Osteosarcoma CT image lesion area detection method and system based on feature correlation

Abstract

The present invention discloses a method and system for detecting lesion areas in osteosarcoma CT images based on feature correlation, which relates to image technology in the medical field. The method includes: obtaining osteosarcoma CT images and making a data set; constructing a target detection model based on feature correlation and using the data set for training, and using the trained target detection model to realize lesion area detection in osteosarcoma CT images. The target detection model is a convolutional neural network, and the Backbone part includes a number of alternating modules connected in sequence for extracting features; the Neck part uses a mutually fused FPN network and a PANET network to receive the output of the Backbone part and output multiple dimensional feature maps; the Head part receives the multiple dimensional feature maps output by the Neck part and outputs a detection image based on global information. The present invention uses a combination of knowledge distillation and target detection to optimize the problem of false detection and missed detection in osteosarcoma CT images, and improve the recognition rate and accuracy of detection.

Description

Technical Field

[0001] The present invention relates to the field of medical images, and in particular to a method and system for detecting lesion areas in osteosarcoma CT images based on feature correlation. Background Art

[0002] Osteosarcoma is a common primary malignant tumor in bone tissue. It has the highest incidence in children, adolescents, and the elderly, and has a bimodal distribution. The average survival time of advanced metastatic osteosarcoma is less than five years. Therefore, early diagnosis, detection of lesions, and treatment are of great significance to reducing the incidence of osteosarcoma and improving patient survival. Among them, CT image examination is widely used in the field of osteosarcoma CT image diagnosis and treatment due to its advantages such as non-invasiveness and fast diagnosis. The osteosarcoma CT image detection method based on deep learning uses medical image processing technology and combines computer analysis and calculation to assist doctors in discovering and diagnosing lesions. It has the advantages of high accuracy, high resolution, and strong observability. It has important significance and application value in the clinical diagnosis and treatment of osteosarcoma.

[0003] In recent years, research on osteosarcoma lesion detection based on deep learning has made great progress. However, since the current imaging method of CT images is easily affected by artifacts, speckle noise, and echo disturbance, there are problems such as low image quality, large speckle noise, and low contrast, resulting in low distinction between the foreground and background of the image, and it is easily disturbed by individual subjective factors during diagnosis, which has a mixed effect on the diagnosis, leading to misdiagnosis and missed diagnosis, resulting in inaccurate detection of osteosarcoma CT images. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of inaccurate detection of osteosarcoma CT images in the prior art.

[0005] The technical solution adopted by the present invention to solve the technical problem is: to provide a method for detecting lesion areas in osteosarcoma CT images based on feature correlation, comprising the following steps:

[0006] Obtain grayscale images of the lesion area of ​​osteosarcoma CT images, preprocess and annotate the images to create a data set;

[0007] Build a target detection model based on feature correlation;

[0008] Use the data set to train the target detection model to obtain a trained target detection model;

[0009] The trained target detection model is used to detect the lesion area in osteosarcoma CT images;

[0010] The target detection model is a convolutional neural network. The Backbone part of the convolutional neural network includes a first unit and several alternating modules connected in sequence. The first unit and the alternating modules both include a CBS structure, which is used to extract features of the input osteosarcoma CT image multiple times; the Neck part of the convolutional neural network adopts a mutually fused FPN network and PANET network. The Neck part receives the features output by the Backbone part, obtains and outputs multiple dimensional feature maps; the Head part of the convolutional neural network adopts a CBS structure, receives the multiple dimensional feature maps output by the Neck part, obtains and outputs a detection image based on global information.

[0011] Preferably, the images are preprocessed and annotated to form a data set, and the preprocessing includes removing annotations from osteosarcoma CT images and enhancing image contrast.

[0012] Preferably, the enhanced image contrast is expressed as:

[0013] y * = C log [ y + 1 ] ;

[0014] in, is the value of the original image pixel, is the pixel value after enhancement processing. Represents a logarithmic function, and the constant C is used to ensure that the grayscale dynamic range of the transformed image falls within a reasonable grayscale range.

