in an image, assigning pixels to different preset categories, thereby obtaining different regions in the image. In this embodiment of the invention, semantic segmentation is performed on a chest X-ray image to divide it into cardiopulmonary and non-cardiopulmonary regions. Then, the pixel values ​​of pixels in the cardiopulmonary region are set to 1, and the pixel values ​​of pixels in the non-cardiopulmonary region are set to 0, resulting in a mask image.

[0044] By performing semantic segmentation on chest X-ray images, the cardiopulmonary region and non-cardiopulmonary region can be separated. Since pathological information is mainly concentrated in the cardiopulmonary region, subsequent processing can focus on the cardiopulmonary region, avoid interference from information in the non-cardiopulmonary region, and improve the accuracy of cardiopulmonary abnormality identification results.

[0045] For example, such as Figure 2B As shown, in this embodiment of the invention, semantic segmentation of chest X-ray images is performed based on a pre-trained U-Net model to obtain a mask image of size H×W. In the mask image, the value of elements in the cardiopulmonary region is 1, and the value of elements in the non-cardiopulmonary region is 0.

[0046] The U-Net model is an improved fully connected network architecture, named for its U-shaped structure. It consists of a left-hand contracting path and a right-hand expanding path. The contracting path follows a typical convolutional neural network structure, repeatedly using two convolutional layers and one max-pooling layer, doubling the dimension of the feature map with each pooling operation. In the expanding path, a deconvolution operation is performed to halve the dimension of the feature map. Then, the feature maps cropped from the contracting path are concatenated to form a new feature map twice the size. Two more convolutional layers are then used for feature extraction, and this process is repeated. Finally, two convolutional layers map the 64-dimensional feature map to a 2-dimensional output map. The U-Net model's "U-shaped" structure makes the cropping and concatenation process more intuitive and logical. The concatenation of high-level and low-level feature maps, along with the repeated and continuous convolutional operations, allows the model to derive a more accurate output feature map from both contextual and detailed information.

[0047] It should be noted that the embodiment of the present invention does not limit the segmentation model used for semantic segmentation. For example, it can also be a fully convolutional network model, DeepLab model, PSPNet or RefineNet, etc.

[0048] S203. Multiply the mask map with the corresponding elements of the global feature map to obtain a local feature map used to represent information about the cardiopulmonary region.

[0049] In embodiments of the present invention, such as Figure 2B As shown, the mask map is multiplied with the corresponding elements of the global feature map (elementwise multiplication). The values ​​of the elements in the central lung region of the global feature map are multiplied by 1 and retained, while the values ​​of the elements in the non-cardiopulmonary regions are multiplied by 0 and set to zero. In this way, the information of the non-cardiopulmonary regions is suppressed while the information of the cardiopulmonary regions is retained, resulting in a local feature map of size C×H×W.

[0050] S204. Perform principal component analysis on the local feature map to find the principal component feature map used to characterize the local feature map.

[0051] Principal Component Analysis (PCA) is a common data analysis technique often used for dimensionality reduction of high-dimensional data. It can be used to extract the principal feature components of the data. The main idea of ​​PCA is to map n-dimensional features to k-dimensional features. These k-dimensional features are new orthogonal features, also known as principal components, which are reconstructed from the original n-dimensional features.

[0052] In this embodiment of the invention, principal component analysis is performed on the local feature map to find the principal component feature map used to characterize the local feature map. The dimension of the principal component feature map is smaller than the dimension of the local feature map.

[0053] For example, in an embodiment of the present invention, principal component analysis of a local feature map includes the following sub-steps:

[0054] S2041. Calculate the covariance matrix of all elements in the local feature map.

[0055] For high-dimensional data, we use covariance to represent the dispersion of the data. Covariance can represent the correlation between two variables. To ensure that two variables represent as much of the original information as possible, we hope that there is no linear correlation between them, because correlation implies that the two variables are not completely independent, and there will inevitably be duplicated information.

[0056] The covariance matrix is ​​used to represent the correlation between each element of a local feature map and any other arbitrary element. The formula for calculating the covariance matrix is ​​as follows:

[0057]

[0058] Where m is the total number of elements in the local feature map, x i For the i-th element in the local feature map, (x i ) T For x i The transpose of .

[0059] S2042. Perform singular value decomposition on the covariance matrix to obtain multiple eigencomponents of the covariance matrix.

[0060] After obtaining the covariance matrix, singular value decomposition is performed on it. Each column of the resulting U vector represents the eigencomponents of the covariance matrix. Specifically, the decomposition process is as follows:

[0061]

[0062] Where U is an orthogonal matrix, u1…u n These are characteristic components.

[0063] S2043. Select the first k feature components as principal components.

[0064] Specifically, the top k feature components (u1…u) k ) is used as the principal component.

