High-precision detection method based on multi-module adenoid image automatic processing

Abstract

The invention discloses a high-precision detection method based on multi-module adenoid body image automatic processing, and aims to solve the problems that in adenoid body full-automatic diagnosis, as morphological operation of fabric defect detection is very sensitive to the size and shape of defects, priori knowledge is difficult to incorporate, and interference is caused to image detection. The invention discloses a high-precision detection method based on multi-module adenoid image automatic processing, and the method comprises the steps: firstly calculating the local uniformity of each pixel, constructing a new uniformity image, and recording the uniformity image as (H-image); the method comprises the steps that firstly, an H image is obtained, then a classic histogram of the H image is calculated, an optimal threshold value is selected to generate a corresponding binary image, the binary image is used for extracting the size and shape of an optimal structural element (SE) of mathematical morphology, and prior knowledge is fused in the process. Secondly, the SE is used to perform a series of morphological operations on the image, and interferences in different medical images are processed to detect possible fabric defects.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence algorithms and image signal processing cross-processing technology, and in particular to a high-precision detection method based on automatic processing of multi-module adenoid images. Background Art

[0002] With the rapid development of artificial intelligence technology, computer vision has become an important component of modern image signal processing. Medical image processing technology aims to enable computers to acquire, process, analyze and understand information in digital images, thereby realizing the function of fully automatic scientific diagnosis. Traditional image processing technology includes steps such as image enhancement, edge detection and feature extraction. These methods have a series of challenges in medical images, such as complex features and long processing time.

[0003] Problems faced by adenoids detection include insufficient image resolution, subjective interpretation, poor matching between images and actual conditions, and insufficient sensitivity to small changes, all of which may affect the accuracy of diagnosis. This technology combines prior knowledge of organ shape and position to improve the diagnosis rate of adenoids. This method integrates a new regularization model into CNN with anatomical prior knowledge, and the model is trained end-to-end. The new framework encourages the model to follow the global anatomical properties of the underlying anatomy by learning nonlinear representations of shape. Not only that, this method integrates prior information of shape and label structure into neural networks for medical image analysis tasks. In this way, the training process of the neural network can be constrained and the neural network can be guided to make more anatomically meaningful predictions, especially when the input image data information is insufficient or inconsistent, it can easily adapt to different analysis tasks and improve the prediction accuracy of the most advanced models.

[0004] In recent years, with the rise of deep learning and neural network technology, especially the application of computer vision and convolutional neural networks, the strong robustness of deep learning and neural networks in the field of image information recognition has been demonstrated, and fuzzy speech semantic recognition technology has made significant progress. In "Anatomically Constrained Neural Networks (ACNNs): Application to Cardiac Image Enhancement and Segmentation", it is proposed how to incorporate prior knowledge inconspicuously and run it as a pixel classifier to solve the problem of difficult information classification. In "Progress in Multimodality Imaging: Truly Simultaneous Ultrasound and Magnetic Resonance Imaging", it is proposed that there is interference between different medical imaging in image processing, and the theoretical basis for multi-module image generation is given. In "Fabric Defect Detection Using Local Homogeneity and Morphological Image Processing", it is proposed in the field of automatic detection. The morphological operation of fabric defect detection is very sensitive to the size and shape of the defect. Therefore, when applying morphological image processing technology to fabric images for defect detection, software-based morphological operations will not work well. With the continuous innovation and development of technology, computer vision technology will be more accurate and efficient. However, there are huge challenges in the application of zero-error, low-cost and high-performance in the field of medical imaging, so new methods need to be proposed. Summary of the invention

[0005] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a high-precision detection method based on automatic processing of multi-module adenoids images. This method aims at the problem that in the fully automatic diagnosis of adenoids, the morphological operation of fabric defect detection is very sensitive to the size and shape of the defect, which makes it difficult to incorporate prior knowledge and interferes with image detection. The present invention provides a high-precision detection method based on automatic processing of multi-module adenoids images. The method first calculates the local uniformity of each pixel and constructs a new uniformity image, which is recorded as (H-image). Then the classical histogram of the H image is calculated, and an optimal threshold is selected to generate a corresponding binary image. The binary image will be used to extract the size and shape of the optimal structuring element (Structuring Element, SE) of mathematical morphology, and prior knowledge is incorporated in this process. Secondly, the SE is used to perform a series of morphological operations on the image, and the interference in different medical imaging is processed to detect possible fabric defects.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a high-precision detection method based on automatic processing of multi-module adenoid images, the method comprising the following steps:

[0007] Step 1: Fabric defect detection algorithm based on local uniformity and mathematical morphology.

[0008] The original image is scanned with a square window, the local uniformity of each pixel is calculated, and the H image is constructed. Then, each image is histogram equalized and thresholded to obtain a binary image, which will be used to extract the size and shape of the optimal structural element of mathematical morphology. According to the structural element, the binary image is opened and closed by morphological operations to realize fabric defect detection.

[0009] Step 2: MRI-compatible ultrasound imaging system.

