An Image Processing Method for Rabbit Airway In Situ Tumors

Abstract

The present invention provides a method for processing images of rabbit airway in-situ tumors, which relates to the field of image enhancement. The aim is to construct an image feature enhancement network for rabbit airway in-situ tumors, improve the size of the traditional attention window by using the dynamic window method, adaptively calculate the attention weights of the dynamic window, and enhance the spatial feature extraction ability of the image; propose a BCDM model for extracting the global dependence relationship between feature channels, improve the extraction of the global dependence relationship between feature channels, and enhance the expression ability of channel features; combine the improved spatial attention weights and channel attention weights to perform weighted fusion on spatial features and channel features, and improve the image enhancement effect.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement, and particularly relates to a method for processing images of rabbit airway in-situ tumors. Background Art

[0002] With the rapid development of medical imaging technology, especially the wide application of small animal imaging devices such as small animal CT and small animal MRI, the early detection and treatment monitoring of tumors have become increasingly important. As an important experimental animal, rabbits have been widely used in tumor biology research, especially playing an irreplaceable role in the establishment of tumor models and drug screening. The rabbit airway in-situ tumor model is a commonly used model for studying lung diseases, tumor treatment methods, and drug action mechanisms. However, the quality of rabbit airway in-situ tumor images still faces some challenges. Although modern imaging technology can provide high-resolution images, due to the complexity of the tumor site, motion artifacts of organs, and noise and low contrast problems during the imaging process, it is often difficult to clearly display the details of the tumor and surrounding tissues. The insufficient image quality seriously affects the accuracy of tumor diagnosis and the evaluation of treatment effects. Therefore, the technology for enhancing rabbit airway in-situ tumor images is particularly important in medical image processing.

[0003] In terms of image feature extraction, spatial features and channel features are important components of images. Spatial features determine the structural information of images, while channel features reveal attributes such as the color and texture of images. Traditional image processing methods often cannot fully extract these two types of features and lack the modeling of the global dependence relationship between channels. To solve this problem, recent research has optimized the feature extraction process by introducing adaptive mechanisms and attention mechanisms. In particular, the combination of the dynamic window method and the global dependence modeling method has become an effective way to improve the image feature expression ability and enhance image quality. Summary of the Invention

[0004] The present invention provides a method for processing images of rabbit airway in-situ tumors, aiming to construct a feature enhancement network for rabbit airway in-situ tumor images, improve the size of the traditional attention window using the dynamic window method, enhance the ability to extract spatial features of images by adaptively calculating the attention weights of the dynamic window; propose a BCDM model for extracting the global dependence relationship between feature channels, improve the extraction of the global dependence relationship between feature channels, and enhance the expression ability of channel features; combine the improved spatial attention weights and channel attention weights to perform weighted fusion on spatial features and channel features, and enhance the image enhancement effect.

[0005] To achieve the above object, the present invention provides the following technical solution: A method for processing images of rabbit airway in-situ tumors, including the following steps.

[0006] S1. Make a dataset of rabbit airway in-situ tumor images.

[0007] S2. Improve the traditional attention window size using the dynamic window method, adaptively adjust the height and width of the window according to the feature gradient complexity, introduce self-attention to calculate the spatial dynamic window self-attention, and construct the tial-fea extraction block by combining multi-scale features; the specific steps of the tial-fea extraction block are as follows:

[0008] S21. Input the first in-situ tumor image feature F of the rabbit airway R ∈R H×W×C , where H, W, and C are the height, width, and total number of channels of the input image respectively. Divide F R into two parts to perform spatial dynamic window self-attention calculation and multi-scale feature extraction respectively;

[0009] S22. For the spatial dynamic window self-attention calculation, first use a 3×3 convolution operation to generate position information Conv 3×3 is a 3×3 convolution. Add the position information to element by element The spatial dynamic window self-attention adaptively adjusts the window size according to the feature gradient complexity T to obtain the width and height H h of the dynamic window, and calculate the spatial self-attention within the H h ×W w window The mathematical model of the feature gradient complexity T is:

[0010]

[0011] The width W w of the dynamic window and the height H h of the dynamic window have the following mathematical model:

[0012] W w = W min + (W max - W min )·σ(d·(T thresh - T));