[0065] S2044. Project the values ​​of each element in the local feature map to the principal components to obtain the principal component feature map used to characterize the principal components of the local feature map.

[0066] Specifically, the values ​​of each element in the local feature map are projected onto the principal components to achieve dimensionality reduction of the local feature map, resulting in a principal component feature map used to characterize the principal components of the local feature map.

[0067] S205. Whiten the principal component feature map to obtain a dimensionality-reduced feature map that provides effective information for representing local feature maps.

[0068] Since adjacent pixels in a chest X-ray image are correlated, meaning elements in the principal component feature map are correlated, some information in the principal component feature map is redundant. Whitening aims to reduce this redundancy. Specifically, in this embodiment of the invention, the principal component feature map is whitened. Data whitening must meet two conditions: first, the correlation between different features must be minimal, close to 0; second, the variance of all features must be equal (not necessarily 1).

[0069] As mentioned earlier, after PCA processing, a principal component feature map is obtained, which includes k principal components. These principal components are independent and uncorrelated, satisfying the first condition for whitening. Therefore, simply dividing each principal component in the principal component feature map by its standard deviation yields a variance of 1 for each principal component, meaning the variances are equal. Thus, standard deviation normalization is performed on the data of each of the k principal components in the local feature map to obtain a dimensionality-reduced feature map that effectively represents the local feature map.

[0070] Figure 2C This is a schematic diagram of a dimensionality-reduced feature map provided in an embodiment of the present invention, exemplarily, as shown below. Figure 2C As shown, the total dimension of the dimensionality-reduced feature map is 8×128, meaning that the effective information of the local feature map can be represented using 8×128 dimensions. We can understand this as the dimensionality-reduced feature map comprising 8 principal components (i.e., k=8), with each principal component represented by 128 dimensions.

[0071] S206. Calculate the similarity between each principal component of the dimensionality-reduced feature map and the N cluster centers in the clustering code, and obtain the N similarity values ​​corresponding to each principal component.

[0072] In this embodiment of the invention, a large number of negative samples (i.e., chest X-ray images with normal heart and lungs) are pre-clustered and encoded to obtain a clustering codebook based on the negative samples. Specifically, firstly, a large number of negative samples are acquired, and a global feature map representing global information is extracted from each negative sample. Next, elements in non-heart and lung regions of the global feature map of the negative samples are suppressed to obtain a dimensionality-reduced feature map that represents the effective information of the local feature map of the negative samples. Then, the dimensionality-reduced feature maps of the large number of negative samples are clustered and encoded to obtain a clustering codebook (PQCodebook) based on the negative samples. Specifically, the process of obtaining the dimensionality-reduced feature map of the negative samples can refer to the aforementioned process of obtaining the dimensionality-reduced feature map of the chest X-ray image, and will not be repeated here in this embodiment of the invention.

[0073] Similarly, the dimensionality-reduced feature map of negative samples includes 8 principal components (k=8), each represented by 128 dimensions. Assuming a total of 50,000 negative samples, the resulting matrix is ​​50k×8×128. Next, clustering is performed on each of the 8 principal components with 50,000 negative samples. Assuming the number of clusters is 256, this will result in 256 cluster centers (N=256), yielding an 8×256×128 clustering code. This clustering code indicates that each negative sample has 256 cluster centers in each of the 8 principal components, each represented by 128 dimensions.

[0074] The similarity between each principal component of the dimensionality-reduced feature map and the 256 cluster centers in the clustering code is calculated, resulting in 256 similarity values ​​for each principal component. Specifically, the similarity can be Consine similarity; the higher the similarity, the greater the probability of a negative anomaly identification result.

[0075] S207. Take the maximum value among the N similarity values ​​corresponding to each principal component as the intermediate similarity value, and obtain the k intermediate similarity values ​​corresponding to the k principal components.

[0076] For example, for each principal component, the maximum value among 256 similarity values ​​is taken as the intermediate similarity value, thus obtaining 8 intermediate similarity values ​​corresponding to 8 principal components.

[0077] S208. Take the minimum value among the k intermediate similarity values ​​as the target similarity value.

[0078] For example, the minimum value among the eight intermediate similarity values ​​is taken as the target similarity value.

[0079] S209. Determine whether the target similarity value is less than the preset value. If yes, proceed to S210; otherwise, proceed to S211.

[0080] For example, the target similarity value is compared with a preset value to determine whether the target similarity value is less than the preset value. For example, the preset value can be set according to the task, and is usually set to 0.5.

[0081] S210, The cardiopulmonary abnormality identification result is confirmed as positive.

[0082] If the target similarity value is less than the preset value, the cardiopulmonary abnormality identification result is considered positive. That is, as long as the similarity between any principal component and one of the cluster centers is less than the preset value, the cardiopulmonary abnormality identification result is considered positive.