[0010] Before acquiring the two imaging modes, the spatial standard between the two imaging planes is established. Based on the lateral fusion technology, the (Magnetic Resonance, MR) imaging is taken to intersect with the transducer and fit along the direction of the ultrasound. Based on the multi-dimensional fusion technology, the reference image of the beads arranged in a triangular manner on two orthogonal planes is used for imaging of the two modes, and each position is adjusted until the spatial standard of the imaging plane is achieved.

[0011] Step 3: Dissect the constrained neural network.

[0012] Image segmentation and super-resolution methods based on convolutional neural networks extend these CNN models with a global training objective to constrain the output space by imposing anatomical shape priors. The new framework encourages the model to follow the global anatomical properties of the underlying anatomy by learning nonlinear representations of shape. Moreover, the prior information of shape and label structure is integrated into the neural network for medical image analysis tasks.

[0013] Furthermore, the step 1 comprises the following steps:

[0014] Step 1-1: Discontinuity test.

[0015] In the texture image, uniformity detection is adopted, where P is represented as a square window, that is, an expanded state with a width of 2N+1 centered on a pixel, and the vector is represented as I c (x c ,y c ), in the window P, we can define different vectors from the central pixel as P i,c =(x j -x c ,y j -y c ), where index i is an integer defined as:

[0016]

[0017] (1) Where P i,c =(x j -x c ,y j -y c ) is the coordinate of the pixel point, (x c ,y c ) represents the central pixel P c The coordinates of .

[0018] According to the definition of the continuous central pixel and the arbitrary pixel, a new vector f is constructed. i as follows:

[0019]

[0020] Furthermore, for a module with width (2N+1), we have (2N+1) 2 -1 vector, all of which satisfy f i :1≤i≤(2N+1) 2 , let f be the sum of all vectors defined in window P, then:

[0021]

[0022] Furthermore, the measured value of H (uniformity value) is defined as the norm of f, then:

[0023] H=||f|| Formula (4)

[0024] Furthermore, in order to limit the large variation of local uniformity, the normalized uniformity is adopted.

[0025]

[0026] Where H max and H min They are the maximum and minimum values ​​obtained by uniformity, and the obtained H image is a different calculation of the estimated value for each pixel. I c (x c ,y c ), whose pixel value is the H value of the pixel uniform value. Because the H value is very small in the uniform area, but very high in the boundary and non-uniform area, the H image can be regarded as a large change to the original image, transforming the uniform area to the dark area and the non-uniform area to the gray area. It can be seen that the H image can be used as a discontinuity measurement in the image.

[0027] Step 1-2: Morphological algorithm

[0028] Digital morphology is applied to represent images, highlighting their appearance, shape and organization. Digital morphology applies nonlinear processes to images to remove details smaller than a certain reference shape (called structural element). Morphological image processing consists of a set of operators that transform images according to the above characteristics. Digital morphology mainly deals with binary images. The basic operations of binary morphology are defined as follows: Dilation is similar to convolution. The structural element is reflected and moved from left to right and then from top to bottom. The algorithm logic is as follows:

[0029]

[0030] The erosion of X by B is described as follows:

[0031]

[0032] The openness of the construction element B to the set X is expressed as:

[0033]

[0034] The closure of the construction element B to the set X is expressed as:

[0035]

[0036] In this algorithm, dilation and erosion on binary images can be viewed as a form of convolution of the algebra of Boolean operations, which are defined between pixels at corresponding positions in two equidimensional images. In addition, opening and closing are two higher-order operations built on dilation and erosion. Due to the connection with Boolean operations, under certain conditions, erosion and dilation can make black pixels white and white pixels black.

[0037] Furthermore, the step 2 comprises the following steps:

[0038] Step 2-1: Encode the enhanced composite CNN.

[0039] EN 2 The composite CNN consists of a k-space completion network and an image domain refinement network. 2 It has convolution kernels, complex convolution strategies and feature enhancement modular units. Considering the k-space encoding process, the EN 2 Convolution kernel improves learning efficiency. Complex convolution strategy is used in both k-space learning and image domain learning to maximize the extraction of complex information from limited MRI data.

[0040] The k-space completion network consists of P encoding enhancement blocks for k-space feature extraction. Each e block contains Q EN 2 A layer of convolutional kernels, a hyperbolic tangent activation layer, and a k-space data consistency check layer.