[0013] H h = H min + (H max - H min )·σ(d·(T thresh - T));

[0014] In the formula, W min , W max are the minimum and maximum values of the window width, and H min , H maxare the minimum and maximum values of the window height, σ is the Sigmoid function, the control factor d adaptively controls the steepness of the Sigmoid curve, and T thresh is the gradient threshold;

[0015] The mathematical model of the control factor d is:

[0016]

[0017] In the formula, μ hw is the global mean, σ hw is H max ×W max is the mean of the eigenvalue within the window, β is the offset for controlling the steepness of the Sigmoid curve, and τ is a small constant to prevent the denominator from being zero;

[0018] S23. The multi-scale convolutional layer processes in parallel using convolutional kernels of 3, 5, and 7 respectively, and then adds them together to obtain multi-scale features

[0019] S24. Fuse the spatial dynamic window self-attention and multi-scale features, perform operations using the GELU activation function, and finally introduce a residual connection and F R to perform element-wise addition to obtain the spatial features of the rabbit airway in-situ tumor image

[0020] S3. Propose an improved BCDM model to extract the global dependence between feature channels. Specifically, encode the information of spatial positions into the channel relationship, and at the same time design an adaptive adjustment factor to dynamically adjust the importance of channel features.

[0021] S4. Adopt dynamic channel weights to optimize the feature weighting method of traditional channel weights, and combine the BCDM model to construct a cnel-acter extraction block.

[0022] S5. Introduce improved spatial attention weights and channel attention weights, and combine the tial-fea extraction block and the cnel-acter extraction block to construct a Space-channel hybrid module.

[0023] S6. Construct a feature enhancement network for rabbit airway in-situ tumor images. In the first stage, perform multi-scale enhancement, in the second stage, perform feature interaction, and in the third stage, fuse the global features and local features of the rabbit airway in-situ tumor images.

[0024] S7. Construct a rabbit airway in-situ tumor image enhancement model. The model enhances the features of the rabbit airway in-situ tumor image through the rabbit airway in-situ tumor image feature enhancement network and outputs the enhanced rabbit airway in-situ tumor image.

[0025] Preferably, in step S1, a micro-bronchoscope dedicated to animals is used to directly capture tumor images inside the rabbit airway, and a dataset of in-situ tumor images of the rabbit airway with high quality, covering the three growth stages of early, middle, and late stages of the tumor, is made. Low-resolution in-situ tumor images of the rabbit airway are generated from high-resolution in-situ tumor images of the rabbit airway, and different low-resolution images are generated from one high-resolution image with different sampling factors. Gaussian noise with different degrees is added to the low-resolution in-situ tumor images of the rabbit airway to simulate the influence of the slight movement of the rabbit during acquisition. The processed dataset is divided into three parts: a training set, a validation set, and a test set, according to the ratio of 5:2:3.

[0026] Preferably, in step S3, the weighted feature F of the in-situ tumor image of the rabbit airway is input BC ∈R H×W×C , and the BCDM model is used to extract the global dependence relationship between feature channels, which is defined as:

[0027]

[0028] In the formula, is the global dependence relationship between feature channels, F BC is the feature weighted and input into the BCDM model, W c is the fusion weight, Q, K, and V are the query, key, and value of the image feature, W p is the weight matrix, Flatten is the flattening matrix, d k is the normalization factor, is the adaptive adjustment factor;

[0029] The adaptive adjustment factor The mathematical model is:

[0030]

[0031] In the formula, FT is the fast Fourier transform;

[0032] The fusion weight W c The mathematical model is:

[0033]

[0034] In the formula, X(i, h, w) is the eigenvalue of channel i at position (h, w), and max h，w is the largest eigenvalue in channel i.

[0035] Preferably, in step S3, the BCDM model is used to extract the global dependencies between feature channels, encoding the information of spatial positions into the channel relationships to capture the interaction between the spatial structure and global dependencies; the adaptive adjustment factor uses the fast Fourier transform to extract low-frequency information, strengthening the interaction between low-frequency and high-frequency features; a learnable fusion weight is introduced to balance the contribution of the adaptive adjustment factor.