[0083] S211. The result of the cardiopulmonary abnormality identification is negative.

[0084] If the target similarity value is greater than or equal to the preset value, the cardiopulmonary abnormality identification result is considered negative.

[0085] The cardiopulmonary anomaly recognition method provided in this invention suppresses elements in non-cardiopulmonary regions of the global feature map, inhibiting information transmission from these regions and preventing interference from non-cardiopulmonary information in the recognition results, thereby improving the accuracy and efficiency of cardiopulmonary anomaly recognition. By pre-constructing a clustered encoding book based on negative samples and inputting the dimensionality-reduced feature map into the clustered encoding book for retrieval, the cardiopulmonary anomaly recognition results are determined. This solves the problem of scarce positive samples, which prevents effective supervised training and reduces annotation costs. Furthermore, using a retrieval method from the clustered encoding book improves recognition efficiency. Through self-supervised learning of the autoencoder, the feature distribution of a large number of negative samples is learned, further reducing annotation costs. Principal component analysis and whitening are performed on the local feature map to reduce its dimensionality. While preserving the effective information of the local feature map, this reduces the workload during subsequent retrieval in the clustered encoding book, saving computational resources and improving retrieval efficiency.

[0086] Example 3

[0087] Figure 3 This is a schematic diagram of the structure of a cardiopulmonary abnormality recognition device provided in Embodiment 3 of the present invention, as shown below. Figure 3 As shown, the device includes:

[0088] The global feature map extraction module 301 is used to extract global feature maps representing global information from the acquired chest X-ray images.

[0089] The global feature map extraction module 301 processes the acquired chest X-ray image, for example, by inputting the chest X-ray image into a pre-trained global network model for processing, to extract global feature maps that represent global information. A global feature map refers to a chest X-ray image represented by all its features; it is the overall attribute of the image. Common global feature maps include color features, texture features, and shape features. Because global feature maps are low-level visual features at the pixel level, they possess good invariance, are computationally simple, and are intuitively represented.

[0090] It should be noted that the embodiments of the present invention do not limit the network model used for extracting the global feature map, as long as it can obtain a global feature map that represents global information.

[0091] The local feature map extraction module 302 is used to suppress elements in the non-cardiopulmonary region of the global feature map to obtain a local feature map that represents information about the cardiopulmonary region.

[0092] In this embodiment of the invention, the cardiopulmonary region refers to the area containing the heart, left lung, and right lung in the X-ray image, while the non-cardiopulmonary region refers to the area outside the central lung region in the X-ray image. Since cardiopulmonary abnormalities are mainly concentrated in the cardiopulmonary region, in this embodiment, the local feature map extraction module 302 suppresses elements in the non-cardiopulmonary region of the global feature map to obtain a local feature map representing information about the cardiopulmonary region. By suppressing elements in the non-cardiopulmonary region of the global feature map, information transmission from the non-cardiopulmonary region is suppressed, avoiding interference from information in the non-cardiopulmonary region on the cardiopulmonary abnormality identification results, thus improving the accuracy and efficiency of cardiopulmonary abnormality identification.

[0093] It should be noted that, in this embodiment of the invention, the specific suppression method for elements in non-cardiopulmonary regions of the global feature map is not limited. For example, elements in non-cardiopulmonary regions can be given a smaller weight than elements in cardiopulmonary regions, or elements in non-cardiopulmonary regions can be set to zero directly, as long as the suppression of elements in non-cardiopulmonary regions of the global feature map can be achieved.

[0094] The dimensionality reduction processing module 303 is used to perform dimensionality reduction processing on the local feature map to obtain a dimensionality-reduced feature map that represents effective information of the local feature map.

[0095] After suppressing elements in non-cardiopulmonary regions of the global feature map to obtain local feature maps representing information about the cardiopulmonary region, the resulting local feature maps have very high dimensionality. This leads to a large workload and low efficiency when searching in the clustering encoding book. In this embodiment, the dimensionality reduction module 303 performs dimensionality reduction processing on the high-dimensional local feature maps to obtain dimensionality-reduced feature maps that represent effective information about the local feature maps. This reduces the workload during subsequent searching in the clustering encoding book while preserving the effective information of the local feature maps, saving computational resources and improving search efficiency.

[0096] The identification result determination module 304 is used to input the dimensionality reduction feature map into a clustering code pre-trained based on negative samples for retrieval and determine the cardiopulmonary abnormality identification result, which includes positive and negative results.