[0041] Among them EN 2 Convolution kernel In k-space learning, the square convolution kernel maps all data within the square window to a point in the output feature map. 2 Convolution maps k-space data with the same encoding gradient to a feature point in the output feature map. We denote F(EN) and P(EN) as the EN along the frequency encoding and phase encoding directions, respectively. 2 Convolution. In k-space, the encoded gradient field (G x ,G y ) and k-space position (k x ,k y ) is as follows:

[0042]

[0043] Where γ is the gyro magnetic ratio, G x (t) and G y (t) are the phase encoding intensity and frequency encoding gradient as a function of time, and t is the time for collecting data samples. It can be seen from the formula that the k-space data in the same row has the same frequency encoding gradient and contains all signals under the same frequency encoding. Using this prior, the F(EN) convolution is designed to extract k-space features row by row along the frequency encoding direction, and each convolution can use the complete information under frequency encoding. Similarly, the P(EN) convolution is developed to extract k-space features in columns along the phase encoding direction. The convolution formulas of F(EN) and P(EN) are as follows:

[0044] F v (i) = K(i)*E Formula (11)

[0045] Where E represents an EN 2 Convolution kernel, when F(EN) is 1×N x , when P(EN) is N y ×1.F v EN represents the output feature vector, K represents the input K-space data, and i is the row index of F(EN) or the column index of P(EN). 2 The appropriate form of the convolution kernel (F(EN) or P(EN)) will be determined based on the characteristics of the k-space dataset. In the reconstruction network, either F(EN) or P(EN) is selected for k-space learning, rather than using both in sequence. This is different from a single convolution, which decomposes a high-dimensional convolution into several lower-dimensional sequential convolutions.

[0046] Step 2-2: Complex convolution strategy.

[0047] The complex convolution strategy is adopted to improve the feature extraction capability and data utilization efficiency. Assume that the input complex valued k-space data is C=X+iY, where X and Y represent the real component and imaginary component of C respectively. In the complex convolution layer, C is convolved with the complex valued filter W=A+iB, where A and B represent the real part and imaginary part of the filter respectively. The complex convolution is performed as follows:

[0048] C*W=(X+iB)*(A+iB)

[0049] =(X*AY*B)+i(X*B+Y*A) Formula (12)

[0050] Where * denotes the convolution operation. Therefore, the complex convolution can be implemented using four real-valued convolutions that explore the relationship between the real and imaginary parts of the k-space data and the complex-valued filter.

[0051] Step 2-3: Functional enhancement of modular units.

[0052] The function enhancement modular unit adopts residual connection and dense connection. Residual connection can reuse the features of the previous layer with low redundancy, and dense connection can improve the information flow between layers. In the function enhancement modular unit, the input feature map with Y channels is fed to two paths: residual path and dense path. The residual path performs residual connection and modifies the feature map through additive operation. The dense path performs dense connection and adds V new feature channels to the modified feature map through splicing operation. The output of the function enhancement modular unit has Y+V ​​channels, and the formula is:

[0053]

[0054] Where S in and S out They represent the input and output of the functional enhancement modular unit. a and H c are the residual path and the dense path respectively. a and H c Complex convolutional layers with filter sizes of 1×1 and 3×3 are used, respectively, and rectified linear unit activation is adopted. denotes element-wise addition, while || denotes channel-wise connection. Functional enhancement modular units can reuse low-redundant features, making the information flow in the network better.

[0055] Step 2-4: Loss function.

[0056] The loss function of Ls is the sum of the loss function of k-space learning and the loss function of image domain learning with equal weight optimization. The L1 loss function can prevent over-smoothing, and the L2 loss function can ensure robust convergence. The weighting function L w Defined as:

[0057] L w =αL1+L2 Formula (14)

[0058] Where α is a hyperparameter that measures the contribution of L1 loss. Therefore, the total loss function L s It is expressed as:

[0059] L s =L w (y,k rec )+L W (x,I rec ) Formula (15)

[0060] In the formula, k rec k-space data predicted by the k-space completion network. rec This is the final reconstructed adenoid MR image.

[0061] Hyperpolarized MR images may be distorted and degraded by magnetic susceptibility artifacts at the air-tissue interface and motion artifacts caused by respiratory and cardiovascular pulsations. To confirm image quality, we excluded 129Xe images with a signal-to-noise ratio less than 6.6 from downstream analysis. The signal-to-noise ratio was calculated as:

[0062]

[0063] Where, Mean signal Mean is the signal intensity in the adenoids. noise and Std noise are the mean and standard deviation of the background signal intensity outside the adenoids.

[0064] Furthermore, the step 3 comprises the following steps:

[0065] Step 3-1: CNN model segmentation of medical images.

[0066] In most existing methods, the models are supervised purely through local loss functions at the pixel level, without exploiting global dependencies and structures in the output space. For this reason, the predicted global descriptions usually do not follow shape, label, or atlas priors. In contrast, one can include the aforementioned priors in the segmentation model. The proposed framework relies on autoencoders and TL network models to obtain a nonlinear compact representation of the underlying anatomical structure.

