A method and system for constructing an adenoid marker point detection model

Abstract

The application discloses a kind of construction methods and systems of adenoid mark point detection model, it includes the following steps: obtaining training sample set, the training sample set includes image file and corresponding label file containing real mark point;Local self-attention mechanism based neural network model is constructed, and local self-attention mechanism is used to capture long-distance dependence relationship;Training sample set is input into neural network model to carry out supervised training, and obtain adenoid mark point detection model;Using the above technical solution makes the mark point detection model finally obtained can accurately locate and detect mark point in nasopharynx lateral X-ray film, without manual determination of doctor, reduces the calculation deviation of A / N value, improves the accuracy of evaluation result;It has smaller computational complexity simultaneously, improves the accuracy and robustness of model to mark point, provides valuable auxiliary information for subsequent medical analysis and diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for constructing an adenoid marker detection model. Background Technology

[0002] The adenoids are a mass of lymphatic tissue located on the posterior wall of the nasopharynx. Physiological enlargement of the adenoids can occur between the ages of 2 and 10, but local infections and inflammation can cause pathological enlargement. In children, enlarged adenoids encroach on the airway, resulting in a relatively small nasopharynx and an increased frequency of upper respiratory tract infections. Adenoid hypertrophy can also cause other problems, including maxillofacial developmental abnormalities, excessive daytime sleepiness, impaired cognitive function, and poor academic performance. Depending on the child's symptoms and disease progression, mild cases can be managed with conservative symptomatic treatment to suppress further enlargement and allow the adenoids to shrink on their own. However, severe cases may require surgical removal. Adenoid hypertrophy poses a significant threat to children's health, and its incidence has increased significantly in recent years. Timely detection and diagnosis are crucial for early treatment and controlling the disease's progression.

[0003] Traditional diagnostic methods for adenoid hypertrophy currently rely primarily on lateral nasopharyngeal radiographs and flexible nasopharyngoscopy. While nasopharyngoscopy allows direct observation of the adenoids, the procedure is painful and limits its clinical application. Lateral nasopharyngeal radiographs, as a simple, economical, and routine examination, have become the most commonly used method for patients with adenoid hypertrophy. However, this procedure requires radiologists to identify markers on the lateral nasopharyngeal radiograph to measure the A / N ratio. This process is time-consuming and cumbersome. Furthermore, the complex structures of the nasopharynx and skull base make the identification of markers highly dependent on the physician's experience. Different physicians have varying sensitivities to grayscale differences in lateral cephalometric radiographs, leading to a lack of objectivity in the judgment of measurement boundaries. This results in deviations in the calculation of the A / N ratio, affecting the accuracy of the diagnostic results. Summary of the Invention

[0004] Therefore, it is necessary to provide a method for constructing an adenoid marker detection model to address the problem that the determination of markers on nasopharyngeal lateral X-ray images is highly dependent on the doctor's experience, which affects the accuracy of adenoid assessment results.

[0005] To achieve the above objectives, the present invention provides a method for constructing an adenoid marker detection model, which includes the following steps:

[0006] Obtain a training sample set, which includes image files and corresponding label files containing real-world marker points;

[0007] A neural network model based on a local self-attention mechanism is constructed, and parameters such as loss function, optimizer, learning rate, and maximum number of training iterations are set; the local self-attention mechanism is used to capture long-distance dependencies.

[0008] The training sample set is input into the neural network model for supervised training. In each epoch, the image files in the training sample set are input into the neural network model, and the output is a segmentation map of the marker points in 4 channels. The loss is calculated between the map and the corresponding label file. The gradient of the loss function is calculated by the optimizer to minimize the loss value. The model weights corresponding to the minimum loss in the entire training process are recorded and the neural network model is updated to obtain the adenoid marker point detection model.

[0009] Furthermore, obtaining the training sample set includes the following steps:

[0010] Batch acquisition of nasopharyngeal lateral X-ray images and corresponding label files containing real marker points for preprocessing;

[0011] Convert nasopharyngeal lateral X-ray images into image file formats;

[0012] Convert the tag file containing real markers into a tag file in a data exchange format containing real markers;

[0013] The converted image files and their corresponding label files are matched one-to-one to construct a training sample set.