[0097] In this embodiment of the invention, a large number of negative samples (i.e., chest X-ray images with normal cardiopulmonary function) are pre-clustered and encoded to obtain a clustered encoding template based on negative samples. In real-world clinical scenarios, positive samples are scarce, and some diseases are even lacking. Furthermore, the annotation of positive samples requires detailed annotation by professionals on the specific location of lesions, abnormality categories, etc., which often takes a significant amount of time, making effective supervised training impossible. This embodiment of the invention pre-clusters and encodes a large number of negative samples to obtain a clustered encoding template based on negative samples. Then, the recognition result determination module 304 inputs the dimensionality-reduced feature map into the pre-trained clustered encoding template based on negative samples for retrieval to determine the cardiopulmonary abnormality recognition result. This solves the problem of the scarcity of positive samples leading to the inability to conduct effective supervised training and reduces annotation costs.

[0098] Specifically, firstly, a large number of negative samples are obtained, and a global feature map representing global information is extracted from each negative sample. Next, elements in non-cardiopulmonary regions of the global feature map of the negative samples are suppressed to obtain a dimensionality-reduced feature map representing effective information of the local feature map of the negative samples. Then, clustering encoding is performed on the dimensionality-reduced feature maps of a large number of negative samples to obtain a clustered encoding version based on the negative samples. In this embodiment of the invention, the clustering encoding algorithm can be the K-MEANS clustering algorithm or other clustering algorithms; this embodiment of the invention is not limited thereto.

[0099] Clustering is used to cluster a large number of negative samples, resulting in a clustered coding book containing multiple categories. Negative samples in each category share similar characteristics and are represented by clusters. The resulting dimensionality-reduced feature map is then input into the clustered coding book for retrieval. The similarity between the dimensionality-reduced feature map and each cluster center is calculated, and the cardiopulmonary abnormality identification result is determined based on the similarity. The cardiopulmonary abnormality identification result includes positive and negative. A higher similarity indicates a greater probability of a negative cardiopulmonary abnormality identification result on the chest X-ray image. For example, if the similarity between the dimensionality-reduced feature map and each cluster center is greater than or equal to a preset value, the cardiopulmonary abnormality identification result is considered negative. Alternatively, if the similarity between the dimensionality-reduced feature map and any cluster center is less than a preset value, the cardiopulmonary abnormality identification result is considered positive.

[0100] The cardiopulmonary abnormality recognition device provided in this invention comprises a global feature map extraction module that extracts a global feature map representing global information from an acquired chest X-ray image; a local feature map extraction module that suppresses elements in non-cardiopulmonary regions of the global feature map to obtain local feature maps representing information in the cardiopulmonary region; a dimensionality reduction module that performs dimensionality reduction processing on the local feature maps to obtain dimensionality-reduced feature maps representing effective information in the local feature maps; and a recognition result determination module that inputs the dimensionality-reduced feature maps into a pre-trained clustering codebook based on negative samples for retrieval to determine the cardiopulmonary abnormality recognition result, which includes positive and negative results. By suppressing elements in non-cardiopulmonary regions of the global feature map, the transmission of information from non-cardiopulmonary regions is suppressed, avoiding interference from information in non-cardiopulmonary regions on the cardiopulmonary abnormality recognition result, thus improving the accuracy and efficiency of cardiopulmonary abnormality recognition. By pre-constructing a clustering codebook based on negative samples and inputting the dimensionality-reduced feature maps into the clustering codebook for retrieval to determine the cardiopulmonary abnormality recognition result, the problem of scarce positive samples leading to the inability to conduct effective supervised training is solved, reducing annotation costs. Furthermore, the method of determining the cardiopulmonary abnormality recognition result by retrieving from the clustering codebook improves recognition efficiency.

[0101] In some embodiments of the present invention, the global feature map extraction module 301 is used to input the chest X-ray image into a pre-trained autoencoder for processing to obtain a global feature map for characterizing global information.

[0102] In this embodiment of the invention, a large number of negative samples are used to pre-train the autoencoder. After the autoencoder is trained, the chest X-ray image is input into the autoencoder for processing to obtain a global feature map for representing global information.

[0103] An autoencoder is a self-supervised learning model. Based on backpropagation and optimization methods (such as gradient descent), it uses the input data itself as supervision, ensuring the output is approximately equal to the input. Through self-supervised learning, the autoencoder learns the feature distribution of a large number of negative samples, reducing annotation costs.

[0104] For example, an autoencoder consists of two parts: an encoder and a decoder. The encoder extracts the most representative information or features from the input negative samples, reducing the amount of information. The decoder learns the reduced features and outputs a global feature map representing global information. This saves a significant amount of time compared to directly learning the original negative samples. During training, the input to the autoencoder is the negative sample, and the output is the reconstructed negative sample. By comparing the error between the input and output of the autoencoder, backpropagation is performed to gradually improve the accuracy of the autoencoder.