[0015] Preferably, the first unit uses two CBS structures to extract low-level features from the osteosarcoma CT image and introduces the SE attention mechanism. The process of extracting features by the first unit includes the following steps:

[0016] First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows:

[0017] ;

[0018] in, represents the global description vector of the cth channel, H and W represent the height and width of the feature map respectively, Indicates the input feature map in the cth channel pixel value;

[0019] Secondly, after obtaining the global description of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel, which is expressed as:

[0020] ;

[0021] in, represents the recalibrated channel weight vector, W1 and W2 represent the weight matrix of the fully connected layer, represents the Sigmoid activation function, ReLU activation function.

[0022] Finally, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The resulting image size is 320×320×64, expressed as:

[0023] ;

[0024] in, represents the feature map of the cth channel after recalibration by the SE module, represents the calibration weight of the cth channel, Represents the feature map of the cth channel of the input feature map.

[0025] Preferably, the alternating module includes an upper-layer alternating module, a middle-layer alternating module and a lower-layer alternating module connected in sequence; the upper-layer alternating module includes a C3 structure, a CBS structure, and a C3 structure connected in sequence; the middle-layer alternating module is a CBS structure; and the lower-layer alternating module includes a C3 structure, a CBS structure, a C3 structure and a fast spatial pyramid pooling structure connected in sequence.

[0026] Preferably, the process of obtaining multiple dimensional feature maps in the Neck part includes the following steps:

[0027] receiving the output of the upper alternating module, the output of the middle alternating module and the output of the lower alternating module;

[0028] The output of the lower alternating module is sequentially passed through the CBS structure and up-sampled, and then matched with the output of the middle alternating module for dimension splicing. The splicing result is then passed through the C3 structure and the CBS structure to obtain the first fusion feature. The first fusion feature is further up-sampled and matched with the output of the upper alternating module for dimension splicing. The splicing result is then passed through the C3 structure to obtain the second fusion feature, which is output to the Head part as the first layer output feature map.

[0029] The second fusion feature is matched with the first fusion feature after passing through the CBS structure, and then the dimension is spliced. The splicing result is then passed through the C3 structure to obtain the third fusion feature, which is output to the Head part as the second layer output feature map;

[0030] The third fusion feature is passed through the CBS structure and matched with the output of the lower alternating module after passing through the CBS structure, and then dimensionally spliced. The splicing result is then passed through the C3 structure to obtain the fourth fusion feature, which is output to the Head part as the third-layer output feature map.

[0031] Preferably, the method of using the data set to train the target detection model includes using a self-distillation module to receive features output by a Backbone component, constructing a multi-branch distillation path, and guiding the alternating module to obtain performance similar to that of its next alternating module.

[0032] Preferably, the process of the self-distillation module guiding the alternating module to obtain similar performance to the next alternating module comprises the following steps:

[0033] Get the output of the upper alternating module , a 1×1 adaptive convolution adjuster is used to adjust it, which is expressed as:

[0034] ;

[0035] in, represents the convolution regulator of the upper alternating module, represents the convolution parameters;

[0036] Get the output of the middle-level alternating module , a 1×1 adaptive convolution adjuster is used to adjust it, which is expressed as:

[0037] ;

[0038] in, represents the convolution regulator applied to the middle layer alternating module, represents the convolution parameters;

[0039] Get the output of the lower alternating module ,calculate and Between Norm and and Between norm, and construct the self-distillation loss, expressed as:

[0040] ;

[0041] Loss self i = 1 2 || F h int i − F guide || 2 , i ∈ [ 1 , 2 ] ;

[0042] in, represents the total self-distillation loss, represents the ith individual self-distillation loss term, express Norm.

[0043] Preferably, the target detection model is trained using the data set, and the overall loss of the training is It is expressed as:

[0044] ;

[0045] in, represents the confidence loss, represents the classification loss, represents the self-distillation loss, , , and Represent confidence weight, rectangular box weight, classification weight and self-distillation weight respectively; Denotes the GIoU loss, expressed as:

[0046] ;

[0047] Among them, A represents the predicted box, B represents the real box, C represents the minimum closed area containing A and B, and |·| represents the area.

