An infection detection method, apparatus, device and medium

Abstract

The application provides an infection detection method, device, equipment and medium, comprising: acquiring a to-be-detected cross-sectional image corresponding to a to-be-detected part, and inputting the to-be-detected cross-sectional image into a pre-trained infection detection model, wherein the infection detection model comprises a convolution network, a transformer encoder, a transformer decoder and a feedforward network; the first feature information of the to-be-detected cross-sectional image is acquired by using the convolution network and the transformer encoder; whether there is an actual infection area in the to-be-detected cross-sectional image is judged according to the first feature information by using the transformer decoder and the feedforward network; when the judgment result is that there is an actual infection area in the to-be-detected cross-sectional image, the first detection position of the actual infection area in the to-be-detected cross-sectional image is further output by using the feedforward network. By using the above technical scheme, high-accuracy infection judgment and infection position output can be obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical imaging, in particular to an infection detection method, device, equipment and medium. BACKGROUND

[0002] Infection is a common disease that can lead to serious complications such as osteomyelitis and arthritis, causing great pain and health risks to patients. Traditional diagnostic methods include physical examination, blood test and imaging examination, among which imaging examination is one of the most commonly used methods for diagnosing bone infection. MRI images, as a high-resolution imaging method, are widely used in the diagnosis of bone infection. However, due to the complex structure and multi-modal information of MRI images, and the need for experienced doctors to interpret them, the accuracy and efficiency of diagnosis still face challenges. Especially in clinical practice, lesions in transverse images are more difficult to identify than in sagittal images.

[0003] With the development of computer vision and deep learning technology, it has become possible to automatically detect and locate bone infection. In order to improve the accuracy and efficiency of diagnosis, a suitable and effective model is needed to extract key information from limited transverse images to accurately detect and locate spinal infection, providing an effective reference for doctors' diagnosis and achieving accurate identification and early identification. SUMMARY

[0004] In order to overcome the above technical defects, the purpose of the present application is to provide an infection detection method, device, equipment and medium for feature extraction of transverse images for infection detection.

[0005] The present application discloses an infection detection method, comprising the following steps:

[0006] Obtain the transverse image corresponding to the part to be detected, and input the transverse image to be detected into a pre-trained infection detection model, wherein the infection detection model comprises a convolutional network, a transformer encoder, a transformer decoder and a feedforward network.

[0007] The convolutional network and the transformer encoder are used to obtain the first feature information of the transverse image to be detected.

[0008] The transformer decoder and the feedforward network are used to determine whether there is an actual infection area in the transverse image to be detected according to the first feature information.

[0009] When the judgment result is that there is an actual infection area in the transverse image to be detected, the feedforward network is further used to output the first detection position of the actual infection area in the transverse image to be detected.

[0010] Preferably, the first feature information of the to-be-detected cross-sectional image is obtained by:

[0011] The image feature sequence of the to-be-detected cross-sectional image is obtained by using the convolutional network.

[0012] The global feature of the to-be-detected cross-sectional image is obtained by using the transformer encoder to perform position encoding and attention weighted averaging on the image feature sequence, as the first feature information of the to-be-detected cross-sectional image.

[0013] Preferably, the determination of whether there is an actual infection area in the to-be-detected cross-sectional image comprises:

[0014] The transformer encoder is used to aggregate a plurality of local features in the to-be-detected cross-sectional image according to the first feature information, to obtain the relationship between the plurality of local features,

[0015] The feedforward network is used to determine whether there is an actual infection area in the to-be-detected cross-sectional image according to the relationship between the plurality of local features in the to-be-detected cross-sectional image, and to obtain a first detection position where the actual infection area is located in the to-be-detected cross-sectional image.

[0016] Preferably, the infection detection model is obtained by the following steps:

[0017] A plurality of training cross-sectional images and a to-be-trained infection detection model are obtained, any of the plurality of training cross-sectional images has a cross-sectional infection area, and any of the plurality of cross-sectional infection areas has a first relative position in the training cross-sectional image corresponding thereto;

[0018] The convolutional network and the transformer encoder are used to obtain second feature information of any of the plurality of training cross-sectional images;

[0019] The transformer decoder and the feedforward network are used to obtain a second detection position of any of the plurality of cross-sectional infection areas in the training cross-sectional image corresponding thereto according to the plurality of second feature information;

[0020] The first loss function is obtained by comparing the first relative position corresponding to any of the plurality of cross-sectional infection areas with the second detection position, and the back propagation is performed to optimize the parameters of the to-be-trained infection detection model;

[0021] The above steps are repeated until the first loss function converges, and the infection detection model is obtained.

