Image enhancement processing method for monitoring throat and neck development of children

Abstract

The invention provides an image enhancement processing method for child laryngeal and neck development monitoring, and relates to the field of image enhancement, a traditional attention calculation method is improved, local pathological attention is calculated in combination with pathological masks and adaptive Gaussian variational attention, and the attention of an important area is highlighted while the calculation complexity is reduced; a hierarchical recursive distillation module is constructed, features are extracted and enhanced step by step through recursive distillation, the features are divided into two parts through adaptive feature segmentation, the important part is used for attention calculation, the remaining part is directly transmitted to the end for cross-scale structure fusion, and unnecessary calculation is reduced; a traditional information distillation method is improved, a self-adaptive information distillation network is constructed in combination with feature separation, hierarchical recursive distillation and cross-scale structure guide fusion, an image enhancement task for child throat and neck development monitoring is optimized, model efficiency is improved, and high-order information is reserved through multi-level learning.

Description

Technical Field

[0001] The present invention belongs to the field of image enhancement, and particularly relates to an image enhancement processing method for pediatric laryngeal and cervical development monitoring. Background Art

[0002] With the development of medical imaging technology, especially in the field of pediatric laryngeal and cervical development monitoring, image enhancement has been widely applied in clinical diagnosis and has become an important means to evaluate the development status of the pediatric larynx, vocal cords, and trachea. However, when facing the complex anatomical structures and variable development states of the pediatric larynx and neck, the existing image enhancement technologies still face many challenges, mainly reflected in the problem of restoring complex image details. The laryngeal anatomical structure of children is more delicate and complex compared to that of adults, especially the underdeveloped larynx, and the presented image features are difficult to identify. Traditional image enhancement methods often cannot perform precise enhancement for different anatomical structures.

[0003] The current image enhancement methods are generally applied to the entire image and lack local fine enhancement of key regions. This makes it possible for doctors to fail to accurately identify early signs of developmental lag during the development monitoring process. Therefore, how to accurately restore the details of the laryngeal structure and enhance the visualization effect of key regions while ensuring image quality has become an important research direction in medical image processing. To solve the problems of unstable image quality, inaccurate detail restoration, and insufficient enhancement of local region information in the existing technology, through local pathological attention and adaptive information distillation, the detail restoration ability of the image is effectively improved, and the enhancement strategy is automatically adjusted according to the characteristics of different imaging modalities, so as to provide more accurate diagnostic basis for doctors. Summary of the Invention

[0004] The present invention provides an image enhancement processing method for pediatric laryngeal and cervical development monitoring, which improves the traditional attention calculation method, combines a pathological mask and adaptive Gaussian variational attention to calculate local pathological attention, highlighting the attention of important regions while reducing the computational complexity; constructs a hierarchical recursive distillation module to gradually extract and enhance features through recursive distillation, divides the features into two parts through adaptive feature segmentation, the important part is used for attention calculation, and the remaining part is directly passed to the end for cross-scale structure fusion, reducing unnecessary calculations; improves the traditional information distillation method, combines feature separation, hierarchical recursive distillation, and cross-scale structure-guided fusion to construct an adaptive information distillation network, optimizes it for the image enhancement task of pediatric laryngeal and cervical development monitoring, improves the model efficiency, and retains high-order information through multi-level learning.

[0005] To achieve the above object, the present invention provides the following technical solution: An image enhancement processing method for pediatric laryngeal and cervical development monitoring, comprising the following steps.

[0006] S1. Obtain images of pediatric laryngeal and cervical development and form a dataset.

[0007] S2. Construct a data augmentation module for pediatric laryngeal and cervical data, design an adaptive non - linear transformation, contrast adaptive enhancement based on structural features, and simulation noise based on the physical properties of medical imaging to simulate the impacts during the acquisition process of pediatric laryngeal and cervical image data.

[0008] S3. Introduce Gaussian variational inference to design an adaptive Gaussian variational attention, construct a pathological mask - guided attention calculation, calculate local attention based on the pathological mask and improved traditional attention, and combine local attention and adaptive Gaussian variational attention to calculate local pathological attention.

[0009] S4. Design a recursive distillation to gradually extract and enhance features based on a recursive optimization strategy, design an adaptive feature segmentation method based on a self - designed adaptive segmentation coefficient, construct a hierarchical recursive distillation module through recursive distillation and adaptive feature segmentation, and retain important features to the next layer for attention calculation.

[0010] S5. Improve the traditional information distillation method to construct an adaptive information distillation network, including feature separation, hierarchical recursive distillation, and cross - scale structure - guided fusion. Design an attention mask based on local gradients for feature separation, and design adaptive fusion weights based on the attention gating mechanism to fuse the distilled features of different layers.

[0011] S6. Construct an image enhancement model for pediatric laryngeal and cervical development monitoring, and complete data simulation and image enhancement for pediatric laryngeal and cervical development monitoring through the pediatric laryngeal and cervical data augmentation module and the adaptive information distillation network.