[0048] The present invention also provides a system for detecting lesion areas in osteosarcoma CT images based on feature correlation, comprising:

[0049] The data set acquisition module is used to acquire the grayscale image of the lesion area of ​​the osteosarcoma CT image, pre-process the image and annotate it to make a data set;

[0050] Model building module, used to build a target detection model based on feature correlation;

[0051] The model training module is used to train the target detection model using the data set to obtain a trained target detection model;

[0052] The model application module is used to detect the lesion area of ​​CT images of osteosarcoma using the trained target detection model;

[0053] The target detection model is a convolutional neural network. The Backbone part of the convolutional neural network includes a first unit and a plurality of alternating modules connected in sequence. The first unit and the alternating modules both include a CBS structure, which is used to extract features of the input osteosarcoma CT image multiple times. The Neck part of the convolutional neural network adopts a mutually fused FPN network and PANET network. The Neck part receives the features output by the Backbone part, obtains and outputs multiple dimensional feature maps. The Head part of the convolutional neural network adopts a CBS structure, receives the multiple dimensional feature maps output by the Neck part, obtains and outputs a detection image based on global information.

[0054] The first unit uses two CBS structures to extract low-level features from osteosarcoma CT images and introduces the SE attention mechanism. The process of extracting features in the first unit includes the following steps:

[0055] First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows:

[0056] ;

[0057] in, represents the global description vector of the cth channel, H and W represent the height and width of the feature map respectively, Indicates the input feature map in the cth channel pixel value;

[0058] Secondly, after obtaining the global description of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel, which is expressed as:

[0059] ;

[0060] in, represents the recalibrated channel weight vector, W1 and W2 represent the weight matrix of the fully connected layer, represents the Sigmoid activation function, ReLU activation function.

[0061] Finally, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The resulting image size is 320×320×64, expressed as:

[0062] ;

[0063] in, represents the feature map of the cth channel after recalibration by the SE module, represents the calibration weight of the cth channel, Represents the feature map of the cth channel of the input feature map.

[0064] The present invention has the following beneficial effects:

[0065] (1) After initially extracting low-level features, the present invention introduces the SE attention mechanism module, which can adaptively recalibrate channel features, suppress the interference of irrelevant features, and optimize the allocation of computing resources, thereby improving the feature expression ability and detection accuracy of the model, and significantly improving its detection performance in the lesion area of ​​osteosarcoma CT images. It is more robust and efficient, which helps to improve clinical diagnosis results;

[0066] (2) The backbone of the present invention adopts a self-distillation module to improve the correlation of image feature information during model transmission, enhance the focus on osteosarcoma lesion areas, and improve the detection capability of osteosarcoma CT images by improving the robustness of the model in feature extraction;

[0067] (3) The present invention introduces knowledge distillation technology into the target detection model. Through multi-branch self-distillation loss, different modules are guided to obtain performance similar to that of the lower-level alternating modules from the feature level, thereby improving the correlation of image feature information during model transmission, reducing the impact of CT image quality on model performance, and optimizing the false detection and missed detection problems in osteosarcoma CT images.

[0068] (4) When constructing the overall loss function, the present invention introduces the GIoU loss function to replace the rectangular box loss function, providing the model with a more comprehensive overlap measurement, faster convergence speed, and more optimized bounding box adjustment performance, thereby improving the robustness of the model, better adapting to a variety of complex scenarios, and reducing the problem of false detection and missed detection of osteosarcoma CT images.

[0069] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A method step diagram of an embodiment of the present invention;

[0071] Figure 2 A schematic diagram of a model of an embodiment of the present invention;

[0072] Figure 3 4 is a system structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0073] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0074] S101, obtaining a grayscale image of the lesion area of ​​the osteosarcoma CT image, preprocessing the image and annotating it to produce a data set;

[0075] S102, constructing a target detection model based on feature correlation;

[0076] S103, training the target detection model using the data set to obtain a trained target detection model;

[0077] S104, using the trained target detection model to detect the lesion area of ​​the osteosarcoma CT image.

[0078] Specifically, in S101, the images are preprocessed and annotated to form a data set, and the preprocessing includes removing annotations of the osteosarcoma CT image and enhancing the image contrast.

