PET brain image choroid development elimination method and storage medium

Abstract

The invention provides a PET brain image choroid development elimination method and a storage medium, and the method comprises an image acquisition step, a preprocessing step, a model establishment step, a model training step and an input step, and comprises the following steps: acquiring PET negative and positive development images of TSPO (18F-DPA714) and Tau protein (18F-Florootau); in combination with clinical image data of neurodegenerative diseases such as Alzheimer's disease (AD) and progressive suprakaryotic paralysis (PSP), high metabolic signals of choroids are accurately distinguished by using a GAN model; a deep learning generative adversarial network is adopted, a PET image processing flow is optimized, and interference of choroid high-development signals is reduced; by comparing positive and negative developing images, the developing precision of lesion tissues is enhanced, automatic elimination of choroid signals in PET images is realized, and the accuracy of image analysis is improved.

Description

Technical Field

[0001] This application relates to the field of machine learning, and particularly to a method for eliminating choroid imaging in PET brain images and a storage medium. Background Art

[0002] The choroid plexus is a highly vascularized tissue located in the ventricles of the brain. Its main functions are to regulate the composition and pressure of cerebrospinal fluid and play an important role in maintaining the metabolic balance of the brain. Due to its dense vascular distribution and high metabolic activity, the choroid plexus often shows strong radioactive tracer uptake in positron emission tomography (PET) imaging. This phenomenon is mainly due to the high blood supply of the choroid plexus tissue, which leads to the accumulation of radioactive drugs locally, and then shows a strong metabolic activity signal.

[0003] In the clinical and research applications of PET imaging, especially in the research of Alzheimer's disease, Parkinson's disease, brain tumors and other neurodegenerative diseases, accurately analyzing the metabolic status and radioactive drug distribution in specific brain regions is of crucial significance for disease diagnosis and progression monitoring. However, the strong imaging of the choroid plexus often interferes with the observation of the target brain region, bringing a series of technical problems and diagnostic uncertainties.

[0004] Choroid plexus imaging usually appears as a strong signal, covering the lower brain region or adjacent to important brain regions (such as the basal ganglia, hippocampus). This not only masks the radioactive distribution in the lesion area but may also cause the signal in the lesion area to be misinterpreted as the metabolic signal of the choroid plexus. This high imaging often blurs the boundaries of key pathological changes, resulting in a reduction in the resolution and diagnostic accuracy of PET images, especially at the brain edge and the junction of the choroid plexus.

[0005] The presence of high metabolic imaging of the choroid plexus in PET images also significantly affects the determination of the uptake amount of radioactive tracers in the target tissue. In quantitative analysis, the uptake value of radioactive tracers (such as the standardized uptake value, SUV) is usually used to evaluate the metabolic activity of brain tissue. However, when the interference of choroid plexus imaging is close to the target tissue, the SUV value may deviate, leading to misjudgment. For example, when evaluating tau deposition in Alzheimer's disease patients, the high signal of the choroid plexus may be misinterpreted as the radioactive drug uptake in gray matter or cortex, thus affecting the judgment of disease progression.

[0006] The interference of choroidal imaging is particularly obvious in the diagnosis of certain neurological diseases. For example, in the early diagnosis of Alzheimer's disease, researchers rely on PET imaging to detect the distribution of tau protein in the brain. However, due to the presence of choroidal imaging, the deposition area of tau in the PET image may overlap with the choroidal signal, resulting in false positive or false negative results and increasing the risk of misdiagnosis. Similarly, when studying diseases such as inflammation and brain tumors, choroidal imaging may also obscure the true lesion area, leading to misjudgment. In modern medical imaging diagnosis, PET / MRI fusion images are widely used to provide comprehensive information on metabolism and anatomical structure. PET imaging provides the metabolic state of the brain, while MRI is better at showing the details of soft tissues. However, since the choroid appears as a high signal in the PET image and a low signal in the MRI image, this inconsistency in signal characteristics will interfere with the fusion of the two images. The degradation of the fusion image quality not only affects the accurate judgment of brain structure and metabolic activities, but also reduces the diagnostic efficiency and increases the risk of misjudgment. Especially in the diagnosis and preoperative evaluation of neurological diseases, this quality loss will seriously affect the doctor's judgment and thus affect the patient's treatment plan.