[0014] Furthermore, the preprocessing includes image scaling.

[0015] Furthermore, the neural network model includes:

[0016] The encoder module is used to perform feature extraction processing on the training dataset to generate local feature vectors for each local image;

[0017] The decoder module is used to map feature maps back to the original image size and capture long-range dependencies through a local attention mechanism.

[0018] Furthermore, the feature extraction process is a step of generating local feature vectors for each local image, including: performing overall feature extraction processing on the training dataset to generate a feature map of the image file at a certain training step; performing multi-scale sampling and attention processing on the feature map to generate multi-scale attention features; and performing information fusion and post-processing on the multi-scale features to generate a feature vector corresponding to each local image.

[0019] Furthermore, the core calculation formula of the local attention mechanism, namely

[0020]

[0021] Here, Q, K, and V are vectors mapped from the attention mechanism input through three mapping matrices, respectively. Representing the spatial dimension of X, the output of Softmax is a vector with the same length as V, whose elements sum to 1. Multiplying it by V yields the attention result.

[0022] Furthermore, the neural network model is the AdeNet model.

[0023] An automated inference method for adenoid marker points, applying the aforementioned method for constructing an adenoid marker point detection model, includes the following steps:

[0024] Preprocess the nasopharyngeal lateral X-ray image to be analyzed;

[0025] Convert the preprocessed nasopharyngeal lateral X-ray film into an image file format to be inferred;

[0026] Input the image file to be inferred into the adenoid marker detection model, and the adenoid marker detection model will output a 4-channel segmentation map;

[0027] The extreme points in each channel are calculated using the argmax function, and the coordinates of the extreme points are the coordinates of the marker points.

[0028] Use the OpenCV extension package to mark the coordinates of four points onto the image file, and obtain an image file with the marked points.

[0029] An automated method for assessing adenoids, employing the aforementioned automated adenoid marker inference method, includes the following steps:

[0030] The coordinates of four marker points were obtained using an automated inference method for adenoid marker points.

[0031] The adenoid A / N ratio was calculated using the coordinates of four marked points.

[0032] Determine the range of the A / N ratio;

[0033] If the adenoid A / N ratio is ∈ [0, 0.60], then the adenoids are normal;

[0034] If the adenoid A / N ratio is ∈ [0.61, 0.70], then the adenoids are moderately hypertrophic.

[0035] If the adenoid A / N ratio is ∈ [0.71, 1], then the adenoids are pathologically enlarged.

[0036] A system for automatically assessing adenoids, comprising the steps of the above-described method for automatically assessing adenoids.

[0037] Unlike existing technologies, the above-mentioned technical solution constructs a training sample set using image files and corresponding label files containing real markers. This training sample set is then input into a neural network model based on a local self-attention mechanism for training. The resulting marker detection model can automatically and accurately locate and detect markers in nasopharyngeal lateral X-rays without requiring manual determination by the physician. This ensures objectivity in marker determination, reduces the calculation bias of the A / N value, and improves the accuracy of adenoid assessment. Furthermore, the adenoid marker detection model can capture long-distance dependencies through a self-attention mechanism while maintaining low computational complexity, thus improving the model's accuracy and robustness in identifying markers and providing valuable auxiliary information for subsequent medical analysis and diagnosis. Attached Figure Description

[0038] Figure 1 This is a schematic flowchart illustrating the method for constructing the adenoid marker detection model of the present invention;

[0039] Figure 2 The image file to be inferred in the example;

[0040] Figure 3 To illustrate, an image file with marker points was obtained;

[0041] Figure 4 This is a schematic diagram illustrating the evaluation results obtained in an example. Detailed Implementation