[0079] Specifically, the enhanced image contrast is expressed as:

[0080] y * = C log [ y + 1 ] ;

[0081] in, is the value of the original image pixel, is the pixel value after enhancement processing. Represents a logarithmic function. The constant C is used to ensure that the grayscale dynamic range of the transformed image meets the requirements. Specifically, it means to ensure that the grayscale dynamic range of the transformed image is within a reasonable range so that the image can be displayed correctly and retain sufficient contrast and details. For an 8-bit grayscale image, the grayscale value of the processed image should be kept in the range of 0-255. The constant C adjusts the result of the logarithmic transformation so that the processed pixel value falls within this reasonable grayscale range.

[0082] See also Figure 2 As shown, it is a model schematic diagram of an embodiment of the present invention, wherein the target detection model is a convolutional neural network, and the Backbone part of the convolutional neural network includes a plurality of alternating modules connected in sequence for extracting features, and the alternating modules all include a CBS structure; the Backbone part of the convolutional neural network includes a first unit and a plurality of alternating modules connected in sequence.

[0083] Specifically, the first unit uses two CBS structures and SE attention to extract low-level features of osteosarcoma CT images, which specifically includes the following steps:

[0084] First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows:

[0085] ;

[0086] in, represents the global description vector of the cth channel, H and W represent the height and width of the feature map respectively, Indicates the input feature map in the cth channel pixel value;

[0087] Secondly, after obtaining the global description of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel. The specific calculation formula is as follows:

[0088] ;

[0089] in, represents the recalibrated channel weight vector, W1 and W2 represent the weight matrix of the fully connected layer, represents the Sigmoid activation function, ReLU activation function.

[0090] Finally, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The resulting image size is 320×320×64. The specific calculation formula is as follows:

[0091] ;

[0092] in, represents the feature map of the cth channel after recalibration by the SE module, represents the calibration weight of the cth channel, Represents the feature map of the cth channel of the input feature map.

[0093] Specifically, the plurality of alternating modules specifically include an upper alternating module, a middle alternating module and a lower alternating module connected in sequence; the upper alternating module includes a C3 structure, a CBS structure and a C3 structure connected in sequence; the middle alternating module is a CBS structure; the lower alternating module includes a C3 structure, a CBS structure, a C3 structure and a fast spatial pyramid pooling structure connected in sequence. The Neck part of the convolutional neural network adopts a mutually fused FPN network and a PANET network, the Neck part receives the features output by the Backbone part, obtains and outputs multiple dimensional feature maps; the Neck part of the convolutional neural network adopts a mutually fused FPN network and a PANET network, the Neck part receives the features output by the Backbone part, obtains and outputs multiple dimensional feature maps; the Head part of the convolutional neural network adopts a CBS structure, receives the multiple dimensional feature maps output by the Neck part, obtains and outputs a detection image based on global information.

[0094] Specifically, the process of obtaining multiple dimensional feature maps in the Neck part includes the following steps:

[0095] S201, receiving the output of the upper layer alternating module, the output of the middle layer alternating module and the output of the lower layer alternating module;

[0096] S202, the output of the lower alternating module is sequentially passed through the CBS structure and up-sampled, and then matched with the output of the middle alternating module, and then dimension splicing is performed. The splicing result is then sequentially passed through the C3 structure and the CBS structure to obtain the first fusion feature; the first fusion feature is further up-sampled, matched with the output of the upper alternating module, and then dimension splicing is performed. The splicing result is then passed through the C3 structure to obtain the second fusion feature, which is output to the Head part as the first layer output feature map;

[0097] S203, the second fusion feature is matched with the first fusion feature after passing through the CBS structure, and then the dimension is spliced. The splicing result is then passed through the C3 structure to obtain the third fusion feature, which is output to the Head part as the second layer output feature map;

[0098] S204, the third fusion feature is passed through the CBS structure and matched with the output of the lower alternating module through the CBS structure, and then dimensionally spliced. The splicing result is then passed through the C3 structure to obtain the fourth fusion feature, which is output to the Head part as the third layer output feature map.