[0105] For example, in this embodiment of the invention, the encoder uses a DenseNet-121 network as the backbone network. The feature map size of the last convolutional network layer of the encoder is 1024×7×7, and the decoder uses 5 layers of deconvolutional upsampling. In application, the input of the encoder is a chest X-ray image with an image size of C×H×W. Its pixel values ​​are normalized in advance, and then downsampled by the DenseNet-121 network and upsampled by deconvolution to output a global feature map of the same size as the original image (i.e., the input chest X-ray image).

[0106] In some embodiments of the present invention, the local feature map extraction module 302 includes:

[0107] The semantic segmentation submodule is used to perform semantic segmentation on the chest X-ray image to obtain a mask image including the cardiopulmonary region. In the mask image, the value of the element in the cardiopulmonary region is 1, and the value of the element in the non-cardiopulmonary region is 0.

[0108] Semantic segmentation refers to classifying each pixel in an image, assigning pixels to different preset categories, thereby obtaining different regions in the image. In this embodiment of the invention, semantic segmentation is performed on a chest X-ray image to divide it into cardiopulmonary and non-cardiopulmonary regions. Then, the pixel values ​​of pixels in the cardiopulmonary region are set to 1, and the pixel values ​​of pixels in the non-cardiopulmonary region are set to 0, resulting in a mask image.

[0109] By performing semantic segmentation on chest X-ray images, the cardiopulmonary region and non-cardiopulmonary region can be separated. Since pathological information is mainly concentrated in the cardiopulmonary region, subsequent processing can focus on the cardiopulmonary region, avoid interference from information in the non-cardiopulmonary region, and improve the accuracy of cardiopulmonary abnormality identification results.

[0110] For example, in an embodiment of the present invention, semantic segmentation of a chest X-ray image is performed based on a pre-trained U-Net model to obtain a mask image of size H×W. In the mask image, the value of the element in the cardiopulmonary region is 1, and the value of the element in the non-cardiopulmonary region is 0.

[0111] The U-Net model is an improved fully connected network architecture, named for its U-shaped structure. It consists of a left-hand contracting path and a right-hand expanding path. The contracting path follows a typical convolutional neural network structure, repeatedly using two convolutional layers and one max-pooling layer, doubling the dimension of the feature map with each pooling operation. In the expanding path, a deconvolution operation is performed to halve the dimension of the feature map. Then, the feature maps cropped from the contracting path are concatenated to form a new feature map twice the size. Two more convolutional layers are then used for feature extraction, and this process is repeated. Finally, two convolutional layers map the 64-dimensional feature map to a 2-dimensional output map. The U-Net model's "U-shaped" structure makes the cropping and concatenation process more intuitive and logical. The concatenation of high-level and low-level feature maps, along with the repeated and continuous convolutional operations, allows the model to derive a more accurate output feature map from both contextual and detailed information.

[0112] It should be noted that the embodiment of the present invention does not limit the segmentation model used for semantic segmentation. For example, it can also be a fully convolutional network model, DeepLab model, PSPNet or RefineNet, etc.

[0113] The element-multiplication submodule is used to multiply the mask image with the corresponding elements of the global feature map to obtain a local feature map used to characterize the cardiopulmonary region.

[0114] In this embodiment of the invention, the element-wise multiplication submodule multiplies the mask map with the corresponding elements of the global feature map. The values ​​of the elements in the central lung region of the global feature map are multiplied by 1 and retained, while the values ​​of the elements in the non-cardiopulmonary regions are multiplied by 0 and set to zero. In this way, the information of the non-cardiopulmonary regions is suppressed while the information of the cardiopulmonary regions is retained, resulting in a local feature map of size C×H×W.

[0115] In some embodiments of the present invention, the dimensionality reduction processing module 303 includes: a principal component analysis submodule and a whitening processing submodule.

[0116] The principal component analysis submodule is used to perform principal component analysis on the local feature map to find the principal component feature map that represents the principal components of the local feature map, wherein the dimension of the principal component feature map is smaller than the dimension of the local feature map.

[0117] Principal Component Analysis (PCA) is a common data analysis technique often used for dimensionality reduction of high-dimensional data. It can be used to extract the principal feature components of the data. The main idea of ​​PCA is to map n-dimensional features to k-dimensional features. These k-dimensional features are new orthogonal features, also known as principal components, which are reconstructed from the original n-dimensional features.

[0118] In this embodiment of the invention, the principal component analysis submodule performs principal component analysis on the local feature map to find the principal component feature map used to characterize the local feature map. The dimension of the principal component feature map is smaller than the dimension of the local feature map.

[0119] In some embodiments of the present invention, the principal component analysis submodule includes:

[0120] The covariance matrix calculation unit is used to calculate the covariance matrix of all elements of the local feature map.

[0121] For high-dimensional data, we use covariance to represent the dispersion of the data. Covariance can represent the correlation between two variables. To ensure that two variables represent as much of the original information as possible, we hope that there is no linear correlation between them, because correlation implies that the two variables are not completely independent, and there will inevitably be duplicated information.