[0012] Preferably, in step S1, obtain images of pediatric laryngeal and cervical development monitoring, capture images of the vocal cords, trachea, and laryngeal soft tissues, perform data annotation on the obtained images, including key area annotation, developmental status annotation, and case category annotation. Label the vocal cords, laryngeal cartilage, and trachea as key areas. The developmental status is divided into three types: normal development, developmental lag, and developmental abnormality. The case category annotation is classified according to the pathological conditions, and the annotated data is divided into a training set and a validation set.

[0013] Preferably, in step S2, the specific steps of the pediatric laryngeal and cervical data augmentation module are as follows: S21. Simulate patient pose changes, imaging angle deviations, and equipment jitters through an adaptive non - linear transformation. The specific formula is: where W geo and B geo are spatial mapping matrices based on affine transformation, simulating angle offset and scale change respectively, N is the number of harmonic components, α i is the amplitude parameter, controlling the deformation intensity, β iis the frequency parameter that controls the local change scale, δ i is the phase parameter that controls the initial state of deformation. X is the image input to the pediatric laryngeal neck data enhancement module, and sin is the sine function;

[0014] S22. By adaptively highlighting the details of the airway, glottis, and muscle tissues based on structural features and suppressing background noise, the specific formula is: In the formula, α con is the edge enhancement intensity, tanh is the hyperbolic tangent function, is the Laplacian operator, and ψ(X) are the mean and standard deviation of X respectively, τ con is the contrast threshold, β con is the contrast enhancement intensity. The mathematical formula of the contrast enhancement intensity β con is: In the formula, exp is the exponential function, γ targ is the target brightness value;

[0015] S23. Through the simulation noise based on the physical characteristics of medical imaging, the specific formula is: In the formula, α sn is the device noise intensity, Γ(0.01, 0.05) is the device noise, β sn is the tissue scattering noise intensity, ∏(-0.3, 0.3) is the tissue scattering noise, which follows a uniform distribution, γ sn is the dynamic blur intensity, is the time change rate; The mathematical formula of the device noise intensity ɑ sn is: In the formula, Be is the current load of the device, Be max is the maximum load of the device, Ti is the device usage duration, Ti max is the maximum device usage duration.

[0016] Preferably, in step S2, the pediatric laryngeal neck data enhancement module designs an adaptive non - linear transformation, a contrast - adaptive enhancement based on structural features, and a simulation noise based on the physical characteristics of medical imaging from three aspects: geometric transformation, contrast enhancement, and simulation experiment noise. It performs adaptive enhancement through these three aspects of geometric transformation, contrast enhancement, and simulation experiment noise to simulate various influences in the process of obtaining pediatric laryngeal neck medical images. Through the adaptive non - linear transformation, it simulates patient pose changes, imaging angle deviations, equipment jitters, and angle changes, improving the adaptability of the model to changes in the angles and sizes of different pediatric laryngeal neck medical images. By using the contrast - adaptive enhancement based on structural features, it highlights the details of the airway, glottis, and muscle tissues, suppresses background noise, and optimizes the visualization of low - contrast regions. Through the simulation noise based on the physical characteristics of medical imaging, it simulates the noise in pediatric laryngeal neck medical images, enhancing the robustness of the model to noise and artifacts. These enhancement techniques can provide more diverse data for training. Using the enhanced pediatric laryngeal neck data to train the model can improve the generalization ability of the model in the actual medical environment, and further improve the image quality and diagnostic accuracy of pediatric laryngeal neck development monitoring.

[0017] Preferably, in step S3, the specific steps for calculating local pathological attention are as follows: S31. Input the first - image feature X i ∈R H×W×C , where H, W, and C are the height, width, and channels of the first image respectively. Construct a pathological mask M p to guide the attention calculation, distinguish the key regions of the larynx such as the vocal cords, trachea, and cartilage, the irrelevant region of the background, and the low - information region, reduce the computational complexity, calculate the attention of the important regions. The pathological mask M p is weighted with the optimized Softmax attention to obtain the local attention A PL , and the specific calculation formula of the pathological mask M p is: In the formula, Z is the normalization factor, δ controls the local smoothness, x i (i, j) is the gray - scale value at the pixel position (i, j), Ω is a 5×5 local region, α is the local smooth weight, β is the edge - enhancement weight, and are the gradient values at x and y respectively; The specific calculation formula of the local attention A PL is: In the formula, G is the expansion order, ⊙ is the weighted operation, Q, K, and V are the query, key, and value respectively, which are obtained by the linear transformation matrices W Q , W K , W VGet, O = W O x i , K = W K x i , V = W V x i , where T represents transpose and d is the dimension of a single attention head;