[0036] Preferably, in step S4, the cnel-acter extraction block extracts the channel features of the second rabbit airway in-situ tumor image. The specific steps are as follows:

[0037] S41. Input the second rabbit airway in-situ tumor image feature F C ∈R H×W×C , and dynamically generate the channel weight W C according to F chan . Multiply the channel weight W chan element-wise with F C to obtain the weighted feature F C1 = W chan · F C ;

[0038] The mathematical model of the channel weight W chan is:

[0039] W chan = σ(W arr · GPool(F c ) + b p );

[0040] In the formula, W arr represents the weight matrix of the linear transformation, GPool represents the global pooling layer, and b p represents the bias vector;

[0041] S42. Input the weighted feature into the BCDM model. The output of the BCDM model captures the local channel details through depthwise separable convolution, and then performs batch normalization and GELU activation function operations to extract the channel features

[0042]

[0043] In the formula, BN is the batch normalization layer, DWConv is the depthwise separable convolution, BCDM is the BCDM model, and GELU is the GELU activation function.

[0044] Preferably, in step S4, the cnel-acter extraction block is used to extract the second rabbit airway in-situ tumor image channel features. By introducing dynamic channel weights, the expression ability of image features is effectively improved. The dynamically generated channel weights enable the channel features of each image to be adaptively adjusted according to the input changes, enhancing the correlation between the channel features and the image content. By using the BCDM model and the deep convolutional layer, the details of the channel features can be finely extracted. Through the combination of batch normalization and the GELU activation function, the non-linear modeling ability and training stability are further improved.

[0045] Preferably, in step S5, the Space-channel hybrid module is adopted to fuse the spatial features and the channel features. The specific steps are as follows:

[0046] Input the spatial features of the rabbit airway in-situ tumor image extracted by the tial-fea extraction block Calculate the improved spatial attention weight A S , input the channel features of the rabbit airway in-situ tumor image extracted by the cnel-acter extraction block Calculate the improved channel attention weight A C , combine the improved spatial attention weight and the channel attention weight to and weight and fuse them. The fused features are subjected to batch normalization and convolution operations to obtain the preliminarily enhanced rabbit airway in-situ tumor image features

[0047]

[0048] The improved spatial attention weight A S The mathematical model is:

[0049]

[0050] In the formula, Mean and Max are the average pooling and maximum pooling operations respectively, and Concat is the feature fusion;

[0051] The improved channel attention weight A C The mathematical model is:

[0052]

[0053] In the formula, X i is the eigenvalue of channel i.

[0054] Preferably, in step S5, the Space-channel mixing module mainly consists of a tial-fea extraction block and a cnel-acter extraction block, which are used to extract the features of in-situ tumor images of rabbit airways. By combining spatial features and channel features, the expression and enhancement effect of image features are improved. In terms of spatial feature extraction, by splicing the results after average pooling and maximum pooling with the original feature map, the spatial attention weights are calculated and convolutional operations are performed on them, enhancing the expression ability of spatial features. In terms of channel feature extraction, by combining channel attention weights, the correlation of feature channels is further improved. Through feature weighting, the focus of attention of the image can be adjusted more precisely, so as to highlight key regions and reduce the interference of background noise in image enhancement, improving image details and overall expressiveness.

[0055] Preferably, in step S6, an in-situ tumor image feature enhancement network of rabbit airways is used to optimize the features. The specific steps are as follows:

[0056] S61 Input the initial in-situ tumor image features of rabbit airways In the first stage, the input features are divided into three feature streams, and different convolutional operations are respectively performed in each feature stream to obtain and

[0057]

[0058] where SCFM is the Space-channel mixing module, Conv 5×5 is a 5×5 convolution, and Conv 7×7 is a 7×7 convolution; the enhanced multi-scale features are fused and the Space-channel mixing module is used to obtain the first-stage features:

[0059]

[0060] S62. In the second stage, the output features of the first stage are divided into four feature blocks, Split is feature partitioning, and dynamic convolutions are respectively applied to each group of feature blocks;

[0061]

[0062] where t is a convolution kernel dynamically generated by the global context information of the output features of the first stage, and j = 1, 2, 3, 4;

[0063] An interaction mechanism is introduced to connect different groups of features:

[0064]

[0065] Fuse the cross-group interaction features and use the Space-channel mixing module to obtain the second-stage features:

[0066]

[0067] S63. In the third stage, through deep feature expression, capture the global and local features of the output features of the second stage, combine the global and local features and use the Space-channel mixing module to form the final enhanced features

[0068] Preferably, in step S6, the rabbit airway in-situ tumor image feature enhancement network gradually extracts, fuses and enhances image features in a multi-stage manner. The network is divided into three stages. In the first stage, the initial rabbit airway in-situ tumor image features are divided into three feature streams, and multi-scale enhancement is performed separately in each feature stream to preliminarily enhance the features of the rabbit airway in-situ tumor image. In the second stage, deeper enhancement is performed on the features output by the first stage. The features output by the first stage are divided into four feature blocks, an interaction mechanism is introduced to connect different groups of features, and the features of different groups are interacted to capture more details. In the third stage, through the deep feature expression ability, capture the global pattern and local details of the image to obtain the finally enhanced rabbit airway in-situ tumor image features.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] The technical solution provided by the present invention is a rabbit airway in-situ tumor image feature enhancement network, which improves the size of the traditional attention window by using the dynamic window method, enhances the spatial feature extraction ability of the image by adaptively calculating the attention weights of the dynamic window; proposes a BCDM model for extracting the global dependence relationship between feature channels, improves the extraction of the global dependence relationship between feature channels, and enhances the expression ability of channel features; combines the improved spatial attention weights and channel attention weights to perform weighted fusion on spatial features and channel features, and improves the enhancement effect of the image. Description of the Drawings

[0071] Figure 1 is a flowchart of a rabbit airway in-situ tumor image processing method provided by the present invention.

[0072] Figure 2 is a structural diagram of the tial-fea extraction block provided by the present invention.

[0073] Figure 3 is a structural diagram of the Space-channel mixing module provided by the present invention.

[0074] Figure 4It is the structural diagram of the rabbit airway in-situ tumor image feature enhancement network provided by the present invention.

[0075] Figure 5 It is the image enhancement effect diagram of the low-resolution rabbit airway in-situ tumor image provided by the present invention. Specific embodiments

[0076] Next, in combination with the drawings of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0077] Please refer to Figures 1 to 5 , the present invention provides a method for processing rabbit airway in-situ tumor images, which improves the size of the traditional attention window by using the dynamic window method, adaptively calculates the attention weight of the dynamic window, and enhances the spatial feature extraction ability of the image; proposes a BCDM model for extracting the global dependence relationship between feature channels, improves the extraction of the global dependence relationship between feature channels, and enhances the expression ability of channel features; combines the improved spatial attention weight and channel attention weight to perform weighted fusion on spatial features and channel features to improve the image enhancement effect.

[0078] Please see Figure 1 as shown, a method for processing rabbit airway in-situ tumor images in an embodiment of the present application.

[0079] S1. Make a rabbit airway in-situ tumor image dataset.

[0080] Furthermore, use a special animal microbronchoscope to take 1000 tumor images of the rabbit airway interior, make a high-quality rabbit airway in-situ tumor image dataset covering the three growth stages of early, middle, and late tumors, generate 3000 low-resolution rabbit airway in-situ tumor images from 1000 high-resolution rabbit airway in-situ tumor images, generate 3 low-resolution images with different sampling factors from one high-resolution image, add Gaussian noise with different degrees to the low-resolution rabbit airway in-situ tumor images to simulate the influence of the slight movement of the rabbit during acquisition, and divide the processed dataset into three parts: a training set, a validation set, and a test set according to 5:2:3, with 1500 images in the training set, 600 images in the validation set, and 900 images in the test set.

[0081] S2. Improve the traditional attention window size by using the dynamic window method, adaptively adjust the height and width of the window according to the feature gradient complexity, introduce self-attention to calculate the spatial dynamic window self-attention, and combine multi-scale features to construct a tial-fea extraction block.

[0082] Further, as Figure 2 shown, the tial-fea extraction block extracts the spatial features of the first rabbit airway in-situ tumor image, and the specific steps are as follows.

[0083] S21. Input the first rabbit airway in-situ tumor image feature F R ∈R H×W×C , where H, W, and C are the height, width, and total number of channels of the input image respectively. Divide F R into two parts to perform spatial dynamic window self-attention calculation and multi-scale feature extraction respectively.