[0042] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0043] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0044] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0045] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0046] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0047] Unless otherwise specified, the use of terms such as “comprising,” “including,” “having,” or other similar expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0048] Similar to the interpretation in the Patent Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also interpreted in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0049] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0050] Unless otherwise expressly specified or limited, the terms "installation," "connection," "linking," "fixing," and "setting," as used in the description of the embodiments of this application, should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral setting; it can be a mechanical connection, an electrical connection, or a communication connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two components or the interaction between two components. For those skilled in the art to which this application pertains, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0051] See Figures 1-4 As shown, this invention provides a method for constructing an adenoid marker detection model. A training sample set is constructed using image files and corresponding label files containing real markers. This training sample set is then input into a neural network model based on a local self-attention mechanism for training. The resulting marker detection model can automatically and accurately locate and detect markers in nasopharyngeal lateral X-ray images without manual determination by the physician. This ensures objectivity in marker determination, reduces the calculation bias of the A / N value, and improves the accuracy of adenoid size assessment. The adenoid marker detection model can capture long-distance dependencies through a self-attention mechanism while possessing low computational complexity, improving the model's accuracy and robustness in detecting markers and providing valuable auxiliary information for subsequent medical analysis and diagnosis.

[0052] The following provides a detailed description of a method for constructing an adenoid marker detection model provided by this invention. See also... Figure 1 As shown, a method for constructing an adenoid marker detection model includes the following steps:

[0053] Step S1: Obtain a training sample set, which includes image files and corresponding label files containing real marker points;

[0054] Step S2: Construct a neural network model based on a local self-attention mechanism, and set parameters such as loss function, optimizer, learning rate, and maximum number of training iterations; the local self-attention mechanism is used to capture long-distance dependencies;

[0055] Step S3: Input the training sample set into the neural network model for supervised training. In each epoch, input the image files in the training sample set into the neural network model, output the label point segmentation map of 4 channels, calculate the loss between the map and the corresponding label file, calculate the gradient of the loss function through the optimizer, thereby minimizing the loss value, record the model weight corresponding to the minimum loss in the entire training process, update the neural network model, and obtain the adenoid label point detection model.

[0056] The label file containing real markers mentioned above is an image file with real markers. The loss function mentioned above measures the difference between the neural network model output and the label file, and can use mean squared error (MSE) or cross-entropy loss function, etc. The optimizer and learning rate are used to minimize the loss function and adjust the neural network model parameters. The optimizer can be Adam, SGD, etc. The gradient calculation mentioned above can use methods such as batch gradient descent or stochastic gradient descent. The marker segmentation map of the four output channels mentioned above corresponds to the four markers on the nasopharyngeal lateral X-ray.

[0057] Step S1, which involves obtaining the training sample set, includes the following steps:

[0058] Batch acquisition of nasopharyngeal lateral X-ray images and corresponding label files containing real marker points for preprocessing;

[0059] Convert nasopharyngeal lateral X-ray images into image file formats;

[0060] Convert the tag file containing real markers into a tag file in a data exchange format containing real markers;

[0061] The converted image files and their corresponding label files are matched one-to-one to construct a training sample set.

[0062] The aforementioned nasopharyngeal lateral X-ray images are obtained from image diagnostic equipment, such as X-ray equipment. The aforementioned tag file containing real marker points is the nasopharyngeal lateral X-ray image with real marker points. The aforementioned preprocessing refers to basic image processing such as image enhancement, denoising, and grayscale conversion performed on the nasopharyngeal lateral X-ray image and the corresponding tag file containing real marker points; the preprocessing also includes image scaling, which can reduce the size of the image, thereby reducing the dimensionality of the input features, reducing the parameters that the neural network model needs to process, reducing the computational load and complexity of the neural network model, and thus improving the inference speed of the neural network model. The aforementioned image scaling includes proportional scaling, non-proportional scaling, etc., and can be implemented using methods such as image pyramid, bilinear interpolation, nearest neighbor interpolation, and bicubic interpolation. Image scaling can be implemented using functions and interfaces provided by image processing libraries (such as OpenCV, PIL) or deep learning frameworks (such as TensorFlow, PyTorch). The aforementioned conversion of nasopharyngeal lateral X-ray images into image file formats refers to converting nasopharyngeal lateral X-ray images from DICOM format to common image file formats. These image file formats are those used for storing and transmitting image data, such as PNG, JPG, GIF, BMP, TIFF, etc., with JPG format being preferred as it significantly reduces file size while maintaining image quality. The aforementioned conversion of tag files containing real markers into tag files in a data exchange format containing real markers refers to converting tag files containing real markers from DICOM format to a data exchange format. A data exchange format is a format used for transmitting and sharing data between different systems, such as JSON, XML, CSV, etc.; JSON format is preferred as it is a lightweight data exchange format commonly used to transfer data from one program or system to another. It is based on a subset of JavaScript syntax, making it easy to understand and use, and possesses good readability and scalability.