[0022] Preferably, the pre-training step further comprises:

[0023] obtaining a plurality of training sagittal images corresponding to the plurality of training transverse images, any of the plurality of training sagittal images has a sagittal infection region, obtaining a second relative position of any of the plurality of sagittal infection regions in the training sagittal image corresponding thereto;

[0024] the convolutional network and the transformer encoder obtain third feature information of any of the plurality of training sagittal images, and input the plurality of third feature information and the plurality of second feature information into the transformer decoder one by one respectively;

[0025] the transformer decoder and the feedforward network obtain a second detection position of any of the plurality of transverse infection regions in the training transverse image corresponding thereto according to the plurality of second feature information and the plurality of third feature information;

[0026] the transformer decoder and the feedforward network obtain a third detection position of any of the plurality of sagittal infection regions in the training sagittal image corresponding thereto according to the plurality of third feature information;

[0027] comparing the first relative position and the second detection position of any of the plurality of transverse infection regions respectively to obtain a first loss function, comparing the second relative position and the third detection position of any of the plurality of sagittal infection regions respectively to obtain a second loss function, and performing back propagation to optimize the parameters of the to-be-trained infection detection model;

[0028] repeating the above steps until the first loss function and the second loss function converge to obtain the infection detection model.

[0029] Preferably, the obtaining of the first loss function comprises:

[0030] The first loss function comprises a first classification loss function and a first positioning loss function, the first cross-entropy is obtained as the first classification loss function by comparing the first relative position and the second detection position of any of the plurality of transverse infection regions respectively, and the first mean absolute error is obtained as the first positioning loss function.

[0031] Preferably, the obtaining of the second loss function comprises:

[0032] The second loss function comprises a second classification loss function and a second positioning loss function, and the second classification loss function is obtained by comparing any corresponding second relative position in the plurality of sagittal infection regions with a third detection position, and the second positioning loss function is obtained by taking the second average absolute value error.

[0033] The application further discloses an infection detection device, comprising:

[0034] A preprocessing module is configured to acquire a to-be-detected transverse image corresponding to a to-be-detected part and input the to-be-detected transverse image into a pre-trained infection detection model, wherein the infection detection model comprises a convolution network, a transformer encoder, a transformer decoder and a feedforward network.

[0035] A feature extraction module is configured to acquire first feature information of the to-be-detected transverse image by using the convolution network and the transformer encoder.

[0036] A judgment module is configured to judge whether an actual infection region exists in the to-be-detected transverse image according to the first feature information by using the transformer decoder and the feedforward network.

[0037] A position detection module is configured to further output a first detection position of the actual infection region in the to-be-detected transverse image by using the feedforward network when the judgment result is that the actual infection region exists in the to-be-detected transverse image.

[0038] The application further discloses an electronic device comprising a memory storing computer executable instructions and a processor, and when the instructions are executed by the processor, the electronic device implements the infection detection method.

[0039] The application further discloses a computer readable storage medium storing a computer program, and when the instructions are executed on a computer, the computer executes the infection detection method.

[0040] Compared with the prior art, the above technical scheme has the following beneficial effects:

[0041] 1. The infection discrimination and infection position output with high accuracy can be obtained, and the application is a very effective clinical diagnosis auxiliary system.

[0042] 2. The design of the infection detection model is simplified, the global feature extraction can be realized efficiently, the detection speed is faster while the accuracy is ensured, and

[0043] 3. The feature information of the sagittal image is involved in the cross-attention mechanism in the transverse image, multi-modal feature information fusion is realized, and the feature information of the sagittal image which is easy to identify the lesion can be used to assist the transverse image which is difficult to identify the lesion to perform infection detection. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 A flowchart of an infection detection method disclosed by the present application is shown in the figure.

[0045] Figure 2 A structure diagram of an infection detection model in an infection detection method disclosed by the present application is shown in the figure. DETAILED DESCRIPTION

[0046] The advantages of the present application are further described below in combination with the accompanying drawings and specific embodiments.