[0122] The covariance matrix is ​​used to represent the correlation between each element of a local feature map and any other arbitrary element. The formula for calculating the covariance matrix is ​​as follows:

[0123]

[0124] Where m is the total number of elements in the local feature map, x i For the i-th element in the local feature map, (x i ) T For x i The transpose of .

[0125] The singular value decomposition unit is used to perform singular value decomposition on the covariance matrix to obtain multiple eigencomponents of the covariance matrix.

[0126] After obtaining the covariance matrix, the singular value decomposition unit performs singular value decomposition on the covariance matrix, and each column of the resulting U vector is an eigencomponent of the covariance matrix. Specifically, the decomposition process is as follows:

[0127]

[0128] Where U is an orthogonal matrix, u1…u n These are characteristic components.

[0129] The principal component determination unit is used to select the first k characteristic components as principal components.

[0130] Specifically, the principal component determination unit will prioritize the first k eigencomponents (u1…u ... k ) is used as the principal component.

[0131] The projection unit is used to project the values ​​of each element in the local feature map onto the principal components to obtain a principal component feature map that characterizes the principal components of the local feature map.

[0132] Specifically, the projection unit projects the values ​​of each element in the local feature map to the principal components, thereby reducing the dimensionality of the local feature map and obtaining a principal component feature map that represents the principal components of the local feature map.

[0133] The whitening processing submodule is used to perform whitening processing on the principal component feature map to obtain a dimensionality-reduced feature map that can effectively represent the local feature map.

[0134] Since adjacent pixels in a chest X-ray image are correlated, meaning elements in the principal component feature map are correlated, some information in the principal component feature map is redundant. Whitening aims to reduce this redundancy. Specifically, in this embodiment of the invention, the principal component feature map is whitened.

[0135] In some embodiments of the present invention, the whitening process submodule includes:

[0136] The standard deviation normalization unit is used to perform standard deviation normalization on the data of each of the k principal components of the local feature map to obtain a dimensionality-reduced feature map that can effectively represent the local feature map.

[0137] Data whitening must meet two conditions: first, the correlation between different features must be minimal, close to 0; second, the variance of all features must be equal (not necessarily 1).

[0138] As mentioned earlier, after PCA processing, a principal component feature map is obtained, which includes k principal components. These principal components are independent and uncorrelated, satisfying the first condition for whitening. Therefore, dividing each principal component in the principal component feature map by its standard deviation yields a variance of 1 for each principal component, meaning the variances are equal. Thus, the standard deviation normalization unit normalizes the standard deviation of each of the k principal components in the local feature map, resulting in a dimensionality-reduced feature map that effectively represents the local feature map.

[0139] For example, the total dimension of the dimensionality-reduced feature map is 8×128, meaning that the effective information of the local feature map can be represented by 8×128 dimensions. We can understand this as the dimensionality-reduced feature map including 8 principal components (i.e., k=8), and each principal component is represented by 128 dimensions.

[0140] In some embodiments of the present invention, the dimensionality-reduced feature map includes k principal components, and the recognition result determination module 304 includes:

[0141] The similarity calculation submodule is used to calculate the similarity between each principal component of the dimensionality-reduced feature map and N cluster centers in the clustering code, and obtain N similarity values ​​corresponding to each principal component.

[0142] In this embodiment of the invention, a large number of negative samples (i.e., chest X-ray images with normal heart and lungs) are pre-clustered and encoded to obtain a clustering codebook based on the negative samples. Specifically, firstly, a large number of negative samples are acquired, and a global feature map representing global information is extracted from each negative sample. Next, elements in non-heart and lung regions of the global feature map of the negative samples are suppressed to obtain a dimensionality-reduced feature map that represents the effective information of the local feature map of the negative samples. Then, the dimensionality-reduced feature maps of the large number of negative samples are clustered and encoded to obtain a clustering codebook (PQCodebook) based on the negative samples. Specifically, the process of obtaining the dimensionality-reduced feature map of the negative samples can refer to the aforementioned process of obtaining the dimensionality-reduced feature map of the chest X-ray image, and will not be repeated here in this embodiment of the invention.

[0143] Similarly, the dimensionality-reduced feature map of negative samples includes 8 principal components (k=8), each represented by 128 dimensions. Assuming a total of 50,000 negative samples, the resulting matrix is ​​50k×8×128. Next, clustering is performed on each of the 8 principal components with 50,000 negative samples. Assuming the number of clusters is 256, this will result in 256 cluster centers (N=256), yielding an 8×256×128 clustering code. This clustering code indicates that each negative sample has 256 cluster centers in each of the 8 principal components, each represented by 128 dimensions.