[0067] To represent the class label images of different tissue types, further set X = {x i∈R, i∈S} is the observed intensity image. The goal of image segmentation is to estimate y given the observed x. In CNN-based segmentation models, this task is accomplished by learning a discriminant function that is specific to the underlying conditional probability distribution P(y s |x) to model. The convolution function is represented by:

[0068]

[0069] Where C represents the group of eigenvalues, represents the optimal eigenvalue function for a row, Represents the function of finding the optimal eigenvalue of a column. In segmentation training, a block diagram of a stacked convolutional autoencoder network for segmentation label training. The AE model is combined with a predictor network to obtain a compact nonlinear representation that can be extracted from intensity and segmentation images. The SR model is referred to as ACNN-SR, and its training function is as follows:

[0070]

[0071] where λ1 and λ2 determine the weights of the shape prior and weight decay terms, and the smooth l-norm loss function ψ quantifies the reconstruction error. The global loss L hp It is defined as the Euclidean distance between the codes generated from the synthesized HR image and the true HR image.

[0072] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0073] 1. The present invention uses autoencoders and TL networks as regularizers to train neural network models. When testing the image prior, the predictions made by the neural network are consistent with the learned shape model of the basic anatomical structure up to 82.4%. In the case of image damage and artifacts, the regularized model adopted by the present invention can achieve an image restoration rate of 67.1%.

[0074] 2. The fabric defect detection algorithm based on local uniformity and mathematical morphology adopted by the present invention has a high overlap degree between the white area (theoretical defect part) of the output image and the corresponding defect area in the obtained fabric image in the simulation experiment.

[0075] 95.6%. Despite the different sizes and directions of the defects, the algorithm’s false alarm rate was only 2.2% and the missed detection rate was only 2.2%. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a 2D cross-section image processed with high resolution according to the present invention.

[0077] Figure 2 This is a comparison chart of the multimodal imaging performance of the present invention.

[0078] Figure 3 This is a simulation diagram of the A / N fully automatic diagnosis principle of the present invention.

[0079] Figure 4 The figure is a graph showing the accuracy and loss of the model of the present invention on the training set and the validation set. DETAILED DESCRIPTION

[0080] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0081] Example 1

[0082] like Figure 1 As shown, a high-precision detection method based on automatic processing of multi-module adenoid images in this embodiment includes the following steps:

[0083] Step 1: Fabric defect detection algorithm based on local uniformity and mathematical morphology

[0084] The original image is scanned with a square window, the local uniformity of each pixel is calculated, and the H image is constructed. Then, each image is histogram-equalized and thresholded to obtain a binary image, which will be used to extract the size and shape of the optimal structural element of mathematical morphology. According to the structural element, the binary image is opened and closed by morphological operations to realize fabric defect detection;

[0085] Step 2: MRI-compatible ultrasound imaging system

[0086] This method establishes the spatial standard between the two imaging planes before acquiring the two imaging modes. Based on the lateral fusion technology, the MR imaging is taken to intersect with the transducer and fit along the direction of the ultrasound. Based on the multi-dimensional fusion technology, the reference image of the beads arranged in a triangular manner on two orthogonal planes is used for two modal imaging. Each position is adjusted until the spatial standard of the imaging plane is reached.

[0087] Step 3: Dissecting the Constrained Neural Network

[0088] This method is based on image segmentation and super-resolution methods of convolutional neural networks, extending these CNN models with a global training objective to constrain the output space by imposing anatomical shape priors. The new framework encourages the model to follow the global anatomical properties of the underlying anatomy by learning nonlinear representations of shape. Not only that, the prior information of shape and label structure is integrated into the neural network for medical image analysis tasks.

[0089] Example 2

[0090] A high-precision detection method based on automatic processing of multi-module adenoid images in this embodiment, based on Embodiment 1, a fabric defect detection algorithm based on local uniformity and mathematical morphology includes the following steps:

[0091] Step 1-1: Discontinuity test

[0092] In the texture image, uniformity detection is adopted, where P is represented as a square window, that is, an expanded state with a width of 2N+1 centered on a pixel, and the vector is represented as I c (x c ,y c ), in the window P, we can define different vectors from the central pixel as P i,c =(x j -x c ,y j -y c ), where index i is an integer defined as:

[0093]

[0094] (1) Where P i,c =(x j -x c ,y j -y c ) is the coordinate of the pixel point, (x c ,y c ) represents the central pixel P c The coordinates of .

[0095] According to the definition of the continuous central pixel and the arbitrary pixel, a new vector f is constructed. i as follows:

[0096]

[0097] Furthermore, for a module with width (2N+1), we have (2N+1) 2 -1 vector, all of which satisfy f i :1≤i≤(2N+1) 2 , let f be the sum of all vectors defined in window P, then:

[0098]

[0099] Furthermore, the measured value of H (uniformity value) is defined as the norm of f, then:

[0100] H=||f|| Formula (4)

[0101] Furthermore, in order to limit the large variation of local uniformity, the normalized uniformity is adopted.

[0102]

[0103] Where H max and Hmin They are the maximum and minimum values ​​obtained by uniformity, and the obtained H image is a different calculation of the estimated value for each pixel. I c (x c ,y c ), whose pixel value is the H value of the pixel uniform value. Because the H value is very small in the uniform area, but very high in the boundary and non-uniform area, the H image can be regarded as a large change to the original image, transforming the uniform area to the dark area and the non-uniform area to the gray area. It can be seen that the H image can be used as a discontinuity measurement in the image.