[0047] The exemplary embodiments will be described in detail herein below with reference to the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments are not meant to represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0048] The terminology used in the present disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. As used in the present disclosure and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0049] It will be understood that, although the terms first, second, third, etc. can be used herein to describe various information, these terms are not intended to denote a particular order or hierarchy. These terms are used only to distinguish one from another. For example, a first information can be termed a second information, and, similarly, a second information can be termed a first information, without departing from the scope of the present disclosure. Depending on the context, the word "if' as used herein can be interpreted as meaning "when" or "in response to determining" or "in response to ascertaining".

[0050] In the description of the present application, it should be understood that the terms "longitudinal", "transverse", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0051] In the description of the present application, unless otherwise specified and limited, it should be noted that the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be mechanical connection or electrical connection, it can be the communication between two elements, it can be direct connection or indirect connection through intermediate medium, and the specific meaning of the above terms can be understood by those skilled in the art according to the specific circumstances.

[0052] In the subsequent description, the suffix such as "module", "component" or "unit" used to represent elements is only for the convenience of the description of the present application, and has no specific meaning. Therefore, "module" and "component" can be used interchangeably.

[0053] Referring to Figure 1 The present application discloses an infection detection method, comprising the following steps:

[0054] S100, acquiring a to-be-detected transverse section image corresponding to a to-be-detected part, and inputting the to-be-detected transverse section image into a pre-trained infection detection model, wherein the infection detection model comprises a convolution network, a transformer encoder, a transformer decoder and a feedforward network;

[0055] S200, acquiring first feature information of the to-be-detected transverse section image by using the convolution network and the transformer encoder;

[0056] S300, judging whether there is an actual infection area in the to-be-detected transverse section image according to the first feature information by using the transformer decoder and the feedforward network;

[0057] S400, when the judgment result is that there is an actual infection area in the to-be-detected transverse section image, further outputting a first detection position where the actual infection area is located in the to-be-detected transverse section image by using the feedforward network.

[0058] Preferably, the first feature information of the to-be-detected transverse section image comprises:

[0059] obtaining an image feature sequence of the to-be-detected cross-sectional image by using the convolutional network;

[0060] performing position encoding and attention weighted averaging on the image feature sequence by using the transformer encoder to obtain a global feature of the to-be-detected cross-sectional image as first feature information of the to-be-detected cross-sectional image.

[0061] Preferably, the determining whether an actual infection area exists in the to-be-detected cross-sectional image comprises:

[0062] performing aggregation on a plurality of local features in the to-be-detected cross-sectional image according to the first feature information by using the transformer encoder to obtain a relationship between the plurality of local features,

[0063] determining whether an actual infection area exists in the to-be-detected cross-sectional image according to the relationship between the plurality of local features in the to-be-detected cross-sectional image by using the feedforward network, and obtaining a first detection position where the actual infection area is located in the to-be-detected cross-sectional image.

[0064] Referring to Figure 2 Preferably, the infection detection model is obtained by the following steps:

[0065] obtaining a plurality of training cross-sectional images and a to-be-trained infection detection model, any of the plurality of training cross-sectional images has a cross-sectional infection area, and any of the plurality of cross-sectional infection areas has a first relative position in a training cross-sectional image corresponding to the cross-sectional infection area;

[0066] obtaining second feature information of any of the plurality of training cross-sectional images by using the convolutional network and the transformer encoder;

[0067] obtaining a second detection position where any of the plurality of cross-sectional infection areas is located in the training cross-sectional image corresponding to the cross-sectional infection area by using the transformer decoder and the feedforward network according to the plurality of second feature information;

[0068] comparing the first relative position corresponding to any of the plurality of cross-sectional infection areas with the second detection position respectively to obtain a first loss function, and performing back propagation to optimize parameters of the to-be-trained infection detection model;

[0069] repeating the above steps until the first loss function converges, and obtaining the infection detection model.

[0070] Specifically, a plurality of training transverse images are obtained by performing a transverse direction scan on a plurality of diagnosed patients determined to have an infection. A first relative position of a transverse infection region in each training transverse image is labeled, for example, a professional physician can delineate the transverse infection region in the training transverse image using labeling software, and the boundary coordinates of the delineated transverse infection region are obtained as the first relative position of the transverse infection region.