[0144] The similarity calculation submodule calculates the similarity between each principal component of the dimensionality-reduced feature map and the 256 cluster centers in the clustering code, resulting in 256 similarity values ​​for each principal component. Specifically, the similarity can be Consine similarity; the higher the similarity, the greater the probability of a negative representational anomaly identification result.

[0145] The intermediate similarity value selection submodule is used to take the maximum value among the N similarity values ​​corresponding to each principal component as the intermediate similarity value, so as to obtain k intermediate similarity values ​​corresponding to k principal components.

[0146] For example, for each principal component, the intermediate similarity value selection submodule takes the maximum value among 256 similarity values ​​as the intermediate similarity value, thus obtaining 8 intermediate similarity values ​​corresponding to the 8 principal components.

[0147] The target similarity value selection submodule is used to select the minimum value among the k intermediate similarity values ​​as the target similarity value.

[0148] For example, the target similarity value selection submodule takes the minimum value among 8 intermediate similarity values ​​as the target similarity value.

[0149] The judgment submodule is used to determine whether the target similarity value is less than a preset value.

[0150] For example, the judgment submodule compares the target similarity value with a preset value to determine whether the target similarity value is less than the preset value. For example, the preset value can be set according to the task, and is usually set to 0.5.

[0151] The positive result determination submodule is used to determine the cardiopulmonary abnormality identification result as positive when the target similarity value is less than the preset value.

[0152] If the target similarity value is less than a preset value, the positive result determination submodule determines the cardiopulmonary abnormality identification result as positive. That is, as long as the similarity between any principal component and one of the cluster centers is less than the preset value, the cardiopulmonary abnormality identification result is considered positive.

[0153] The negative result determination submodule is used to determine that the cardiopulmonary abnormality identification result is negative when the target similarity value is greater than or equal to the preset value.

[0154] If the target similarity value is greater than or equal to the preset value, the negative result determination submodule determines the cardiopulmonary abnormality identification result as negative.

[0155] The cardiopulmonary abnormality recognition device provided in this invention suppresses elements in non-cardiopulmonary regions of the global feature map, inhibiting information transmission from these regions and preventing interference from non-cardiopulmonary information on the recognition results, thereby improving the accuracy and efficiency of cardiopulmonary abnormality recognition. By pre-constructing a clustered encoding book based on negative samples and inputting the dimensionality-reduced feature map into the clustered encoding book for retrieval, the cardiopulmonary abnormality recognition results are determined. This solves the problem of scarce positive samples, which prevents effective supervised training and reduces annotation costs. Furthermore, using a retrieval method from the clustered encoding book improves recognition efficiency. Through self-supervised learning of the autoencoder, the feature distribution of a large number of negative samples is learned, further reducing annotation costs. Principal component analysis and whitening are performed on the local feature map to reduce its dimensionality. While preserving the effective information of the local feature map, this reduces the workload during subsequent retrieval in the clustered encoding book, saving computational resources and improving retrieval efficiency.

[0156] Example 4

[0157] Embodiment 4 of the present invention provides a computer device, Figure 4 This is a schematic diagram of the structure of a computer device provided in Embodiment 4 of the present invention, as shown below. Figure 4 As shown, the computer device includes:

[0158] Processor 401, memory 402, communication module 403, input device 404, and output device 405; the number of processors 401 in the computer device can be one or more. Figure 4 Taking a processor 401 as an example; the processor 401, memory 402, communication module 403, input device 404, and output device 405 in a computer device can be connected via a bus or other means. Figure 4 Taking a bus connection as an example, the processor 401, memory 402, communication module 403, input device 404, and output device 405 mentioned above can be integrated into a computer device.

[0159] The memory 402, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the module corresponding to the cardiopulmonary abnormality recognition method in the above embodiment. The processor 401 executes various functional applications and data processing of the computer device by running the software programs, instructions, and modules stored in the memory 402, thereby realizing the cardiopulmonary abnormality recognition method described above.

[0160] Memory 402 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the microcomputer. Furthermore, memory 402 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, memory 402 may further include memory remotely located relative to processor 401, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0161] The communication module 403 is used to establish a connection with external devices (such as smart terminals) and to realize data interaction with external devices. The input device 404 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the computer device.

[0162] The computer device provided in this embodiment can execute the cardiopulmonary abnormality identification method provided in any of the above embodiments of the present invention, and has corresponding functions and beneficial effects.

[0163] Example 5

[0164] Embodiment 5 of the present invention provides a storage medium containing computer-executable instructions, on which a computer program is stored. When executed by a processor, the program implements the cardiopulmonary abnormality identification method provided in any of the above embodiments of the present invention. The method includes:

[0165] Extract global feature maps from the acquired chest X-ray images to represent global information;

[0166] Suppressing elements in non-cardiopulmonary regions of the global feature map yields a local feature map used to characterize the cardiopulmonary region.