[0104] Step 1-2: Morphological algorithm

[0105] Digital morphology is applied to represent images, highlighting their appearance, shape and organization. Digital morphology applies nonlinear processes to images to remove details smaller than a certain reference shape (called structural element). Morphological image processing consists of a set of operators that transform images according to the above characteristics. Digital morphology mainly deals with binary images. The basic operations of binary morphology are defined as follows: Dilation is similar to convolution. The structural element is reflected and moved from left to right and then from top to bottom. The algorithm logic is as follows:

[0106]

[0107] The erosion of X by B is described as follows:

[0108]

[0109] The openness of the construction element B to the set X is expressed as:

[0110]

[0111] The closure of the construction element B to the set X is expressed as:

[0112]

[0113] In this algorithm, dilation and erosion on binary images can be viewed as a form of convolution of the algebra of Boolean operations, which are defined between pixels at corresponding positions in two equidimensional images. In addition, opening and closing are two higher-order operations built on dilation and erosion. Due to the connection with Boolean operations, under certain conditions, erosion and dilation can make black pixels white and white pixels black.

[0114] Example 3

[0115] This embodiment is a high-precision detection method based on multi-module adenoids image automatic processing. Based on Example 2, a nuclear magnetic resonance compatible ultrasound imaging system is established, wherein EN 2It has convolution kernels, complex convolution strategies and feature enhancement modular units. Considering the k-space encoding process, the EN 2 Convolution kernel improves learning efficiency. Complex convolution strategy is used in both k-space learning and image domain learning to maximize the extraction of complex information from limited MRI data.

[0116] Among them EN 2 Convolution kernel In k-space learning, the square convolution kernel maps all data within the square window to a point in the output feature map. 2 Convolution maps k-space data with the same encoding gradient to a feature point in the output feature map. We denote F(EN) and P(EN) as the EN along the frequency encoding and phase encoding directions, respectively. 2 Convolution. In k-space, the encoded gradient field (G x ,G y ) and k-space position (k x ,k y ) is as follows:

[0117]

[0118] Where γ is the gyro magnetic ratio, G x (t) and G y (t) are the phase encoding intensity and frequency encoding gradient as a function of time, and t is the time for collecting data samples. It can be seen from the formula that the k-space data in the same row has the same frequency encoding gradient and contains all signals under the same frequency encoding. Using this prior, the F(EN) convolution is designed to extract k-space features row by row along the frequency encoding direction, and each convolution can use the complete information under frequency encoding. Similarly, the P(EN) convolution is developed to extract k-space features in columns along the phase encoding direction. The convolution formulas of F(EN) and P(EN) are as follows:

[0119] F v (i) = K(i)*E Formula (11)

[0120] Where E represents an EN 2 Convolution kernel, when F(EN) is 1×N x , when P(EN) is N y ×1.F v EN represents the output feature vector, K represents the input K-space data, and i is the row index of F(EN) or the column index of P(EN). 2The appropriate form of the convolution kernel (F(EN) or P(EN)) will be determined based on the characteristics of the k-space dataset. In the reconstruction network, either F(EN) or P(EN) is selected for k-space learning, rather than using both in sequence. This is different from a single convolution, which decomposes a high-dimensional convolution into several lower-dimensional sequential convolutions.

[0121] Step 2-2: Complex convolution strategy

[0122] The complex convolution strategy is adopted to improve the feature extraction capability and data utilization efficiency. Assume that the input complex valued k-space data is C=X+iY, where X and Y represent the real component and imaginary component of C respectively. In the complex convolution layer, C is convolved with the complex valued filter W=A+iB, where A and B represent the real part and imaginary part of the filter respectively. The complex convolution is performed as follows:

[0123] C*W=(X+iB)*(A+iB)

[0124] =(X*AY*B)+i(X*B+Y*A) Formula (12)

[0125] Where * denotes the convolution operation. Therefore, the complex convolution can be implemented using four real-valued convolutions that explore the relationship between the real and imaginary parts of the k-space data and the complex-valued filter.

[0126] Step 2-3: Functionality Enhancement Modular Units

[0127] The function enhancement modular unit adopts residual connection and dense connection. Residual connection can reuse the features of the previous layer with low redundancy, and dense connection can improve the information flow between layers. In the function enhancement modular unit, the input feature map with Y channels is fed to two paths: residual path and dense path. The residual path performs residual connection and modifies the feature map through additive operation. The dense path performs dense connection and adds V new feature channels to the modified feature map through splicing operation. The output of the function enhancement modular unit has Y+V ​​channels, and the formula is:

[0128]

[0129] Where S in and S out They represent the input and output of the functional enhancement modular unit. a and H c are the residual path and the dense path respectively. a and H c Complex convolutional layers with filter sizes of 1×1 and 3×3 are used, respectively, and rectified linear unit activation is adopted. denotes element-wise addition, while || denotes channel-wise connection. Functional enhancement modular units can reuse low-redundant features, making the information flow in the network better.