[0084] S22. For spatial dynamic window self-attention calculation, first use a 3×3 convolution operation to generate position information Conv 3×3 is a 3×3 convolution. Add the position information to the elements to obtain the position-weighted feature The spatial dynamic window self-attention adaptively adjusts the window size according to the feature gradient complexity T. The value range of the feature gradient complexity T is from 0.1 to 0.6, and the initial value is set to 0.4, to obtain the width W w and height H h of the dynamic window. The value range of W w is from 2 to 8, and the value range of H h is from 4 to 16. Calculate the spatial self-attention within the window of H h ×W w .

[0085] The mathematical model of the feature gradient complexity T is:

[0086]

[0087] In the formula, and are the gradients of the feature map in the height and width directions, and the value range is from 0 to 255;

[0088] The mathematical model of the width W w and height H h of the dynamic window is:

[0089] W w =W min +(W max -W mim )·σ(d·(T thresh -T));

[0090] H h =H min +(H max -H min )·σ(d·(Tthresh -T));

[0091] Wherein, W min , W max are the minimum and maximum values of the window width, the minimum value is 2, the maximum value is 8, H min , H max are the minimum and maximum values of the window height, the minimum value is 4, the maximum value is 16, σ is the Sigmoid function, the control factor d adaptively controls the steepness of the Sigmoid curve, and the value range of the control factor d is from 0.3 to 0.5, and the initial value is set to 0.3, T thresh is the gradient threshold, set to 0.4;

[0092] The mathematical model of the control factor d is:

[0093]

[0094] Wherein, μ hw is the global mean value σ hw is for H max ×W max the mean value of the eigenvalue within the window is the eigenvalue at the position of the i-th row and the j-th column in, β are respectively the offsets for controlling the steepness of the Sigmoid curve, β is set to 0.8, τ is a small constant to prevent the denominator from being zero, and τ is set to 10 -6 .

[0095] S23. The multi-scale convolutional layer uses convolutions with convolution kernels of 3, 5, and 7 for parallel processing respectively, and then adds them to obtain multi-scale features

[0096] S24. Fuse the spatial dynamic window self-attention and multi-scale features, use the GELU activation function operation, and finally introduce a residual connection and F R to perform element-wise addition to obtain the spatial features of the rabbit airway in-situ tumor image

[0097] S3. Propose an improved BCDM model to extract the global dependence relationship between feature channels. Specifically, encode the information of the spatial position into the channel relationship, and at the same time design an adaptive adjustment factor to dynamically adjust the importance of channel features.

[0098] Furthermore, input the weighted rabbit airway in-situ tumor image feature F BC ∈R H×W×C , and the BCDM model is used for the global dependence relationship between feature channels, which is defined as:

[0099]

[0100] In the formula, is the global dependence between feature channels, and F BC is the feature after weighting the input to the BCDM model, and W c is the fusion weight with a value ranging from 0.25 to 0.75 to balance the contribution of the adaptive adjustment factor. Q, K, and V are the query, key, and value of the image feature, and W p is the weight matrix, Flatten is the flattened matrix, and d k is the normalization factor, and d k is set to 64, is the adaptive adjustment factor, with a value range from 0 to 1 and an initial value set to 0.5;

[0101] The said adaptive adjustment factor The mathematical model is:

[0102]

[0103] In the formula, FT is the fast Fourier transform;

[0104] The said fusion weight W c The mathematical model is:

[0105]

[0106] In the formula, X(i, h, w) is the eigenvalue of channel i at position (h, w), and max h，w is the largest eigenvalue in channel i.

[0107] S4. Adopt a feature weighting method that optimizes the traditional channel weight with dynamic channel weights, and combine with the BCDM model to construct a cnel-acter extraction block.

[0108] Furthermore, the cnel-acter extraction block extracts the channel features of the second rabbit airway in-situ tumor image, and the specific steps are as follows.

[0109] S41. Input the second rabbit airway in-situ tumor image feature F C ∈R H×W×C , and dynamically generate the channel weight W C according to F chan . Multiply the channel weight W chan element-wise with F C to obtain the weighted feature F C1 = W chan ·F C ; The mathematical model of the said channel weight W chan is:

[0110] W chan = σ(Warr ·GPool(F C )+b p );

[0111] Wherein, W arr represents the weight matrix of the linear transformation, GPool represents the global pooling layer, and b p represents the bias vector with a value range from -0.1 to 0.1, and the initial value is set to 0.