[0018] S32. Introduce Gaussian variational inference to optimize attention calculation to obtain adaptive Gaussian variational attention A GVA , and the specific formula is: In the formula, Q i is the Q variable of the Gaussian distribution, Q is the query, μ Q is the mean of Q, ε represents the noise of the standard normal distribution, θ Q is the variance of Q, G i is the adaptive Gaussian compensation factor, p(Q i ) is the variational probability density of Q, σ T is the medical modality adjustment factor, σ T = 0.5·Var(x i ) + 0.5·Mean(x i ), Var(x i ) is the local variance of X i , Mean(x i ) is the local mean of X i ; The adaptive Gaussian compensation factor G i has the following specific calculation formula: In the formula, W G is the linear transformation matrix, b G is the Gaussian bias term that controls the benchmark for compensating the mean, λ1 and λ2 are learning parameters, and λ1 + λ2 = 1;

[0019] S33. Combine the local attention A PL and the adaptive Gaussian variational attention A GVA to obtain the local pathological attention A LPSA , and the specific formula is: A LPSA = τ·A PL + (1 - τ)·A GVA ; In the formula, τ is the attention fusion weight, which is adaptively learned by the neural network.

[0020] Preferably, in step S3, local pathological attention plays an important role in pediatric laryngeal and cervical image enhancement by introducing a pathological mask and adaptive Gaussian variational attention. It can accurately focus on key regions such as vocal cords, laryngeal cartilages, and trachea in pediatric laryngeal and cervical development monitoring, improve the detail restoration of local regions, and at the same time optimize noise suppression and interference from irrelevant regions. Through adaptive Gaussian variational inference, the model can adaptively adjust attention calculation according to different medical imaging modalities, enhancing the robustness and generalization ability of the model.

[0021] Preferably, in step S4, the specific steps of the hierarchical recursive distillation module are as follows: S41. The specific calculation formula for the t-th recursive distillation is: In the formula, α HRD is a hyperparameter that controls the recursive weight, is the output of the t-th recursive distillation, and Conv is the convolution operation;

[0022] S42. Input the second image feature H t 、W t and C are the height, width, and channels of the second image respectively. Adaptive feature segmentation divides the feature into two parts, X en and X fu according to the adaptive segmentation coefficient. The specific calculation formula for the adaptive segmentation coefficient is: In the formula, X c is the feature of the c-th channel of feature X t , c ∈ (1, C), and Sig is the sigmoid function; The formula for adaptive feature segmentation is: In the formula, X en is the input for the next layer of attention calculation, and X fu is used for the final feature fusion.

[0023] Preferably, in step S4, the hierarchical recursive distillation module can gradually refine the key information and structural details in the pediatric laryngeal and cervical images by recursively extracting and enhancing features layer by layer. The layer-by-layer property of recursive distillation can ensure the gradual enhancement of low-level features and high-level features, prevent information loss in the deep network, and thus maintain the integrity of the details in the pediatric laryngeal and cervical images. After distillation in each layer, the module divides the extracted features into two parts through an adaptive feature segmentation mechanism: one part is passed to the next layer for further refinement, and the other part is used for the final cross-scale structure fusion. This can not only finely enhance local details in the deep network but also improve the overall quality of the pediatric laryngeal and cervical images through global information fusion. The combination of recursive distillation and adaptive feature segmentation can effectively improve the accuracy and computational efficiency of pediatric laryngeal and cervical image super-resolution, ensuring that details and the global structure can be enhanced simultaneously.

[0024] Preferably, in step S5, the specific steps of the adaptive information distillation network are as follows: S51. Decompose the pediatric laryngeal and cervical medical image into a key region and a background region, calculate the attention mask of the image, and based on the attention mask, divide the pediatric laryngeal and cervical medical image \(I\in R\) H×W×C into the key region \(I\) crux and the background region \(I\) cont , where \(H\), \(W\), and \(C\) are the height, width, and channels of the pediatric laryngeal and cervical medical image respectively. The specific calculation formula for the attention mask \(M\) AM is as follows: In the formula, is the local gradient value, \(x1\in[0, H]\), \(y1\in[0, W]\), and MAX is to take the maximum value; The calculation formula for the key region \(I\) crux is: \(I\) crux \(=\) \(M\) AM \(\odot\) \(I\); The calculation formula for the background region \(I\) cont is: \(I\) cont \(=\) \((1 - M\) AM ) \(\odot\) \(I\);

[0025] S52. The information in the key region is more important than that in the background region. A higher distillation weight needs to be assigned to the key region \(I\) crux . According to the distillation weight, obtain the region-adaptive distillation feature. The distillation weight formula is: \(W\) RAD \(=\) \(Sig(W\) 1R \(I\) crux \(+\) \(W\) 2R \(I\) cont ); In the formula, \(W\)1R and W 2R is a trainable weight matrix; The region adaptive distilled feature F RAD The calculation formula is: In the formula, ||I crux ⊙I cont || 2 is the L2 norm of the feature map, and e is the base of the natural logarithm;

[0026] S53. Use local pathological attention for attention calculation. The specific formula is: F LPSA = A LPSA (F RAD ); In the formula, A LPSA is the local pathological attention;

[0027] S54. Use hierarchical recursive distillation to gradually extract and enhance features, and divide the features into two parts through an adaptive feature segmentation method, which are respectively used for the next layer of attention calculation and the final feature fusion;

[0028] S55. Fuse the distilled features of different layers. The fusion formula is: In the formula, L is the distillation level, v i is the adaptive fusion weight, is the distilled feature of the i-th layer; The specific calculation formula of the adaptive fusion weight v i is: In the formula, and are learned through the attention gating mechanism.