[0130] Step 2-4: Loss Function

[0131] The loss function of Ls is the sum of the loss function of k-space learning and the loss function of image domain learning with equal weight optimization. The L1 loss function can prevent over-smoothing, and the L2 loss function can ensure robust convergence. The weighting function L w Defined as:

[0132] L w =αL1+L2 Formula (14)

[0133] Where α is a hyperparameter that measures the contribution of L1 loss. Therefore, the total loss function L s It is expressed as:

[0134] L s =L w (y,k rec )+L W (x,I rec ) Formula (15)

[0135] In the formula, k rec k-space data predicted by the k-space completion network. rec This is the final reconstructed adenoid MR image.

[0136] Hyperpolarized MR images may be distorted and degraded by magnetic susceptibility artifacts at the air-tissue interface and motion artifacts caused by respiratory and cardiovascular pulsations. To confirm image quality, we excluded 129Xe images with a signal-to-noise ratio less than 6.6 from downstream analysis. The signal-to-noise ratio was calculated as:

[0137]

[0138] Where, Mean signal Mean is the signal intensity in the adenoids. noise and Std noise are the mean and standard deviation of the background signal intensity outside the adenoids.

[0139] Example 4

[0140] This embodiment provides a high-precision detection method based on automatic processing of multi-module adenoid images, based on Embodiment 3, using an anatomically constrained neural network.

[0141] In most existing methods, the models are supervised purely through local loss functions at the pixel level, without exploiting global dependencies and structures in the output space. For this reason, the predicted global descriptions usually do not follow shape, label, or atlas priors. In contrast, one can include the aforementioned priors in the segmentation model. The proposed framework relies on autoencoders and TL network models to obtain a nonlinear compact representation of the underlying anatomical structure.

[0142] To represent the class label images of different tissue types, further set X = {x i ∈R, i∈S} is the observed intensity image. The goal of image segmentation is to estimate y given the observed x. In CNN-based segmentation models, this task is accomplished by learning a discriminant function that is specific to the underlying conditional probability distribution P(y s |x) for modeling. The convolution function is represented by:

[0143]

[0144] Where C represents the group of eigenvalues, represents the optimal eigenvalue function for a row, Represents the function of finding the optimal eigenvalue of a column. In segmentation training, a block diagram of a stacked convolutional autoencoder network for segmentation label training. The AE model is combined with a predictor network to obtain a compact nonlinear representation that can be extracted from intensity and segmentation images. The SR model is referred to as ACNN-SR, and its training function is as follows:

[0145]

[0146] where λ1 and λ2 determine the weights of the shape prior and weight decay terms, and the smooth l-norm loss function ψ quantifies the reconstruction error. The global loss L hp It is defined as the Euclidean distance between the codes generated by the synthetic HR image and the real HR image. In the training project, the discriminator uses the A / N indicator to distinguish and combines the existing prophetic knowledge to improve the accuracy of adenoid diagnosis.

[0147] Example 5

[0148] A high-precision detection method based on automatic processing of multi-module adenoid images in this embodiment is based on Embodiment 4 to perform medical data diagnosis simulation. First, simulation hardware conditions are prepared, and then simulation system parameters are set, which specifically include the following:

[0149] 1. Simulation hardware conditions

[0150] The following simulation experiment is conducted on the simulation platform of Python 3.12.4 and TensorFlow 2.70. The computer CPU model is E5-2680 v4, the number is 6, and the GPU model is NVIDIA Geforce RTX 3060. The video memory is 12.6GB.

[0151] 2. Simulation system parameters

[0152] The dataset consists of 1200 images Figure 2 The 2D stack consists of short-axis and cine 3D high-resolution cardiac MR images. Each image pair was acquired from healthy subjects using a standard imaging protocol. The 2D stack was acquired at different breath holds and contains motion artifacts. Similarly, 3D imaging is essential in a standard clinical setting due to the need for long image acquisition. The voxel resolution of the 2D stacked low-resolution images is fixed at 1.25×1.25×10.00mm, and the voxel resolution of the HR images is fixed at 1.25×1.25×2.00mm.

[0153] 3. Simulation content

[0154] Figure 1 The anatomical constrained SR model diagram obtained by the fabric defect detection algorithm based on the anatomical constrained neural network, local uniformity and mathematical morphology. The resolution increases from left to right. Figure 2 The researchers performed simultaneous MRI and ultrasound imaging of the adenoids, using high-resolution MRI. Although larger chords (0.4 mm) can be distinguished on MRI images, ultrasound images are much more sensitive to them. This highlights the versatility and efficiency of multimodal fusion. Figure 3 It shows that in non-model diagnosis, the accuracy of point annotation can reach 87%. Figure 4 It represents the accuracy and loss curve of the model on the training set and the validation set. It can be seen from the figure that the accuracy is stable at around 80% and the loss rate is below 10%.