[0007] In the field of PET imaging, traditional image processing methods face major challenges from the interference of choroidal imaging. Conventional methods, such as thresholding and morphological operations, although having certain effects in denoising or segmentation, often have unsatisfactory results when dealing with the choroid with high contrast and complex structure. First of all, the thresholding method relies on setting a fixed intensity threshold to segment the high signal area. However, since the intensity of the choroid is similar to that of the metabolic signals in other brain regions, the thresholding method is prone to losing important information or retaining false signals. Morphological operations are also prone to generating artifacts in some scenarios, distorting the image.

[0008] These traditional methods lack a deep understanding of the image structure and signals, and it is difficult to accurately identify and remove the choroidal imaging signal without damaging the information of adjacent brain regions. This method is obviously inadequate when dealing with highly complex brain images, especially when performing accurate quantitative analysis on the lesion area. To meet the needs of clinical and scientific research, a more intelligent solution that can flexibly adapt to different image conditions is required. Summary of the Invention

[0009] This application provides a method for eliminating choroidal imaging in PET brain images and a storage medium to solve the problem of choroidal imaging in PET brain images.

[0010] This application provides a method for eliminating choroidal imaging in PET brain images, including an image acquisition step, a preprocessing step, a model establishment step, a model training step, and an input step.

[0011] The image acquisition step is used to acquire at least one set of PET brain images, and the PET brain images include choroid imaging regions; the preprocessing step is to convert the PET brain images from the DCM format to the NII format, and then preprocess the PET brain images in the NII format to obtain an initial image set, and the initial image set includes at least one initial image; the model establishment step is used to construct an elimination model, and the elimination model includes a generator and a discriminator, and the generator and the discriminator form an adversarial mechanism. Input the initial image into the elimination model, extract image features of the initial image by introducing dilated convolution and attention mechanism, extract the imaging features and background features of each initial image, and based on the imaging features and the background features, use the adversarial mechanism to perform choroid imaging elimination and repair processing on the initial image to obtain a final PET brain image with choroid imaging removed; the model training step inputs the initial image set into the elimination model, the generator is used to generate a reconstructed image with choroid imaging removed, the discriminator compares the reconstructed image with the real image, and improves the image quality of the reconstructed image output by the generator through the adversarial mechanism. When the image quality meets the standard, stop training to obtain a trained elimination model, where the real image is an image corresponding to the initial image without choroid imaging; the input step is to input the initial image into the trained elimination model to obtain the final PET brain image.

[0012] Further, the preprocessing step specifically includes an origin correction step, a spatial normalization step, a smoothing process step, and a sliding window sampling step.

[0013] The origin correction step is to perform AC-PC correction on the brain part in the PET brain image, unify the spatial origin of different PET brain images, set the origin position of the PET brain image at the AC to obtain a first image; the spatial normalization step is to register the first image to the Montreal Neurological Institute standard brain template space in sequence to unify the coordinate space of all first images to obtain a second image; the smoothing process step is to smooth the second image using a Gaussian filter to obtain a third image; the sliding window sampling step is to sample the third image with a sliding window of a preset size and a preset step length to obtain an initial image.

[0014] Further, the model training step specifically includes a dilated convolution step, a weight update step, and an imaging elimination step.