[0112] S42. Input the weighted features into the BCDM model. The output of the BCDM model captures local channel details through depthwise separable convolution, and then performs batch normalization and GELU activation function operations to extract channel features

[0113] Wherein, BN is the batch normalization layer, DWConv is the depthwise separable convolution, BCDM is the BCDM model, and GELU is the GELU activation function.

[0114] S5. Introduce the improved spatial attention weight and channel attention weight, and construct a Space-channel hybrid module by combining the tial-fea extraction block and the cnel-acter extraction block.

[0115] Furthermore, as Figure 3 shown, use the Space-channel hybrid module to fuse the spatial features and channel features. The specific steps are as follows:

[0116] Input the spatial features of the in-situ tumor image of the rabbit airway extracted by the tial-fea extraction block Calculate the improved spatial attention weight A S , A S ranges from 0 to 1. Input the channel features of the in-situ tumor image of the rabbit airway extracted by the cnel-acter extraction block Calculate the improved channel attention weight A C , A C ranges from 0 to 1, and A S and A C add up to 1. Combine the improved spatial attention weight and channel attention weight to and weight and fuse them. The fused features are subjected to batch normalization and convolution operations to obtain the preliminarily enhanced features of the in-situ tumor image of the rabbit airway

[0117] The mathematical model of the improved spatial attention weight A S is:

[0118]

[0119] In the formula, Mean and Max are average pooling and max pooling operations respectively, and Concat is feature fusion;

[0120] The improved channel attention weight A C The mathematical model is:

[0121]

[0122] In the formula, X i is the eigenvalue of channel i.

[0123] S6. Construct a feature enhancement network for in-situ rabbit airway tumor images. In the first stage, multi-scale enhancement is performed. In the second stage, feature interaction is carried out. In the third stage, the global features and local features of the in-situ rabbit airway tumor images are fused.

[0124] Furthermore, as Figure 4 shown, the feature enhancement network for in-situ rabbit airway tumor images is used to optimize the features, and the specific steps are as follows.

[0125] S61. Input the initial features of the in-situ rabbit airway tumor image into the first stage. The input features are divided into three feature streams, and different convolution operations are respectively performed in each feature stream to obtain and

[0126] In the formula, SCFM is the Space-channel mixing module, and Conv 5×5 is a 5×5 convolution, and Conv 7×7 is a 7×7 convolution;

[0127] The enhanced multi-scale features are fused and the Space-channel mixing module is used to obtain the first-stage features:

[0128]

[0129] S62. In the second stage, the output features of the first stage are divided into four feature blocks, Split is feature partitioning, and dynamic convolution is applied to each group of feature blocks;

[0130]

[0131] In the formula, t is the convolution kernel dynamically generated by the global context information of the input features. The initial values of t are set to 3, 5, and 9 respectively, and j = 1, 2, 3, 4;

[0132] Introduce an interaction mechanism to connect different groups of features:

[0133]

[0134] Fuse the cross-group interaction features and use the Space-channel mixing module to obtain the second-stage features:

[0135]

[0136] S63. In the third stage, through deep feature expression, capture the global and local features of the output features of the second stage, combine the global and local features and use the Space-channel mixing module to form the final enhanced features

[0137] S7. Construct a rabbit airway in-situ tumor image enhancement model. The model enhances the features of the rabbit airway in-situ tumor image through the rabbit airway in-situ tumor image feature enhancement network and outputs the enhanced rabbit airway in-situ tumor image.

[0138] Furthermore, in step S7, the rabbit airway in-situ tumor image enhancement model includes an input, initial feature extraction, a rabbit airway in-situ tumor image feature enhancement network, an image reconstruction module, and an output. Input the rabbit airway in-situ tumor image into the model. The model enhances the rabbit airway in-situ tumor image through the rabbit airway in-situ tumor image feature enhancement network. The enhanced rabbit airway in-situ tumor image features are reconstructed into a rabbit airway in-situ tumor image through the image reconstruction module, and the enhanced rabbit airway in-situ tumor image is output.