[0071] Since the infection detection model to be trained in the present application aggregates the features of the sequence, before inputting the plurality of training transverse images into the infection detection model to be trained, the plurality of training transverse images need to be preprocessed to be converted into a sequence. For example, after unifying the number of layers, length, and width of the plurality of training transverse images, the plurality of training transverse images are respectively converted into transverse image matrices with fixed size and dimension and input into a convolutional network to extract training image features. The convolutional network can use ResNet50. The convolutional network outputs a plurality of transverse image feature matrices, and a convolution kernel is used to reduce the dimension of each transverse image feature matrix to obtain a plurality of one-dimensional transverse image feature sequences, and the plurality of transverse image feature sequences are input into a transformer encoder. In the transformer encoder, a fixed position encoding is added to each attention layer, each position in the transverse image feature sequence interacts with other positions to establish a global dependency relationship, and a multi-head attention mechanism is used to calculate attention weights, thereby performing attention weighted averaging on the input transverse image feature sequence. After multiple layers of processing by the transformer encoder, a plurality of sequences representing the global features of each training transverse image are generated, i.e., the second feature information of each training transverse image.

[0072] The transformer decoder takes the second feature information as input, learns the relationship between the global features of the second feature information through a self-attention mechanism, learns the relationship between local features through a multi-head attention mechanism, and adds a clustering position encoding to each attention layer to cluster the local features. The output of the transformer decoder is decoded into a second detection position and a class label of the transverse infection region by a feedforward network.

[0073] Referring to Figure 2 Preferably, the pre-training step further comprises:

[0074] A plurality of training sagittal images corresponding one-to-one to the plurality of training transverse images are obtained, any of the plurality of training sagittal images exists a sagittal infection region, and a second relative position of any of the plurality of sagittal infection regions in the training sagittal image corresponding thereto is obtained.

[0075] The convolutional network and the transformer encoder obtain third feature information of any of the plurality of training sagittal images, and input the plurality of third feature information and the plurality of second feature information respectively one by one into the transformer decoder;

[0076] The transformer decoder and the feedforward network obtain, according to the plurality of second feature information and the plurality of third feature information, a second detection position of any of the plurality of transverse infection regions in the training transverse image corresponding thereto;

[0077] The transformer decoder and the feedforward network obtain, according to the plurality of third feature information, a third detection position of any of the plurality of sagittal infection regions in the training sagittal image corresponding thereto;

[0078] The first loss function is obtained by comparing the corresponding first relative position and the second detection position of any of the plurality of transverse infection regions, the second loss function is obtained by comparing the corresponding second relative position and the third detection position of any of the plurality of sagittal infection regions, and back propagation is performed to optimize the parameters of the to-be-trained infection detection model;

[0079] The above steps are repeated until the first loss function and the second loss function converge, and the infection detection model is obtained.

[0080] Specifically, in order to utilize the sagittal image which is easy to identify the lesion to assist the diagnosis of the transverse image which is difficult to identify the lesion, the sagittal image is added to the training of the infection detection model. A plurality of training sagittal images are obtained by scanning the sagittal direction of a plurality of diagnosed patients who have been diagnosed as having an infection. The first relative position of the sagittal infection region in each training sagittal image is labeled, for example, a professional physician can use labeling software to outline the sagittal infection region, and the boundary coordinates of the outlined sagittal infection region are obtained as the second relative position of the sagittal infection region.

[0081] The multiple training sagittal images are preprocessed to form a sequence. For example, after the number of layers, length and width of the multiple training sagittal images are unified, the multiple training sagittal images are respectively converted into sagittal image matrices with fixed size and dimension, and input into a convolutional network to extract training image features. The convolutional network can be ResNet50. The convolutional network outputs multiple sagittal image feature matrices, and a convolution kernel is used to reduce the dimension of the multiple sagittal image feature matrices respectively, to obtain multiple one-dimensional sagittal image feature sequences, and the multiple sagittal image feature sequences are input into a transformer encoder. In the transformer encoder, a fixed position encoding is added to each attention layer, each position in the sagittal image feature sequence interacts with other positions to establish a global dependency relationship, and a multi-head attention mechanism is used to calculate attention weights, so as to perform attention weighted average on the input sagittal image feature sequence. After multiple layer processing of the transformer encoder, multiple groups of sequences representing the global features of each training sagittal image are generated, that is, the third feature information of each training sagittal image.

[0082] The transformer decoder takes the third feature information as input, learns the relationship of the global features of the third feature information through a self-attention mechanism, learns the relationship between local features through a multi-head attention mechanism, and adds a clustering position encoding to each attention layer to cluster the local features, so that a feedforward network decodes the output of the transformer decoder into a third detection position and a class label of the sagittal infected area.