[0063] The neural network model described above includes:

[0064] The encoder module is used to perform feature extraction processing on the training dataset to generate local feature vectors for each local image;

[0065] The decoder module is used to map feature maps back to the original image size and capture long-range dependencies through a local attention mechanism.

[0066] The encoder module typically consists of multiple convolutional and pooling layers, used to extract image features layer by layer. By progressively reducing the image size, the encoder module generates local feature vectors for each local image, containing both low-level and high-level feature information. The encoder module can effectively extract features at various scales and levels from the input image, which helps the model better understand the semantic information of the image. The decoder module, on the other hand, maps the feature maps back to the original image size through transposed convolutional layers or other upsampling operations, thereby restoring the spatial resolution of the segmentation result. This helps generate accurate segmentation boundaries and details, and, combined with a local attention mechanism, captures long-range dependencies, improving the accuracy and coherence of the segmentation results.

[0067] Specifically, the feature extraction process is a step of generating local feature vectors for each local image, including: performing overall feature extraction processing on the training dataset to generate a feature map of the image file at a certain training step; performing multi-scale sampling and attention processing on the feature map to generate multi-scale attention features; and performing information fusion and post-processing on the multi-scale features to generate a feature vector corresponding to each local image.

[0068] Multi-scale refers to signal sampling at different scales, allowing for the observation of different features at each scale. Multi-scale is not the same as multi-resolution, which simply refers to the size of the image. Multi-scale can be understood as the different features perceived by a person when the distance to a fixed object varies; that is, when an object is imaged at different sizes in the field of view (different scales), its features also differ. Of course, multi-resolution can achieve multi-scale in certain specific situations. In neural network models, multi-scale can specifically manifest as scaling the output feature maps of different convolutional layers to a uniform size, so that they contain both global information and local detail information. Different scales of images are suitable for different tasks. If the image task is simple, such as determining whether the original image is a solid color or has a foreground, a small-scale image is sufficient. If the image task is of medium difficulty, such as requiring image classification, at least a medium-scale image is needed. If the image task is difficult, such as requiring semantic segmentation or image description, large-scale images are necessary for good results. Multi-scale feature generation and feature vector generation can further improve the expressive power and robustness of features, enhance model performance and effectiveness, and thus adapt to variations in patient head size.

[0069] Specifically, the core calculation formula of the local attention mechanism is,

[0070]

[0071] Here, Q, K, and V are vectors mapped from the attention mechanism input through three mapping matrices, respectively. Representing the spatial dimension of X, the output of Softmax is a vector with the same length as V, whose elements sum to 1. Multiplying it by V yields the attention result.

[0072] Local attention calculates the result of the above formula within a single window. It improves the local attention calculation mechanism and combines it with a multi-scale mechanism to expand the attention calculation window from a single scale to multiple scales. This adds contextual information with different resolutions and features to the model, enriches the model's feature representation ability, and improves the model's generalization ability.

[0073] The neural network model described above is preferably the AdeNet model. In the AdeNet model, the encoder module consists of multiple dense blocks, each of which comprises multiple convolutional layers, attention layers, and layer normalization. The encoder module connects the output of the previous layer with the input of the current layer through dense connections to facilitate information transfer and feature reuse. The decoder module can consist of upsampling operations and the same dense blocks, used to map the features from the encoder module back to the original image size. In the decoder, the AdeNet model introduces a local attention mechanism, which learns to generate pixel-level weight maps to weight features, thereby capturing long-range dependencies and processing details in the image. The AdeNet model effectively extracts features through the characteristics of the U-Net architecture. The introduced local attention mechanism can weight features according to the pixel-level weight maps, enabling the model to better focus on useful information in the image and suppress noise. This helps to capture long-range dependencies and process details in the image, improving the accuracy of the segmentation results while maintaining high computational efficiency.