[0139] Furthermore, in step S7, the rabbit airway in-situ tumor image enhancement model is developed using the Python language through the Pycharm application and implemented based on the Pytorch framework. Use 1500 training sets in the rabbit airway in-situ tumor image dataset to train the model. Start training using the self-supervised learning method. Compare the training results with the validation set and test the model with the test set. The model extracts the spatial and channel features in the image, enhances the rabbit airway in-situ tumor image features in the spatial and channel directions, and constructs the enhanced rabbit airway in-situ tumor image through the enhanced features.

[0140] Furthermore, as Figure 5 shown, Figure 5 the left half in it is a low-resolution rabbit airway in-situ tumor image, Figure 5 and the right half in it is the enhanced rabbit airway in-situ tumor image obtained after being processed by the rabbit airway in-situ tumor image enhancement model, significantly enhancing the clarity and recognizability of the rabbit airway in-situ tumor image.

[0141] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the inventive concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention.

Claims

1. A rabbit airway in situ tumor image processing method, characterized in that: The following steps are involved: S1. Create a rabbit airway in situ tumor image dataset; S2. Use the dynamic window method to improve the traditional attention window size, adaptively adjust the height and width of the window according to the feature gradient complexity, introduce the dynamic window self-attention in the self-attention calculation space, and combine multi-scale features to construct the tial-fea extraction block; The specific steps of the tial-fea extraction block are: S21, input the first rabbit airway in situ tumor image feature F R ∈R H×W×C , H, W, C are the height, width, and total number of channels of the input image respectively. R Divide The two parts perform spatial dynamic window self-attention calculation and multi-scale feature extraction respectively; S22, spatial dynamic window self-attention calculation first uses a 3×3 convolution operation to generate position information Conv 3×3 It is a 3×3 convolution, which combines the position information with Adding elements The spatial dynamic window self-attention adaptively adjusts the window size according to the feature gradient complexity T to obtain the width W of the dynamic window w and high H h , calculate H h ×W w Spatial self-attention within a window The mathematical model of the characteristic gradient complexity T is: In the formula, and is the gradient of the feature map in height and width; The width W of the dynamic window w and high H of dynamic window h The mathematical model is: W w =W min +(W max -W min )·σ(d·(T thresh -T)); H h =H min +(H max -H min )·σ(d·(T thresh -T)); Where W min , W max is the minimum and maximum value of the window width, H min , H max is the minimum and maximum value of the window height, σ is the Sigmoid function, and the control factor d adaptively controls the steepness of the Sigmoid curve. thresh is the gradient threshold; The mathematical model of the control factor d is: In the formula, μ hw is the global mean, σ hw H max ×W max The mean of the eigenvalues ​​in the window, β is the offset that controls the steepness of the Sigmoid curve, To prevent a small constant with a zero denominator; S23, multi-scale convolution layer uses convolution kernels of 3, 5, and 7 for parallel processing, and then adds them together to obtain multi-scale features S24, the spatial dynamic window self-attention and multi-scale features are integrated, the GELU activation function is used, and finally the residual connection and F are introduced. R Add the elements to get the spatial characteristics of the rabbit airway in situ tumor image S3. A BCDM model is proposed to improve the global dependency between feature channels. Specifically, the spatial position information is encoded into the channel relationship, and an adaptive adjustment factor is designed to dynamically adjust the importance of channel features. S4, using dynamic channel weights to optimize the feature weighting method of traditional channel weights, and combining the BCDM model to build the cnel-acter extraction block; S5, introduce improved spatial attention weights and channel attention weights, and combine the tial-fea extraction block and the cnel-acter extraction block to build a Space-channel hybrid module; S6. Construct a rabbit airway tumor in situ image feature enhancement network. The first stage performs multi-scale enhancement, the second stage performs feature interaction, and the third stage fuses the global and local features of the rabbit airway tumor in situ image. S7. Construct a rabbit airway in situ tumor image enhancement model. The model enhances the features of the rabbit airway in situ tumor image through the rabbit airway in situ tumor image feature enhancement network, and outputs the enhanced rabbit airway in situ tumor image.