[0029] Preferably, in step S5, the adaptive information distillation network can precisely process the key regions and detailed information in pediatric laryngeal and cervical medical images by combining feature separation, hierarchical recursive distillation, and cross-scale structure-guided fusion. In the image enhancement task of pediatric laryngeal and cervical development monitoring, first, the pediatric laryngeal and cervical medical image is divided into key regions and background regions through the feature separation mechanism to ensure that the key regions receive more computing resources; then, the hierarchical recursive distillation gradually extracts and optimizes the detailed information of the pediatric laryngeal and cervical medical image, refining the features of important regions in each layer of recursion to avoid information loss; finally, through cross-scale structure-guided fusion, the global structure and local details are effectively combined to achieve the best effect in both details and global structure of the pediatric laryngeal and cervical medical image; the adaptive information distillation can significantly improve the quality of super-resolution of pediatric laryngeal and cervical medical images by dynamically adjusting the information flow and feature processing.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides an image enhancement processing method for pediatric laryngeal and cervical development monitoring, improves the traditional attention calculation method, combines the pathological mask and adaptive Gaussian variational attention to calculate the local pathological attention, highlighting the attention of important regions while reducing the computational complexity; constructs a hierarchical recursive distillation module to gradually extract and enhance features through recursive distillation, divides the features into two parts through adaptive feature segmentation, the important part is used for attention calculation, and the remaining part is directly transmitted to the end for cross-scale structure fusion, reducing unnecessary calculations; improves the traditional information distillation method, combines feature separation, hierarchical recursive distillation, and cross-scale structure-guided fusion to construct an adaptive information distillation network, optimizes for the image enhancement task of pediatric laryngeal and cervical development monitoring, improves the model efficiency and retains high-order information through multi-level learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flowchart of an image enhancement processing method for pediatric laryngeal and cervical development monitoring provided by the present invention.

[0032] Figure 2 is a structural diagram of local pathological attention provided by the present invention.

[0033] Figure 3 is a structural diagram of the adaptive information distillation network provided by the present invention.

[0034] Figure 4 is an enhanced effect diagram of the laryngeal coronal CT image provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0036] Please refer to Figures 1 to 4 , the present invention provides an image enhancement processing method for pediatric laryngeal and cervical development monitoring, improves the traditional attention calculation method, combines pathological masks and adaptive Gaussian variational attention to calculate local pathological attention, highlights the attention of important regions while reducing computational complexity; constructs a hierarchical recursive distillation module to gradually extract and enhance features through recursive distillation, divides the features into two parts through adaptive feature segmentation, the important part is used for attention calculation, and the remaining part is directly passed to the end for cross-scale structure fusion, reducing unnecessary calculations; improves the traditional information distillation method, combines feature separation, hierarchical recursive distillation and cross-scale structure-guided fusion to construct an adaptive information distillation network, optimizes for the image enhancement task of pediatric laryngeal and cervical development monitoring, improves the model efficiency and retains high-order information through multi-level learning.

[0037] Please see Figure 1 as shown, an image enhancement processing method in an embodiment of the present application for pediatric laryngeal and cervical development monitoring.

[0038] S1. Obtain images of pediatric laryngeal and cervical development monitoring and make them into a data set.

[0039] Further, obtain 2000 images of pediatric laryngeal and cervical development monitoring, capture images of the vocal cords, trachea, and laryngeal soft tissues, perform data annotation on the obtained images, including key region annotation, development status annotation, and case category annotation, label the vocal cords, laryngeal cartilage, and trachea as key regions, the development status is divided into three types: normal development, developmental lag, and developmental abnormality, and the case category annotation is classified according to the pathological conditions. Divide the labeled data into 1400 training sets and 600 validation sets.

[0040] S2. Construct a pediatric laryngeal and cervical data enhancement module, design an adaptive non-linear transformation, contrast adaptive enhancement based on structural features, and simulation noise based on the physical characteristics of medical imaging to simulate the influences in the process of obtaining pediatric laryngeal and cervical image data.

[0041] Furthermore, the pediatric laryngeal and cervical data enhancement module designs adaptive non - linear transformation, contrast self - adaptive enhancement based on structural features, and simulation noise based on the physical characteristics of medical imaging from three aspects: geometric transformation, contrast enhancement, and simulation experiment noise, simulates the impacts during the acquisition process of pediatric laryngeal and cervical image data, and uses the enhanced pediatric laryngeal and cervical data to train the model. The specific steps of the pediatric laryngeal and cervical data enhancement module are as follows.