[0083] In addition, the transformer decoder takes the third feature information and the second feature information as input, that is, the third feature information is added to the cross-attention mechanism in the step of learning the second feature information, to complete the fusion of the two modal image feature information.

[0084] Preferably, the first loss function comprises:

[0085] The first loss function comprises a first classification loss function and a first positioning loss function. The first classification loss function is obtained by comparing any corresponding first relative position in the multiple transverse infected areas with the second detection position, and the first average absolute error is obtained as the first positioning loss function.

[0086] Preferably, the second loss function comprises:

[0087] The second loss function comprises a second classification loss function and a second positioning loss function, and the second classification loss function and the second positioning loss function are respectively obtained by comparing any corresponding second relative position in the plurality of sagittal infection regions with a third detection position, and the second classification loss function is obtained by taking the second cross-entropy as the second classification loss function, and the second positioning loss function is obtained by taking the second mean absolute error as the second positioning loss function.

[0088] Specifically, the first loss function and the second loss function are recorded respectively, and the infection detection model is repeatedly trained until the first loss function and the second loss function converge, so that the infection detection model can sufficiently learn the feature information of the training transverse plane image, and learn the fusion of the feature information of the training sagittal plane image and the feature information of the training transverse plane image. The classification loss function refers to a loss function used to measure the classification detection result, and the cross-entropy can be used. The positioning loss function refers to a loss function used to measure the position detection result, and the mean absolute error can be used. If the detection position output by the infection detection model is inconsistent with the actual position, a positioning loss will be generated. The first classification loss function and the first positioning loss function are weighted and added to obtain the first loss function, and the second classification loss function and the second positioning loss function are weighted and added to obtain the second loss function.

[0089] In addition, the infection detection model can also introduce a position encoding loss GIoU Loss for punishing position encoding errors. The inconsistency of the target position encoding is represented, so as to further improve the detection accuracy.

[0090] The application also discloses an infection detection device, comprising:

[0091] A preprocessing module is configured to obtain a to-be-detected transverse plane image corresponding to a to-be-detected part, and input the to-be-detected transverse plane image into a pre-trained infection detection model, wherein the infection detection model comprises a convolutional network, a transformer encoder, a transformer decoder and a feedforward network.

[0092] A feature extraction module is configured to obtain first feature information of the to-be-detected transverse plane image by using the convolutional network and the transformer encoder.

[0093] A judgment module is configured to judge whether an actual infection region exists in the to-be-detected transverse plane image according to the first feature information by using the transformer decoder and the feedforward network.

[0094] A position detection module is configured to further output a first detection position of the actual infection region in the to-be-detected transverse plane image by using the feedforward network when the judgment result is that the actual infection region exists in the to-be-detected transverse plane image.

[0095] The application further discloses an electronic device, which comprises a memory storing computer executable instructions and a processor, and when the instructions are executed by the processor, the electronic device implements the infection detection method.

[0096] The application further discloses a computer readable storage medium, which stores a computer program, and when the instructions are executed on a computer, the computer executes the infection detection method.

[0097] It should be noted that the embodiments of the present application have better implementation, and do not limit the present application in any form, and any skilled person in the art can change or modify the above disclosed technical content into equivalent effective embodiments, as long as the content is not deviated from the technical solution of the present application, and any modification or equivalent change and modification of the above embodiments according to the technical essence of the present application still belongs to the scope of the technical solution of the present application.

Claims

1. An infection detection method, characterized in that, Includes the following steps: The cross-sectional image corresponding to the part to be inspected is obtained, and the cross-sectional image is input into a pre-trained infection detection model, wherein the infection detection model includes a convolutional network, a transformer encoder, a transformer decoder, and a feedforward network. The first feature information of the transverse section image to be detected is obtained by using the convolutional network and the transformer encoder. The transformer decoder and the feedforward network are used to determine whether there is an actual infected area in the transverse section image to be inspected based on the first feature information. When the determination result indicates that there is an actual infected area in the transverse section image to be inspected, the feedforward network is used to further output the first detection position where the actual infected area is located in the transverse section image to be inspected.

2. The infection detection method according to claim 1, characterized in that, The acquisition of the first feature information of the transverse section image to be inspected includes: The convolutional network is used to obtain the image feature sequence of the transverse section image to be detected; The image feature sequence is positionally encoded and attention-weighted averaged using the transformer encoder to obtain the global features of the transverse section image to be inspected, which serve as the first feature information of the transverse section image to be inspected.