[0099] Specifically, the target detection model further includes a self-distillation module, which receives features output by the Backbone part, constructs a multi-branch distillation path, and guides the alternating module to obtain performance similar to that of its next alternating module, including the following steps:

[0100] S301, obtain the output of the upper alternating module ,at this time The size of is (80×80×256), and it is adjusted using a 1×1 adaptive convolution adjuster, expressed as:

[0101] ;

[0102] in, represents the convolution regulator applied to the upper alternating module, Represents the convolution parameters, adjusted The size is (20×20×1024).

[0103] S302, obtain the output of the middle layer alternating module ,at this time The size of is (40×40×512), and it is adjusted using a 1×1 adaptive convolution adjuster, expressed as:

[0104] ;

[0105] in, represents the convolution regulator applied to the middle layer alternating module, Represents the convolution parameters, adjusted The size is (20×20×1024).

[0106] S303, obtaining the output of the lower alternating module (size is 20×20×1024), calculate and Between Norm and and Between norm, and construct the self-distillation loss, expressed as:

[0107] ;

[0108] Loss self i = 1 2 || F h int i − F guide || 2 , i ∈ [ 1 , 2 ] ;

[0109] in, represents the total self-distillation loss, Represents each individual self-distillation loss term, which is the output of the current layer With the feature output of the i-th previous layer The L2 norm difference between ; express Norm.

[0110] By calculating the self-distillation loss between the output of the current layer and the feature output of the previous layer, and adding it to the total loss function for optimization, the output of the feature extraction module will tend to be consistent. This can guide the feature extraction modules of each layer to obtain similar performance, thereby improving the overall performance of the model.

[0111] Specifically, the target detection model is trained using the data set, and the overall loss of the osteosarcoma CT image lesion area detection method based on global information is , and its calculation formula is as follows:

[0112] ;

[0113] in, represents the confidence loss, represents the rectangular box loss, represents the classification loss, represents the self-distillation loss, , , and They represent confidence weight, rectangular box weight, classification weight and self-distillation weight respectively.

[0114] Furthermore, in the overall loss function GIoU loss is introduced in the calculation stage, and the original rectangular box loss The calculation formula becomes:

[0115] ;

[0116] Secondly, the GIoU loss is calculated according to the formula:

[0117] ;

[0118] Where A represents the predicted box, B represents the real box, C represents the minimum closed area containing A and B, and | · | represents the area;

[0119] Constructing an overall loss method for lesion region detection in osteosarcoma CT images based on global information , and its calculation formula is as follows:

[0120] ;

[0121] in, represents the confidence loss, represents the GIoU loss, represents the classification loss, Represents the loss from distillation.

[0122] See also Figure 3 FIG. 1 is a system structure diagram of an embodiment of the present invention, including:

[0123] The data set acquisition module 301 is used to acquire the grayscale image of the lesion area of ​​the osteosarcoma CT image, pre-process the image and annotate it to produce a data set;

[0124] A model building module 302 is used to build a target detection model based on feature correlation;

[0125] The model training module 303 is used to train the target detection model using the data set to obtain a trained target detection model;

[0126] A model application module 304 is used to implement lesion area detection of osteosarcoma CT images using the trained target detection model;

[0127] The present invention utilizes the combination of knowledge distillation and target detection to optimize the problems of false detection and missed detection in osteosarcoma CT images, thereby improving the recognition rate and accuracy of detection.