[0074] See Figures 2-3 As shown, an automatic inference method for adenoid marker points, applying the above-mentioned method for constructing an adenoid marker point detection model, includes the following steps:

[0075] Preprocess the nasopharyngeal lateral X-ray image to be analyzed;

[0076] Convert the preprocessed nasopharyngeal lateral X-ray film into an image file format to be inferred;

[0077] Input the image file to be inferred into the adenoid marker detection model, and the adenoid marker detection model will output a 4-channel segmentation map;

[0078] The extreme points in each channel are calculated using the argmax function, and the coordinates of the extreme points are the coordinates of the marker points.

[0079] Use the OpenCV extension package to mark the coordinates of four points onto the image file, and obtain an image file with the marked points.

[0080] After inputting the image file to be inferred into the trained adenoid marker detection model, the model outputs an image file with markers. The location of the markers varies from patient to patient, but they are all key points for assessing adenoid hypertrophy and can help doctors make specific diagnostic results. The whole process does not require doctors to manually mark the markers; they only need to briefly confirm the output detection image, which greatly improves the efficiency of clinicians in receiving patients.

[0081] An automated method for assessing adenoids, employing the aforementioned automated adenoid marker inference method, includes the following steps:

[0082] The coordinates of four marker points were obtained using an automated inference method for adenoid marker points.

[0083] The adenoid A / N ratio was calculated using the coordinates of four marked points.

[0084] Determine the range of the A / N ratio;

[0085] If the adenoid A / N ratio is ∈ [0, 0.60], then the adenoids are normal;

[0086] If the adenoid A / N ratio is ∈ [0.61, 0.70], then the adenoids are moderately hypertrophic.

[0087] If the adenoid A / N ratio is ∈ [0.71, 1], then the adenoids are pathologically enlarged.

[0088] In this context, A refers to the vertical distance from the most prominent point of the adenoid in the nasopharynx to the tangent line of the skull base, representing the maximum thickness of the adenoid. N is the distance within the plane of A, from the tangent line of the skull to the posterior edge of the soft palate, representing the width of the nasopharynx at the most prominent part of the adenoid. A / N is the ratio of these two values, reflecting the maximum degree to which the adenoid tip obstructs the space between the anterior and posterior nasopharynx. Based on a given range, the A / N ratio of the adenoid can be classified. According to the range of the A / N ratio, adenoids are divided into three categories: normal, moderately enlarged, and pathologically enlarged. By using an automated method to infer adenoid marker points and calculate the A / N ratio, rapid and accurate assessment and analysis of the adenoid can be achieved. This can help doctors understand the degree of adenoid hypertrophy in clinical diagnosis and research, thereby guiding treatment and further research. Furthermore, automated methods can reduce the influence of manual operation and subjectivity, improving the consistency and reproducibility of results.

[0089] See Figures 2-4 As shown, in order to further understand the content of the present invention, the present invention provides a system for automatically assessing adenoids, and the steps of the above method.

[0090] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method. The computer program involved in the embodiments can be stored in a computer device readable storage medium, including but not limited to disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures capable of performing similar or equivalent functions to the storage media listed above, such as DNA, RNA, proteins, etc., which have information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-described medium types or a combination of the above-described medium types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium or distributed across multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into a device or connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0091] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0092] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of the present invention, or equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of patent protection of the present invention.