[0155] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A high-precision detection method based on automatic processing of multi-module adenoid images, characterized in that: The method comprises the following steps: Step 1: Fabric defect detection algorithm based on local uniformity and mathematical morphology; The original image is scanned with a square window, the local uniformity of each pixel is calculated, and the H image is constructed. Then, each image is histogram equalized and thresholded to obtain a binary image. The binary image will be used to extract the size and shape of the optimal structural element of mathematical morphology. The binary image is opened and closed according to the structural element to realize fabric defect detection. Step 2: MRI-compatible ultrasound imaging system; Before acquiring the two imaging modes, the spatial standard between the two imaging planes is established. Based on the lateral fusion technology, the (Magnetic Resonance, MR) imaging is taken to intersect with the transducer and fit along the direction of the ultrasound. Based on the multi-dimensional fusion technology, the reference image of the beads arranged in a triangular manner on two orthogonal planes is used for imaging of the two modes. Each position is adjusted until the spatial standard of the imaging plane is achieved. Step 3: Dissecting the Constrained Neural Network, Image segmentation and super-resolution methods based on convolutional neural networks extend these CNN models with a global training objective. By imposing anatomical shape priors to constrain the output space, the new framework encourages the model to follow the global anatomical properties of the underlying anatomy by learning nonlinear representations of shape. Not only that, the prior information of shape and label structure is integrated into the neural network for medical image analysis tasks.

2. A high-precision detection method based on automatic processing of multi-module adenoid images according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1-1: Discontinuity test; In the texture image, uniformity detection is adopted, where P is represented as a square window, that is, an expanded state with a width of 2N+1 centered on a pixel, and the vector is represented as I c (x c ,y c ), in the window P, we can define different vectors from the central pixel as P i,c =(x j -x c ,y j -y c ), where index i is an integer defined as: (1) Where P i,c =(x j -x c ,y j -y c ) is the coordinate of the pixel point, (x c ,y c ) represents the central pixel P c The coordinates of According to the definition of the continuous central pixel and the arbitrary pixel, a new vector f is constructed. i as follows: For a module with width (2N+1), we have (2N+1) 2 -1 vector, all of which satisfy f i :1≤i≤(2N+1) 2 , let f be the sum of all vectors defined in window P, then: The measured value of H (uniformity value) is defined as the norm of f, then: H=||f|| Formula (4) In order to limit the large variation of local uniformity, the normalized uniformity is adopted Where H max and H min They are the maximum and minimum values ​​obtained by uniformity, and the obtained H image is a different calculation of the estimated value for each pixel. I c (x c ,y c ), whose pixel value is the H value of the pixel uniform value. Because the H value is very small in the uniform area, but very high in the boundary and non-uniform area, the H image can be regarded as a large change to the original image, transforming the uniform area to the dark area and the non-uniform area to the gray area. It can be seen that the H image can be used as a discontinuity measurement in the image. Step 1-2: Morphological Algorithm Digital morphology is applied to represent images, highlighting their appearance, shape and organization. Digital morphology is applied to nonlinear processes of images to remove details smaller than a certain reference shape (called structural element). Morphological image processing consists of a set of operators that transform images according to the above characteristics. Digital morphology mainly processes binary images. The basic operations of binary morphology are defined as follows: Dilation is similar to convolution. The structural element is reflected and moved from left to right and then from top to bottom. The algorithm logic is as follows: The erosion of X by B is described as follows: The openness of the construction element B to the set X is expressed as: The closure of the construction element B to the set X is expressed as: In this algorithm, dilation and erosion on binary images can be seen as a form of convolution of Boolean algebra, which are operations defined between pixels at corresponding positions in two equal-dimensional images. In addition, opening and closing are two higher-order operations based on dilation and erosion. Due to the connection with Boolean operations, under certain conditions, erosion and dilation can make black pixels white and white pixels black.