[0128] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting lesion areas in osteosarcoma CT images based on feature correlation, characterized in that: The following steps are involved: Obtain grayscale images of the lesion area of ​​osteosarcoma CT images, preprocess and annotate the images to create a data set; Build a target detection model based on feature correlation; Use the data set to train the target detection model to obtain a trained target detection model; The trained target detection model is used to detect the lesion area in osteosarcoma CT images; The target detection model is a convolutional neural network. The Backbone part of the convolutional neural network includes a first unit and a plurality of alternating modules connected in sequence. The first unit and the alternating modules both include a CBS structure, which is used to extract features of the input osteosarcoma CT image multiple times. The Neck part of the convolutional neural network adopts a mutually fused FPN network and PANET network. The Neck part receives the features output by the Backbone part, obtains and outputs multiple dimensional feature maps. The Head part of the convolutional neural network adopts a CBS structure, receives the multiple dimensional feature maps output by the Neck part, obtains and outputs a detection image based on global information. The first unit uses two CBS structures to extract low-level features from osteosarcoma CT images and introduces the SE attention mechanism. The process of extracting features in the first unit includes the following steps: First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows: Among them, z c represents the global description vector of the cth channel, H and W represent the height and width of the feature map respectively, x c (i,j) represents the (i,j)th pixel value of the cth channel of the input feature map; Secondly, after obtaining the global description of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel, which is expressed as: s c =σ(W2·δ(W1·z c )); Among them, s c represents the recalibrated channel weight vector, W1 and W2 represent the weight matrices of the fully connected layer, σ represents the Sigmoid activation function, and δ represents the ReLU activation function; Finally, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The resulting image size is 320×320×64, expressed as: in, represents the feature map of the cth channel after recalibration by the SE module, x c Represents the feature map of the cth channel of the input feature map; The object detection model is trained by using the data set, including using a self-distillation module to receive features of a backbone part of the output, constructing a multi-branch distillation path, and guiding the alternating module to obtain performance similar to that of its next alternating module; The process of the self-distillation module guiding the alternating module to obtain performance similar to that of the next alternating module comprises the following steps: Get the output of the upper alternating module A 1×1 adaptive convolution adjuster is used to adjust it, which is expressed as: in, represents the convolution regulator of the upper alternating module, represents the convolution parameters; Get the output of the middle-level alternating module A 1×1 adaptive convolution adjuster is used to adjust it, which is expressed as: in, represents the convolution regulator applied to the middle layer alternating module, represents the convolution parameters; Get the output F of the lower alternating module guide , calculate F guide and The L2 norm between guide and The L2 norm between , and construct the self-distillation loss, expressed as: Among them, Loss self represents the total self-distillation loss, represents the kth individual self-distillation loss term, and ||·||2 represents the L2 norm; The target detection model is trained using the data set, and the overall loss of the training is expressed as: Loss=αLoss obj +βLoss GIoU +λLoss clc +ηLoss self ; Among them, Loss obj Represents confidence loss, Loss clc Represents classification loss, Loss self represents the self-distillation loss, α, β, λ and η represent the confidence weight, rectangular box weight, classification weight and self-distillation weight respectively; Loss GIoU Denotes the GIoU loss, expressed as: Among them, A represents the predicted box, B represents the real box, C represents the minimum closed area containing A and B, and |·| represents the area.

2. The method for detecting lesion areas in osteosarcoma CT images based on feature correlation according to claim 1, characterized in that: The images are preprocessed and annotated to form a data set, and the preprocessing includes removing the annotations of the osteosarcoma CT image and enhancing the image contrast.

3. The method for detecting lesion areas in osteosarcoma CT images based on feature correlation according to claim 2, characterized in that: The enhanced image contrast is expressed as: and * =Zlog[y+1]; Where y is the value of the original image pixel, y* is the corresponding pixel value after enhancement, log represents the logarithmic function, and the constant Z is used to ensure that the grayscale dynamic range of the transformed image falls within a reasonable grayscale range.

4. The method for detecting lesion areas in osteosarcoma CT images based on feature correlation according to claim 1, characterized in that: The alternating module includes an upper alternating module, a middle alternating module and a lower alternating module connected in sequence; the upper alternating module includes a C3 structure, a CBS structure and a C3 structure connected in sequence; the middle alternating module is a CBS structure; the lower alternating module includes a C3 structure, a CBS structure, a C3 structure and a fast spatial pyramid pooling structure connected in sequence.

5. The method for detecting lesion areas in osteosarcoma CT images based on feature correlation according to claim 4, characterized in that: The process of obtaining multiple dimensional feature maps in the Neck part includes the following steps: receiving the output of the upper alternating module, the output of the middle alternating module and the output of the lower alternating module; The output of the lower alternating module is sequentially passed through the CBS structure and up-sampled, and then matched with the output of the middle alternating module for dimension splicing. The splicing result is then passed through the C3 structure and the CBS structure to obtain the first fusion feature. The first fusion feature is further up-sampled and matched with the output of the upper alternating module for dimension splicing. The splicing result is then passed through the C3 structure to obtain the second fusion feature, which is output to the Head part as the first layer output feature map. The second fusion feature is matched with the first fusion feature after passing through the CBS structure, and then the dimension is spliced. The splicing result is then passed through the C3 structure to obtain the third fusion feature, which is output to the Head part as the second layer output feature map; The third fusion feature is passed through the CBS structure and matched with the output of the lower alternating module after passing through the CBS structure, and then dimensionally spliced. The splicing result is then passed through the C3 structure to obtain the fourth fusion feature, which is output to the Head part as the third-layer output feature map.