2. The rabbit airway in situ tumor image processing method according to claim 1, characterized in that: In the step S1, a high-quality rabbit airway in situ tumor image dataset covering the three growth stages of the tumor, namely, early stage, middle stage and late stage, is produced, a low-resolution rabbit airway in situ tumor image is generated using a high-resolution rabbit airway in situ tumor image, a high-resolution image is used to generate a low-resolution image with different sampling factors, different degrees of Gaussian noise are added to the low-resolution rabbit airway in situ tumor image to simulate the influence of the rabbit's micro-motion during acquisition, and the processed dataset is divided into three parts: a training set, a validation set and a test set.

3. The rabbit airway in situ tumor image processing method according to claim 2, characterized in that: In the step S3, the weighted rabbit airway tumor in situ image feature F is input. BC ∈R H×W×C , H, W, C are F BC The height, width, and total number of channels of the BCDM model are defined as: In the formula, is the global dependency between feature channels, W c is the fusion weight, Q, K and V are the weighted query, key and value of the rabbit airway in situ tumor image features, and W p is the weight matrix, Flatten is the flattened matrix, d k is the normalization factor, is the adaptive adjustment factor; The adaptive adjustment factor The mathematical model is: Where FT is fast Fourier transform; The fusion weight W c The mathematical model is: Where X(i, h, w) is the eigenvalue of channel i at position (h, w), is the largest eigenvalue in the channel.

4. The rabbit airway in situ tumor image processing method according to claim 3, characterized in that: In the step S4, the specific steps of cnel-acter extraction block are: S41, input the second rabbit airway in situ tumor image feature F C ∈R H×W×C , according to F C Dynamically generate channel weights W chan , the channel weight W chan With F C Multiply element by element to get the weighted feature F C1 =W chan ·F C ; The channel weight W chan The mathematical model is: W chan =σ(W arr ·GPool(F C )+b p ); Where W arr represents the weight matrix of the linear transformation, GPool represents the global pooling layer, and b p represents the bias vector; S42. The weighted features are input into the BCDM model. The output of the BCDM model captures local channel details through deep separable convolution, and then batch normalization and GELU activation function operations are performed to extract channel features. Where BN is the batch normalization layer, DWConv is the depthwise separable convolution, BCDM is the BCDM model, and GELU is the GELU activation function.

5. The rabbit airway in situ tumor image processing method according to claim 4, characterized in that: In the step S5, the specific steps of the Space-channel mixing module are: Input spatial features of rabbit airway in situ tumor images extracted by tial-fea extraction block Calculate the improved spatial attention weight A S , input the channel features of the rabbit airway in situ tumor image extracted by the cnel-acter extraction block Calculate the improved channel attention weight A C , combined with the improved spatial attention weight and channel attention weight and Weighted and fused, the fused features are batch normalized and convolved to obtain the initial enhanced rabbit airway in situ tumor image features The improved spatial attention weight A S The mathematical model is: In the formula, Mean and Max are average pooling and maximum pooling operations respectively, and Concat is feature fusion; The improved channel attention weight A C The mathematical model is: Where, X i is the eigenvalue of channel i.

6. A rabbit airway tumor in situ image processing method according to claim 5, characterized in that: In step S6, the specific steps of the rabbit airway in situ tumor image feature enhancement network are as follows: S61 input initial rabbit airway in situ tumor image features In the first stage, the input features are divided into three feature streams, and different convolution operations are performed in each feature stream to obtain and In the formula, SCFM is the Space-channel hybrid module, Conv 5×5 is a 5×5 convolution, Conv 7×7 It is a 7×7 convolution; the enhanced multi-scale features are fused and the first-stage multi-scale enhanced features are obtained using the Space-channel mixing module: S62, the second stage converts the output features of the first stage Divided into four feature blocks, Split is feature partitioning, and dynamic convolution is applied to each group of feature blocks; Where t is the convolution kernel dynamically generated by the global context information of the output features of the first stage, j = 1, 2, 3, 4; Introduce an interaction mechanism to connect different groups of features: The cross-group interaction features are fused and the second-stage features are obtained using the Space-channel mixing module: S63. The third stage captures the global and local features of the output features of the second stage through deep feature expression, combines the global and local features and uses the Space-channel hybrid module to obtain the final enhanced features.

Images