[0015] The dilated convolution step is to expand the receptive field of the convolution kernel by setting a dilation factor, and can effectively capture the global features and local features in the initial image with complex geometric shapes while keeping the number of convolution kernel parameters unchanged. Its formula is

[0016]

[0017] Among them, Y(t) represents a feature map with a larger receptive field, t represents the position of the image pixel, k represents the size of the convolutional kernel, i represents the index of the convolutional kernel, X represents the initial image, and M represents the convolutional kernel;

[0018] The weight update step is to optimize the feature extraction process by constructing an adaptive attention weight distribution to obtain a final feature map, obtain the contour of the choroid development region based on the residual map, enhance the features of the development region through the final feature map, and suppress the features of the background interference region to obtain the development features and the background features; the development elimination step is used to calculate and obtain the third weight of the elimination model for specific voxels of the image, update the adversarial mechanism through the third weight to obtain the final adversarial mechanism, and based on the development features and the background features, use the final adversarial mechanism to perform choroid development elimination and repair processing on the initial image to obtain the final PET brain image without choroid development.

[0019] Further, the weight update step specifically includes a feature extraction step, a first weight distribution calculation step, a second weight distribution calculation step, and a feature weight adjustment step.

[0020] The feature extraction step is to input the initial image into the elimination model to extract a multi-channel feature map F, where F ∈ R C×H×W , where C represents the number of feature channels, H represents the height of the feature map, W represents the width of the feature map, and R is the set of real numbers; the first weight distribution calculation step is to calculate the first global average pooling and the first global maximum pooling on each channel for the feature map F to obtain the global information describing the channel characteristics and obtain the importance weight of the channel dimension, and its formula is

[0021]

[0022]

[0023]

[0024] Among them, represents the first global average pooling, represents the first global maximum pooling, MLP represents the multi-layer perceptron, σ represents the sigmoid activation function, α c ∈[0,1], and α cThe first weight representing the channel dimension; the second weight distribution calculation step is to calculate the second global average pooling and the second global max pooling on the spatial dimension of the feature map F to generate a two-dimensional feature map and obtain the spatial importance weight, and its formula is

[0025]

[0026]

[0027]

[0028] wherein, represents the second global average pooling, represents the second global max pooling, CONV represents the convolution operation, represents the feature stacking operation, β(i,j) ∈ [0,1], and β(i,j) represents the second weight of the spatial position (i,j); the feature weight adjustment step is based on the obtained first weight and second weight to adjust the channel dimension and the spatial dimension of the feature map F to obtain the adjusted final feature map.

[0029] Furthermore, the development elimination step specifically includes a matrix calculation step, a score calculation step, a normalization step, and a weighted summation step.

[0030] The matrix calculation step is used to calculate and obtain a query matrix, a key matrix, and a value matrix, and its formula is

[0031] Q = XW q ;

[0032] K = XW k ;

[0033] V = XW v ;

[0034] wherein, Q represents the query matrix; K represents the key matrix; V represents the value matrix; W q represents the query weight matrix for projecting the input data matrix X into the query matrix Q; W k represents the key weight matrix for generating the key matrix K; W v represents the value weight matrix for generating the value matrix V, and X represents the data matrix;

[0035] The score calculation step is used to calculate the dot product score between the query matrix and the key matrix and scale the dot product score, and its formula is

[0036]

[0037] Among them, scaled score(Q,K) represents the dot product score of the query matrix and the key matrix d, and d k represents the dimensions of the query matrix and the key matrix;

[0038] The normalization step is to perform Softmax normalization on the obtained dot product score to obtain the attention weights, and its formula is

[0039]

[0040] Among them, a ij represents the attention weights; the weighted summation step is to perform weighted summation on the value matrix based on the obtained attention weights to obtain the third weights, and its formula is

[0041]

[0042] Among them, Attention(Q,K,V) represents the third weights.

[0043] Furthermore, the feature weight adjustment step specifically includes a channel dimension adjustment step and a spatial dimension adjustment step.

[0044] The channel dimension adjustment step is to adjust the channel dimension of the feature map F based on the obtained first weights, and its formula is

[0045] F′ C = α C ·F C

[0046] Among them, F C represents the feature map of the c-th channel, and F' C represents the adjusted feature map of the c-th channel; the spatial dimension adjustment step is to adjust the spatial dimension of the feature map F based on the obtained second weights, and its formula is

[0047] F′ ij = β(i,j)·F ij

[0048] Among them, F ij represents the value at the position (i,j) in the feature map F, and F' ij represents the adjusted value at the position (i,j) in the feature map F; after the feature map F performs the channel dimension adjustment step and the spatial dimension adjustment step, the final feature map is obtained.