3. A high-precision detection method based on automatic processing of multi-module adenoid images according to claim 1, characterized in that: The step 2 comprises the following steps: Step 2-1: Encoding enhanced composite CNN; EN 2 The composite CNN consists of a k-space completion network and an image domain refinement network. 2 It has convolution kernels, complex convolution strategies and feature enhancement modular units. Considering the k-space encoding process, the EN 2 Convolution kernel improves learning efficiency and adopts complex convolution strategy in both k-space learning and image domain learning to maximize the extraction of complex information from limited MRI data; The k-space completion network consists of P encoding enhancement blocks for k-space feature extraction, each e block contains Q EN 2 A layer of convolution kernels, a hyperbolic tangent activation layer, and a k-space data consistency check layer; Among them EN 2 In k-space learning, the square convolution kernel maps all data within the square window to a point in the output feature map, illustrating the EN 2 Convolution maps k-space data with the same encoding gradient to a feature point in the output feature map. We denote F(EN) and P(EN) as the EN along the frequency encoding and phase encoding directions, respectively. 2 Convolution, in k-space, encodes the gradient field (G x ,G y ) and k-space position (k x ,k y ) is as follows: Where γ is the gyro magnetic ratio, G x (t) and G y (t) are the phase encoding intensity and frequency encoding gradient as a function of time, t is the time of collecting data samples, and the k-space data in the same row has the same frequency encoding gradient, containing all signals under the same frequency encoding. Using this prior, the F(EN) convolution is designed to extract k-space features row by row along the frequency encoding direction. Each convolution can use the complete information under frequency encoding. Similarly, the P(EN) convolution is developed to extract k-space features in columns along the phase encoding direction. The convolution formulas of F(EN) and P(EN) are as follows: F v (i) = K(i) * Ε Equation (11) Where E represents an EN 2 Convolution kernel, when F(EN) is 1×N x , when P(EN) is N y ×1, F v represents the output feature vector, K represents the input K-space data, i is the row index of F(EN) or the column index of P(EN), EN 2 The appropriate form of the convolution kernel (F(EN) or P(EN)) will be determined based on the characteristics of the k-space dataset. In the reconstruction network, F(EN) or P(EN) is selected for k-space learning instead of using both in sequence, which is different from a single convolution, which decomposes a high-dimensional convolution into several low-dimensional sequential convolutions. Step 2-2: complex convolution strategy; The complex convolution strategy is adopted to improve the feature extraction capability and data utilization efficiency. Assume that the input complex-valued k-space data is C=X+iY, where X and Y represent the real component and imaginary component of C respectively. In the complex convolution layer, C is convolved with the complex-valued filter W=A+iB, where A and B represent the real part and imaginary part of the filter respectively. The complex convolution is performed as follows: C*W=(X+iB)*(A+iB) =(X*AY*B)+i(X*B+Y*A) Formula (12) Where * represents the convolution operation, so the complex convolution can be implemented using four real-valued convolutions that explore the relationship between the real and imaginary parts of the k-space data and the complex-valued filter, Step 2-3: Functional enhancement of modular units; The function enhancement modular unit adopts residual connection and dense connection. The residual connection can reuse the features of the previous layer and has low redundancy. The dense connection can improve the information flow between layers. In the function enhancement modular unit, the input feature map with Y channels is fed to two paths: the residual path and the dense path. The residual path performs residual connection and modifies the feature map through additive operation. The dense path performs dense connection and adds V new feature channels to the modified feature map through splicing operation. The output of the function enhancement modular unit has Y+V ​​channels, and the formula is: Where S in and S out They represent the input and output of the functional enhancement modular unit, H a and H c are the residual path and the dense path respectively, H a and H c The complex convolutional layers with filter sizes of 1×1 and 3×3 are used respectively, and the rectified linear unit activation is used. represents addition by element, while || represents connection by channel. Functional enhancement modular units can reuse low-redundant features, making the information flow in the network better; Step 2-4: Loss function; The loss function of Ls is the sum of the loss function of k-space learning and the loss function of image domain learning with equal weight optimization. The L1 loss function can prevent over-smoothing, and the L2 loss function can ensure robust convergence. The weighting function L w Defined as: L w =αL1+L2 Formula (14) Where α is a hyperparameter that measures the contribution of L1 loss, so the total loss function L s It is expressed as: L s =L w (y,k rec )+L W (x,I rec ) Formula (15) In the formula, k rec k-space data predicted by the k-space completion network, I rec The final reconstructed adenoid MR image; Since the magnetic susceptibility artifacts at the air-tissue interface and the motion artifacts caused by respiratory and cardiovascular pulsations may distort and degrade the hyperpolarized MR images, in order to confirm the image quality, 129Xe images with a signal-to-noise ratio less than 6.6 were excluded from downstream analysis. The signal-to-noise ratio was calculated as: Where, Mean signal Mean is the signal intensity in the adenoids. noise and Std noise are the mean and standard deviation of the background signal intensity outside the adenoids.

4. A high-precision detection method based on automatic processing of multi-module adenoid images according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3-1: CNN model segmentation of medical images; A model that can include the above priors in the segmentation model. The proposed framework relies on autoencoders and TL network models to obtain nonlinear and compact representations of the underlying anatomical structures; To represent the class label images of different tissue types, further set X = {x i ∈R, i∈S} is the observed intensity image. The purpose of image segmentation is to estimate the observed y. In the CNN-based segmentation model, this task is accomplished by learning a discriminant function that is responsible for the underlying conditional probability distribution P(y s |x) is modeled, where the convolution function is expressed as: Where C represents the group of eigenvalues, represents the optimal eigenvalue function for a row, Indicates the optimal eigenvalue function for the column. In segmentation training, the block diagram of the stacked convolutional autoencoder network for segmentation label training. The AE model is combined with the predictor network to obtain a compact nonlinear representation that can be extracted from the intensity and segmentation images. The SR model is referred to as ACNN-SR, and its training function is as follows: where λ1 and λ2 determine the weights of the shape prior and weight decay terms, while the smooth l-norm loss function ψ quantifies the reconstruction error and the global loss L hp It is defined as the Euclidean distance between the codes generated from the synthesized HR image and the true HR image.

Images