[0042] S21. Simulate the patient's posture change, imaging angle deviation, and equipment jitter through adaptive non - linear transformation. The specific formula is: In the formula, W geo and B geo are spatial mapping matrices based on affine transformation, simulating angle offset and scale change respectively. N is the number of harmonic components, set to 10. The amplitude parameter α i is initially set to 0.1, and the value range is [0.05, 0.2]. The frequency parameter β i is initially set to 1, and the value range is [0.5, 2]. The phase parameter δ i is initially set to The value range is X is the image input to the pediatric laryngeal and cervical data enhancement module, and sin is the sine function.

[0043] S22. Highlight the details of the airway, glottis, and muscle tissue and suppress background noise through contrast self - adaptation based on structural features. The specific formula is: In the formula, the edge enhancement intensity α con is initially set to 2, and the value range is [0.5, 3]. tanh is the hyperbolic tangent function, is the Laplacian operator, and ψ(X) are the mean and standard deviation of X respectively. The contrast threshold τ con is set to 0.3. The contrast enhancement intensity β con is initially set to 1, and the value range is [0, 2]. The mathematical formula of the contrast enhancement intensity β con is: In the formula, exp is the exponential function, and γ targ is the target brightness value, initially set to 10, and the value range is [0, 50].

[0044] S23. Generate simulation noise based on the physical characteristics of medical imaging. The specific formula is: In the formula, the equipment noise intensity αsn The initial value is set to 0.1, the value range is [0.1, 0.4], Γ(0.01, 0.05) is the device noise, and the tissue scattering noise intensity is β sn The initial value is set to 0.1, the value range is [0, 0.3], Π(-0.3, 0.3) is the tissue scattering noise, which follows a uniform distribution, and the dynamic blur intensity is γ sn The initial value is set to 0.1, the value range is [0, 0.2], is the time change rate; the device noise intensity α sn The mathematical formula is: In the formula, Be is the current load of the device, the initial value is set to 10 kW, the value range is [0, 50] kW, Be max is the maximum load of the device, which is set to 50 kW, Ti is the device usage time, the initial value is set to 0.5 hours, the value range is [0, 12] hours, Ti max is the maximum device usage time, which is set to 12 hours.

[0045] S3. Introduce Gaussian variational inference to design an adaptive Gaussian variational attention, construct a pathological mask to guide attention calculation, calculate local attention based on the pathological mask and improved traditional attention, and calculate local pathological attention by combining local attention and adaptive Gaussian variational attention.

[0046] Furthermore, as Figure 2 shown, calculate the attention of the key area by combining local attention and adaptive Gaussian variational attention. The specific steps of calculating local pathological attention are as follows.

[0047] S31. Input the first image feature X i ∈R H×W×C , where H, W, and C are the height, width, and channels of the first image, which are set to 640×640×3. Construct a pathological mask M p to guide attention calculation, distinguish the key areas of the larynx, vocal cords, trachea, cartilage, the irrelevant area background, and the low-information area, reduce the computational complexity, calculate the attention of the important area, and the pathological mask M p is weighted with the optimized Softmax attention to obtain the local attention A PL , and the pathological mask M p The specific calculation formula is: In the formula, the normalization factor Z is set to 15.6, the local smoothness δ has an initial value of 2, and the value range is [1, 3], X i(i, j) is the grayscale value at the pixel position (i, j), Ω is a 5×5 local region, the initial value of the local smoothing weight α is set to 1, and its value range is [0.5, 2]. The initial value of the edge enhancement weight β is set to 1, and its value range is [0.3, 1.5]. and are the gradient values at x and y respectively; The local attention A PL The specific calculation formula is: In the formula, the expansion order G is set to 4, ⊙ is the weighted operation, Q, K, and V are the query, key, and value respectively, and are obtained by the linear transformation matrices W Q , W K , W V That is, Q = W Q x i , K = W K x i , V = W V x i , T represents the transpose, and the dimension d of a single attention head is set to 64.

[0048] S32. Introduce Gaussian variational inference to optimize the attention calculation to obtain the adaptive Gaussian variational attention A GVA , and the specific formula is: In the formula, Q i is the Q variable of the Gaussian distribution, Q is the query, μ Q is the mean of Q, ε represents the noise of the standard normal distribution, θ Q is the variance of Q, G i is the adaptive Gaussian compensation factor, p(Q i ) is the variational probability density of Q, σ T is the medical modality adjustment factor, σ T = 0.5·Var(x i ) + 0.5·Mean(x i ), Var(x i ) is the local variance of x i , Mean(x i ) is the local mean of X i ; The adaptive Gaussian compensation factor G i The specific calculation formula is: In the formula, W G is the linear transformation matrix, and the Gaussian bias term b GThe initial value is set to 0.1, and the value range is [0.05, 0.2]. The learning parameters λ1 and λ2 are both initially set to 0.5, and the value range is [0.3, 0.7], and λ1 + λ2 = 1.