Claims

1. A method for constructing an adenoid marker point detection model, characterized in that, The method comprises the following steps: obtaining a training sample set comprising image files and corresponding label files containing real marker points; constructing a neural network model based on a local self-attention mechanism, setting a loss function, an optimizer, a learning rate, and a maximum number of training iterations; the neural network model is an AdeNet model; the AdeNet model comprises an encoder module and a decoder module, the encoder module is composed of multiple dense blocks, each dense block is composed of multiple convolutional layers, attention layers, and layer normalization; the encoder module connects the output of the previous layer with the input of the current layer through dense connection to facilitate information transmission and feature reuse; the decoder module is composed of upsampling operations and the same dense blocks, which is used to map the features of the encoder module back to the original image size; the decoder also introduces a local attention mechanism to generate a pixel-level weight map to weight the features to capture long-distance dependencies and handle details in the image; input the training sample set into the neural network model for supervised training; in each epoch, input the image files in the training sample set into the neural network model, output a 4-channel marker segmentation map, calculate the loss between the output and the corresponding label, calculate the gradient for the loss function through the optimizer, thereby minimizing the loss value, record the model weight corresponding to the minimum loss in the entire training process, and update the neural network model to obtain an adenoid marker detection model.

2. The method of claim 1, wherein the method further comprises: determining a plurality of adenoid marker points on the 3D model of the patient's adenoid tissue; and determining a plurality of adenoid marker points on the 3D model of the patient's adenoid tissue. The method comprises the following steps: batch obtaining nasopharyngeal lateral X-ray films and corresponding label files containing real marker points for preprocessing; convert the nasopharyngeal lateral X-ray films into image files in image file format; convert the label files containing real marker points into label files in data interchange format containing real marker points; corresponding label files to construct a training sample set.

3. The method of claim 2, wherein the method further comprises: determining a plurality of adenoid marker points on the 3D model of the patient's adenoid tissue; and determining a plurality of adenoid marker points on the 3D model of the patient's adenoid tissue. The preprocessing includes image scaling processing.

4. The method for constructing an adenoid marker detection model according to claim 1, wherein the encoder module is used for feature extraction processing of the training sample set to generate a local feature vector of each local image.

5. The method of claim 4, wherein the method further comprises: determining a plurality of adenoid marker points on the 3D model of the patient's adenoid tissue; and determining a plurality of adenoid marker points on the 3D model of the patient's adenoid tissue. The feature extraction processing is a step of generating a local feature vector of each local image, comprising: performing overall feature extraction processing on the training sample set to generate a feature map of the image file at a certain training step; performing multi-scale sampling and attention processing on the feature map to generate multi-scale attention features; and performing information fusion and post-processing on the multi-scale attention features to generate a feature vector corresponding to each local image.

6. The method of claim 4, wherein the method further comprises: The core calculation formula of the local attention mechanism is ； Where Q, K and V are vectors mapped from the input of attention mechanism by three mapping matrices respectively, The spatial dimension of X-ray image, the result of Softmax output is a vector with the same length as V, the sum of each element is 1, and the multiplication of V gets the attention result.

7. An automatic reasoning adenoid marker point method, applying the construction method of the adenoid marker point detection model of any one of claims 1-6, characterized in that, The method comprises the following steps: preprocessing the nasopharyngeal lateral X-ray film to be inferred; convert the preprocessed nasopharyngeal lateral X-ray film into an inference image file in image file format; input the inference image file into the adenoid marker detection model, which will output a 4-channel marker segmentation map; calculate the extreme points in each channel by using the argmax function, and the coordinates of the extreme points are the coordinates of the marker points; The coordinates of the four marker points are marked into the original image file by using the Opencv expansion package to obtain an image file with the marker points.

8. A method for automatically assessing adenoids, using the method for automatically inferring adenoid landmark points according to claim 7, characterized in that, The method comprises the following steps: The coordinates of the four marker points obtained by the automatic reasoning adenoid marker point method; The A / N ratio of the adenoid is calculated by the coordinates of the four marker points; The range of the A / N ratio is determined; If the A / N ratio of the adenoid is ∈ [0, 0.60], the adenoid is normal; If the A / N ratio of the adenoid is ∈ [0.61, 0.70], the adenoid is moderately hypertrophic; If the A / N ratio of the adenoid is ∈ [0.71, 1], the adenoid is pathologically hypertrophic.

9. A system for automatically assessing adenoids, characterized by: The steps of the method for automatically evaluating the adenoid in claim 8 are implemented.

Images