3. The infection detection method according to claim 1, characterized in that, The determination of whether there is an actual infected area within the transverse section image to be inspected includes: Using the transformer encoder, based on the first feature information, multiple local features within the transverse section image to be detected are aggregated to obtain the relationships between the multiple local features. Using the feedforward network, based on the relationship between multiple local features in the cross-sectional image to be inspected, it is determined whether there is an actual infected area in the cross-sectional image to be inspected, and the first detection position of the actual infected area in the cross-sectional image to be inspected is obtained.

4. The infection detection method according to claim 1, characterized in that, The infection detection model was pre-trained through the following steps: A plurality of training cross-sectional images and an infection detection model to be trained are obtained. A cross-sectional infection region exists in any of the plurality of training cross-sectional images. A first relative position of any of the cross-sectional infection regions located in its corresponding training cross-sectional image is obtained. The plurality of training cross-sectional images are input into the infection detection model to be trained. The convolutional network and the transformer encoder are used to obtain the second feature information of any one of the plurality of training transverse images; The transformer decoder and the feedforward network are used to obtain any second detection position in any of the multiple transverse infection regions located within its corresponding training transverse image based on the multiple second feature information; the first relative position and the second detection position corresponding to any of the multiple transverse infection regions are compared respectively to obtain a first loss function, and backpropagation is performed to optimize the parameters of the infection detection model to be trained; Repeat the above steps until the first loss function converges to obtain the infection detection model.

5. The infection detection method according to claim 4, characterized in that, Also includes: Obtain multiple training sagittal images that correspond one-to-one with the multiple training transverse images. A sagittal infection region exists within any of the multiple training sagittal images. Obtain the second relative position of any of the multiple sagittal infection regions located within its corresponding training sagittal image. Input the multiple training sagittal images into the infection detection model to be trained. The convolutional network and the transformer encoder obtain the third feature information of any one of the plurality of training sagittal images, and input the plurality of third feature information and the plurality of second feature information into the transformer decoder in a one-to-one correspondence. The transformer decoder and the feedforward network obtain any second detection position within the corresponding training transverse image of the multiple transverse infection regions based on the multiple second feature information and the multiple third feature information. The transformer decoder and the feedforward network obtain any third detection position in the multiple sagittal infection regions located within their corresponding training sagittal images based on the multiple third feature information. A first loss function is obtained by comparing the first relative position and the second detection position corresponding to any one of the multiple transverse infection regions, and a second loss function is obtained by comparing the second relative position and the third detection position corresponding to any one of the multiple sagittal infection regions. Backpropagation is then performed to optimize the parameters of the infection detection model to be trained. Repeat the above steps until both the first loss function and the second loss function converge to obtain the infection detection model.

6. The infection detection method according to claim 4, characterized in that, The process of obtaining the first loss function includes: The first loss function includes a first classification loss function and a first localization loss function. The first relative position and the second detection position are compared respectively in any of the multiple transverse infection regions to obtain the first cross-entropy as the first classification loss function and the first mean absolute error as the first localization loss function.

7. The infection detection method according to claim 5, characterized in that, The process of obtaining the second loss function includes: The second loss function includes a second classification loss function and a second localization loss function. The second relative position corresponding to any one of the plurality of sagittal infection regions is compared with the third detection position to obtain the second cross-entropy as the second classification loss function, and the second mean absolute error is obtained as the second localization loss function.

8. An infection detection device, characterized in that, include: The preprocessing module is used to acquire the cross-sectional image corresponding to the part to be inspected, and input the cross-sectional image to be inspected into a pre-trained infection detection model, wherein the infection detection model includes a convolutional network, a transformer encoder, a transformer decoder, and a feedforward network. The feature extraction module is used to obtain the first feature information of the transverse section image to be detected by using the convolutional network and the transformer encoder; The judgment module is used to use the transformer decoder and the feedforward network to determine whether there is an actual infected area in the transverse section image to be inspected based on the first feature information; The location detection module is used to further output the first detection position of the actual infected area in the cross-sectional image under test when the judgment result is that there is an actual infected area in the cross-sectional image under test using the feedforward network.

9. An electronic device, characterized in that, The electronic device includes a memory storing computer-executable instructions and a processor, which, when executed by the processor, cause the electronic device to perform the infection detection method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the instructions are executed on a computer, the computer performs the infection detection method according to any one of claims 1-7.

Images