[0049] S33. Combine the local attention A PL and the adaptive Gaussian variational attention A GVA to obtain the local pathological attention A LPSA . The specific formula is: A LPSA = τ·A PL +(1 - τ)·A GVA ; In the formula, the initial value of the attention fusion weight τ is set to 0.4, and the value range is [0.2, 0.8].

[0050] S4. Design a recursive distillation to gradually extract and enhance features based on a recursive optimization strategy, design an adaptive feature segmentation method based on a self-designed adaptive segmentation coefficient, construct a hierarchical recursive distillation module through recursive distillation and adaptive feature segmentation, and retain important features to the next layer for attention calculation.

[0051] Furthermore, the specific steps of the hierarchical recursive distillation module are as follows.

[0052] S41. The specific calculation formula for the t-th recursive distillation is: In the formula, the hyperparameter α HRD controlling the recursive weight is initially set to 0.1, and the value range is (0, 1), is the output of the t-th recursive distillation, and Conv is a 3×3 convolution operation.

[0053] S42. Input the second image feature H t 、W t and C are the height, width, and channels of the second image, respectively, set to 320×320×3. The adaptive feature segmentation divides the feature into X en and X fu in two parts according to the adaptive segmentation coefficient. The specific calculation formula of the adaptive segmentation coefficient is: In the formula, X c is the feature of the c-th channel of the feature X t , c ∈ (1, C), and Sig is the sigmoid function; The adaptive feature segmentation formula is: In the formula, X enThe input for the next-level attention calculation, X fu For the final feature fusion.

[0054] S5. Improve the traditional information distillation method to construct an adaptive information distillation network, including feature separation, hierarchical recursive distillation, and cross-scale structure-guided fusion. Design an attention mask based on the local gradient for feature separation, and design an adaptive fusion weight based on the attention gating mechanism to fuse the distillation features of different layers.

[0055] Furthermore, as Figure 3 shown, the specific steps of the adaptive information distillation network are as follows.

[0056] S51. Decompose the pediatric laryngeal and cervical medical image into a key region and a background region, calculate the attention mask of the image, and based on the attention mask, divide the pediatric laryngeal and cervical medical image feature I ∈ R H×W×C into the key region I crux and the background region I cont , where H, W, and C are the height, width, and channels of the pediatric laryngeal and cervical medical image, set to 640×640×3. The specific calculation formula of the attention mask M AM is: In the formula, is the local gradient value, x1 ∈ [0, H], y1 ∈ [0, W], and MAX is to take the maximum value; The calculation formula of the key region I crux is: I crux = M AM ⊙ I; The calculation formula of the background region I cont is: I cont = (1 - M AM ) ⊙ I.

[0057] S52. The information in the key region is more important than that in the background region. It is necessary to assign a higher distillation weight to the key region I crux and obtain the region-adaptive distillation feature according to the distillation weight. The distillation weight formula is: W RAD = Sig(W 1R I crux + W 2R I cont ); In the formula, W 1R and W 2R are trainable weight matrices; The calculation formula of the region-adaptive distillation feature F RAD is: Wherein, ||I crux ⊙I cont ||2 is the L2 norm of the feature map, and e is the base of the natural logarithm.

[0058] S53. Use local pathological attention for attention calculation. The specific formula is: F LPSA =A LPSA (F RAD ); Wherein, A LPSA is the local pathological attention.

[0059] S54. Use hierarchical recursive distillation to gradually extract and enhance features, and divide the features into two parts through an adaptive feature segmentation method, which are respectively used for the next layer of attention calculation and the final feature fusion.

[0060] S55. Fuse the distilled features of different layers. The fusion formula is: Wherein, the distillation level L is set to 16, v i is the adaptive fusion weight, is the distilled feature of the i-th layer; The specific calculation formula of the adaptive fusion weight v i is: Wherein, and are learned through an attention gating mechanism.

[0061] S6. Construct an image enhancement model for pediatric laryngeal and cervical development monitoring, and complete data simulation and image enhancement for pediatric laryngeal and cervical development monitoring through a pediatric laryngeal and cervical data enhancement module and an adaptive information distillation network.

[0062] Furthermore, in step S6, the image of pediatric laryngeal and cervical development monitoring is input into the image enhancement model for pediatric laryngeal and cervical development monitoring. First, data enhancement is performed through the pediatric laryngeal and cervical data enhancement module, and then detailed information is extracted and optimized through the adaptive information distillation network, significantly improving the quality of the image super-resolution for pediatric laryngeal and cervical development monitoring; the image enhancement model for pediatric laryngeal and cervical development monitoring is based on the Pytorch framework, implemented through the Pycharm application program, and the model is trained using 1200 training sets in the image data set for pediatric laryngeal and cervical development monitoring. The trained model is tested using the test set. In the image enhancement task for pediatric laryngeal and cervical development monitoring, not only is the image enhancement effect improved, but the model efficiency is also optimized.