[0167] The local feature map is subjected to dimensionality reduction processing to obtain a dimensionality-reduced feature map that can effectively represent the local feature map;

[0168] The reduced feature map is input into a clustering code pre-trained based on negative samples for retrieval to determine the cardiopulmonary abnormality identification result, which includes positive and negative results.

[0169] It should be noted that the embodiments of the apparatus, computer equipment, and storage media are basically similar to the method embodiments, so the descriptions are relatively simple. For relevant details, please refer to the descriptions of the method embodiments.

[0170] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a robot, personal computer, server, or network device, etc.) to execute the cardiopulmonary abnormality recognition method described in any embodiment of the present invention.

[0171] It is worth noting that the various modules, sub-modules, and units included in the above-mentioned device are only divided according to functional logic, but are not limited to the above division, as long as they can achieve the corresponding functions; in addition, the specific names of each functional module are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0172] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0173] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0174] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A cardiopulmonary abnormality detection device, characterized in that, include: The global feature map extraction module is used to extract global feature maps representing global information from the acquired chest X-ray images; The local feature map extraction module is used to suppress elements in the non-cardiopulmonary region of the global feature map to obtain a local feature map that represents information about the cardiopulmonary region. The dimension reduction processing module is used to perform dimension reduction processing on the local feature map to obtain a dimension-reduced feature map that represents effective information of the local feature map. The identification result determination module is used to input the dimensionality reduction feature map into a clustering code pre-trained based on negative samples for retrieval, and determine the cardiopulmonary abnormality identification result, which includes positive and negative results; The dimensionality reduction processing module includes: The principal component analysis submodule is used to perform principal component analysis on the local feature map to find the principal component feature map that represents the principal components of the local feature map, wherein the dimension of the principal component feature map is smaller than the dimension of the local feature map. The whitening processing submodule is used to perform whitening processing on the principal component feature map to obtain a dimension-reduced feature map that can effectively represent the local feature map. The identification result determination module includes: The similarity calculation submodule is used to calculate the similarity between each principal component of the dimensionality-reduced feature map and N cluster centers in the clustering code, and obtain N similarity values ​​corresponding to each principal component; The intermediate similarity value selection submodule is used to take the maximum value among the N similarity values ​​corresponding to each principal component as the intermediate similarity value, and obtain k intermediate similarity values ​​corresponding to k principal components; The target similarity value selection submodule is used to take the minimum value among the k intermediate similarity values ​​as the target similarity value; The judgment submodule is used to determine whether the target similarity value is less than a preset value; The positive result determination submodule is used to determine the cardiopulmonary abnormality identification result as positive if the target similarity value is less than the preset value. The negative result determination submodule is used to determine the cardiopulmonary abnormality identification result as negative if the target similarity value is greater than or equal to the preset value.

2. The cardiopulmonary abnormality identification device according to claim 1, characterized in that, The global feature map extraction module is used to input the chest X-ray image into a pre-trained autoencoder for processing to obtain a global feature map that represents global information.

3. The cardiopulmonary abnormality identification device according to claim 1, characterized in that, The local feature map extraction module includes: The semantic segmentation submodule is used to perform semantic segmentation on the chest X-ray image to obtain a mask image including the cardiopulmonary region. In the mask image, the value of the element in the cardiopulmonary region is 1, and the value of the element in the non-cardiopulmonary region is 0. The element-multiplication submodule is used to multiply the mask image with the corresponding elements of the global feature map to obtain a local feature map used to characterize the cardiopulmonary region.

4. The cardiopulmonary abnormality identification device according to claim 3, characterized in that, The semantic segmentation submodule is used to perform semantic segmentation on the chest X-ray image based on a pre-trained U-Net model to obtain a mask image including the cardiopulmonary region. In the mask image, the value of the element in the cardiopulmonary region is 1, and the value of the element in the non-cardiopulmonary region is 0.

5. The cardiopulmonary abnormality identification device according to claim 1, characterized in that, The principal component analysis submodule includes: The covariance matrix calculation unit is used to calculate the covariance matrix of all elements of the local feature map; A singular value decomposition unit is used to perform singular value decomposition on the covariance matrix to obtain multiple eigencomponents of the covariance matrix. The principal component determination unit is used to select the first k feature components as principal components. The projection unit is used to project the values ​​of each element in the local feature map onto the principal components to obtain a principal component feature map that characterizes the principal components of the local feature map.

6. The cardiopulmonary abnormality identification device according to claim 5, characterized in that, The whitening process submodule includes: The standard deviation normalization unit is used to perform standard deviation normalization on the data of each of the k principal components of the local feature map to obtain a dimensionality-reduced feature map that can effectively represent the local feature map.

7. A computer device, characterized in that, Includes the cardiopulmonary abnormality recognition device as described in any one of claims 1-6.

Images