6. A lesion area detection system for osteosarcoma CT images based on feature correlation, characterized in that: include: The data set acquisition module is used to acquire the grayscale image of the lesion area of ​​the osteosarcoma CT image, pre-process the image and annotate it to produce a data set; Model building module, used to build a target detection model based on feature correlation; The model training module is used to train the target detection model using the data set to obtain a trained target detection model; The model application module is used to detect the lesion area of ​​CT images of osteosarcoma using the trained target detection model; The target detection model is a convolutional neural network. The Backbone part of the convolutional neural network includes a first unit and a plurality of alternating modules connected in sequence. The first unit and the alternating modules both include a CBS structure, which is used to extract features of the input osteosarcoma CT image multiple times. The Neck part of the convolutional neural network adopts a mutually fused FPN network and PANET network. The Neck part receives the features output by the Backbone part, obtains and outputs multiple dimensional feature maps. The Head part of the convolutional neural network adopts a CBS structure, receives the multiple dimensional feature maps output by the Neck part, obtains and outputs a detection image based on global information. The first unit uses two CBS structures to extract low-level features from osteosarcoma CT images and introduces the SE attention mechanism. The process of extracting features in the first unit includes the following steps: First, the feature map of each channel is compressed into a scalar through global average pooling to obtain a global feature description. The specific calculation formula is as follows: Among them, z c represents the global description vector of the cth channel, H and W represent the height and width of the feature map respectively, x c (i,j) represents the (i,j)th pixel value of the cth channel of the input feature map; Secondly, after obtaining the global description of each channel, the global description is processed through two fully connected layers and activation functions to generate the weight of each channel, which is expressed as: yes c =σ(W2·δ(W1·z c )); Among them, s c represents the recalibrated channel weight vector, W1 and W2 represent the weight matrices of the fully connected layer, σ represents the Sigmoid activation function, and δ represents the ReLU activation function; Finally, the obtained channel weights are remapped back to the feature map of each channel, and the features of each channel are reweighted by element-by-element multiplication. The resulting image size is 320×320×64, expressed as: in, represents the feature map of the cth channel after recalibration by the SE module, x c Represents the feature map of the cth channel of the input feature map; The object detection model is trained by using the data set, including using a self-distillation module to receive features of a backbone part of the output, constructing a multi-branch distillation path, and guiding the alternating module to obtain performance similar to that of its next alternating module; The process of the self-distillation module guiding the alternating module to obtain performance similar to that of the next alternating module comprises the following steps: Get the output of the upper alternating module A 1×1 adaptive convolution adjuster is used to adjust it, which is expressed as: in, represents the convolution regulator of the upper alternating module, represents the convolution parameters; Get the output of the middle-level alternating module A 1×1 adaptive convolution adjuster is used to adjust it, which is expressed as: in, represents the convolution regulator applied to the middle layer alternating module, represents the convolution parameters; Get the output F of the lower alternating module guide , calculate F guide and The L2 norm between guide and The L2 norm between , and construct the self-distillation loss, expressed as: Among them, Loss self represents the total self-distillation loss, represents the kth individual self-distillation loss term, and ||·||2 represents the L2 norm; The target detection model is trained using the data set, and the overall loss of the training is expressed as: Loss=αLoss obj +βLoss GIoU +λLoss clc +ηLoss self ; Among them, Loss obj Represents confidence loss, Loss clc Represents classification loss, Loss self represents the self-distillation loss, α, β, λ and η represent the confidence weight, rectangular box weight, classification weight and self-distillation weight respectively; Loss GIoU represents the GIoU loss, expressed as: Among them, A represents the predicted box, B represents the real box, C represents the minimum closed area containing A and B, and |·| represents the area.

Images