[0063] Furthermore, as Figure 4As shown Figure 4 The left half in the middle is the initial coronal CT image of the larynx Figure 4 The right half in the middle is the coronal CT image of the larynx after being processed by the image enhancement model for pediatric laryngeal and cervical development monitoring, significantly enhancing the clarity and recognizability of the images for pediatric laryngeal and cervical development monitoring.

[0064] 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 still be made, and these all belong to the protection scope of the present invention.

Claims

1. An image enhancement processing method for monitoring the laryngeal and cervical development of children, characterized in that It includes the following steps: S1. Obtain images for pediatric laryngeal and cervical development monitoring and form a dataset; S2. Construct a data augmentation module for pediatric laryngeal and cervical data, design an adaptive non - linear transformation, contrast adaptive enhancement based on structural features, and simulation noise based on the physical properties of medical imaging to simulate the influences in the process of obtaining pediatric laryngeal and cervical image data; S3. Introduce Gaussian variational inference to design an adaptive Gaussian variational attention, construct a pathological mask - guided attention calculation, calculate local attention based on the pathological mask and improved traditional attention, and combine local attention and adaptive Gaussian variational attention to calculate local pathological attention; S4. Design a recursive distillation based on a recursive optimization strategy to gradually extract and enhance features, design an adaptive feature segmentation method based on a self - designed adaptive segmentation coefficient, construct a hierarchical recursive distillation module through recursive distillation and adaptive feature segmentation, and retain important features for attention calculation in the next layer; S5. Improve the traditional information distillation method to construct an adaptive information distillation network, including feature separation, hierarchical recursive distillation, and cross - scale structure - guided fusion. Design an attention mask based on local gradients for feature separation, and design adaptive fusion weights based on the attention gating mechanism to fuse the distilled features of different layers; S6. Construct an image enhancement model for pediatric laryngeal and cervical development monitoring, and complete data simulation and image enhancement for pediatric laryngeal and cervical development monitoring through the pediatric laryngeal and cervical data augmentation module and the adaptive information distillation network.

2. The image enhancement processing method for pediatric laryngeal and cervical development monitoring according to claim 1, wherein, In the S1 step, to obtain images for pediatric laryngeal and cervical development monitoring, capture images of the vocal cords, trachea, and laryngeal soft tissues, and perform data annotation on the obtained images, including key region annotation, development status annotation, and case category annotation. Label the vocal cords, laryngeal cartilages, and trachea as key regions. The development status is divided into three types: normal development, developmental lag, and developmental abnormality. The case category annotation is classified according to the pathological conditions, and the annotated data is divided into a training set and a validation set.

3. The image enhancement processing method for pediatric laryngeal and cervical development monitoring according to claim 2, characterized in that, In the S2 step, the specific steps of the pediatric laryngeal and cervical data augmentation module are as follows: S21. Simulate the patient's posture change, imaging angle deviation, and equipment jitter through adaptive nonlinear transformation. The specific formula is as follows: where, W geo and B geo are spatial mapping matrices based on affine transformation, simulating angular offset and scale change respectively, N is the number of harmonic components, α i is the amplitude parameter, controlling the deformation intensity, β i is the frequency parameter, controlling the local change scale, δ i is the phase parameter, controlling the starting state of deformation, X is the image input to the pediatric laryngeal neck data augmentation module, and sin is the sine function; S22. Highlight the details of the airway, glottis, and muscle tissues adaptively based on structural features and suppress background noise. The specific formula is: Where α con is the edge enhancement intensity, tanh is the hyperbolic tangent function, is the Laplacian operator, and ψ(X) are the mean and standard deviation of X respectively, τ con is the contrast threshold, β con is the contrast enhancement intensity, and the mathematical formula for the contrast enhancement intensity β con is as follows: where exp is the exponential function, and γ targ is the target brightness value; S23. Generate simulation noise based on the physical properties of medical imaging. The specific formula is: where ɑ sn is the equipment noise intensity, Γ(0.01, 0.05) is the equipment noise, and β sn is the tissue scattering noise intensity, Π(-0.3, 0.3) is the tissue scattering noise, which follows a uniform distribution, and γ sn is the dynamic blur intensity, is the time change rate; the mathematical formula for the equipment noise intensity α sn is as follows: where Be is the current load of the device, and Be max is the maximum load of the device, Ti is the usage duration of the device, and Ti max is the maximum usage duration of the device.

4. An image enhancement processing method for pediatric laryngeal and cervical development monitoring according to claim 3, characterized in that, In the S3 step, the specific steps of calculating local pathological attention are as follows: S31. Input the first image feature X i ∈R H×W×C , where H, W, and C are the height, width, and channels of the first image respectively, construct the pathology mask M p to guide the attention calculation, distinguish the key regions of the larynx, namely the vocal cords, trachea, and cartilage, the irrelevant region background, and the low-information region, reduce the computational complexity, calculate the attention of the important regions, and the pathology mask M p is weighted with the optimized Softmax attention to obtain the local attention A PL , and the pathology mask M p The specific calculation formula is as follows: where Z is the normalization factor, δ controls the local smoothness, and x i (i, j) is the gray value at the pixel position (i, j), Ω is a 5×5 local region, α is the local smoothness weight, β is the edge enhancement weight, and are the gradient values at x and y, respectively; The local attention A PL The specific calculation formula is as follows: where G is the expansion order, ⊙ is the weighted operation, Q, K, and V are the query, key, and value respectively, obtained by the linear transformation matrices W Q , W K , W V such that Q = W Q x i , K = W K x i , V = W V x i , T represents the transpose, and d is the dimension of a single attention head; S32. Introduce Gaussian variational inference to optimize attention calculation to obtain adaptive Gaussian variational attention A GVA , and the specific formula is as follows: Where Q i is the Q variable of the Gaussian distribution, Q is the query, μ Q is the mean of Q, ε represents the noise of the standard normal distribution, θ Q is the variance of Q, G i is the adaptive Gaussian compensation factor, p(Q i ) is the variational probability density of Q, σ T is the medical modality adjustment factor, σ T = 0.5·Var(X i ) + 0.5·Mean(X i ), Var(X i ) is the local variance of X i , Mean(Z i ) is the local mean of X i ; The adaptive Gaussian compensation factor G i The specific calculation formula is as follows: where W G is a linear transformation matrix, b G is a Gaussian bias term that controls the reference for compensating the mean, and λ1 and λ2 are learning parameters, where λ1 + λ2 = 1; S33. Combine the local attention A PL and the adaptive Gaussian variational attention A GVA to obtain the local pathological attention A LPSA . The specific formula is as follows: A LPSA = τ·A PL +(1 - τ)·A GVA ; where τ is the attention fusion weight, which is adaptively learned by the neural network.

5. The image enhancement processing method for pediatric laryngeal and cervical development monitoring according to claim 4, wherein In the S4 step, the specific steps of the hierarchical recursive distillation module are as follows: S41. The specific calculation formula for the t - th recursive distillation is: where ɑ HRD is a hyperparameter for controlling the recursive weight, is the output of the t-th recursive distillation, and Conv is the convolution operation; S42. Input the second image feature H t , W t , and C are the height, width, and channels of the second image respectively. The adaptive feature segmentation divides the feature into X en and X fu two parts. The specific calculation formula of the adaptive segmentation coefficient is as follows: where X c is the feature of the c-th channel of feature X t , c ∈ (1, C), and Sig is the sigmoid function; The formula for the adaptive feature segmentation is: where X en is the input for the next layer's attention calculation, and X fu is used for the final feature fusion.

6. The image enhancement processing method for pediatric laryngeal and cervical development monitoring according to claim 5, wherein, In the S5 step, the specific steps of the adaptive information distillation network are as follows: S51. Decompose the pediatric laryngeal and cervical medical image into a key region and a background region, calculate the attention mask of the image, and based on the attention mask, divide the pediatric laryngeal and cervical medical image feature I ∈ R H×W×C into the key region I crux and the background region I cont , where H, W, and C are the height, width, and channels of the pediatric laryngeal and cervical medical image respectively, and the specific calculation formula for the attention mask M AM is as follows: In the formula, is the local gradient value, x1 ∈ [0, H], y1 ∈ [0, W], and MAX is to take the maximum value; The key area I crux The calculation formula is as follows: I crux = M AM ☉I; The background area I cont The calculation formula is as follows: I cont = (1 - M AM ) ⊙ I; S52. The information of the key region is more important than that of the background region, and a higher distillation weight needs to be assigned to the key region I crux to obtain region-adaptive distillation features according to the distillation weight. The distillation weight formula is as follows: W RAD = Sig(W 1R I crux + W 2R I cont ); where, W 1R and W 2R are trainable weight matrices; The region adaptive distilled feature F RAD The calculation formula is as follows: where, ||I crux ⊙I cont ||2 is the L2 norm of the feature map, and e is the base of the natural logarithm; S53. Use local pathological attention for attention calculation. The specific formula is: F LPSA = A LPSA (F RAD ) where A LPSA is the local pathological attention; S54. Use hierarchical recursive distillation to gradually extract and enhance features, and divide the features into two parts through the adaptive feature segmentation method, which are respectively used for attention calculation in the next layer and the final feature fusion; S55. Fuse the distilled features of different layers. The fusion formula is: where L is the distillation level, υ i is the adaptive fusion weight, is the distillation feature of the i-th layer; The specific calculation formula for the adaptive fusion weight υ i is as follows: In the formula, and are learned through the attention gating mechanism.

Images