[0049] This application also proposes a storage medium storing computer-readable instructions, which, when read by at least one processor, cause the at least one processor to execute at least one step in the multi-device multi-channel power data synthesis method described above.

[0050] This application provides a method for eliminating choroid imaging in PET brain images and a storage medium. This method is mainly applied to the problem of high-signal imaging of the brain choroid in positron emission tomography (PET). By collecting TSPO ( 18 F-DPA714) and Tau protein ( 18 F-Florzolotau) PET negative and positive imaging images, combined with the clinical imaging data of neurodegenerative diseases such as Alzheimer's disease (AD) and progressive supranuclear palsy (PSP), the GAN model is used to accurately distinguish the hypermetabolic signals of the choroid; the generative adversarial network of deep learning is adopted to optimize the PET image processing process and reduce the interference of the high-imaging signals of the choroid; by comparing the positive and negative imaging images, the imaging accuracy of the diseased tissue is enhanced, and the automatic elimination of the choroid signal in the PET image is realized, improving the accuracy of image analysis. Brief Description of the Drawings

[0051] In order to more clearly illustrate the technical solutions in the embodiments of this application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of this application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0052] Figure 1 is the flowchart of the method for eliminating choroid imaging in PET brain images described in the embodiments of this application;

[0053] Figure 2 is the flowchart of the preprocessing steps described in the embodiments of this application;

[0054] Figure 3 is the schematic diagram of the elimination model described in the embodiments of this application;

[0055] Figure 4 is the flowchart of the model training steps described in the embodiments of this application;

[0056] Figure 5 is the flowchart of the weight update steps described in the embodiments of this application;

[0057] Figure 6 is the flowchart of the feature weight adjustment steps described in the embodiments of this application;

[0058] Figure 7 is the flowchart of the imaging elimination steps described in the embodiments of this application;

[0059] Figure 8 is the schematic diagram of the storage medium described in the embodiments of this application.

[0060] Description of the Reference Numerals:

[0061] 100 storage medium, 110 processor, 120 memory. Specific implementation manner

[0062] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0063] As Figure 1 shown, the present application provides a method for eliminating choroid imaging in PET brain images, including step S1) image acquisition step, step S2) preprocessing step, step S3) model establishment step, step S4) model training step, and step S5) input step.

[0064] Step S1) Image acquisition step, acquiring at least one set of PET brain images, where the PET brain images include choroid imaging regions.

[0065] In this embodiment, acquire brain 18 F-Florzolotau and 18 F-DPA714 PET images. Where PET refers to positron emission tomography imaging; 18 F-Florzolotau refers to a compound containing the radioactive isotope fluorine-18, which is mainly used for positron emission tomography (PET) imaging to detect tau protein in the brain.

[0066] Step S2) Preprocessing step, converting the PET brain images from DCM format to NII format, and then preprocessing the NII format PET brain images to obtain an initial image set, where the initial image set includes at least one initial image.

[0067] As Figure 2 shown, step S2) the preprocessing step specifically includes step S21) origin correction step, step S22) spatial normalization step, step S23) smoothing processing step, and step S24) sliding window sampling step.

[0068] Step S21) Origin correction step, performing AC-PC correction on the brain part in the PET brain image, unifying the spatial origin of different PET brain images, and setting the origin position of the PET brain image at the AC to obtain a first image.

[0069] Step S22) Spatial normalization step: sequentially register the first image into the space of the Montreal Neurological Institute standard brain template to unify the coordinate spaces of all the first images, and obtain a second image.

[0070] In this embodiment, due to the different shapes and sizes of the brains of different subjects, the images cannot be well aligned spatially. The first image will be sequentially registered into the space of the Montreal Neurological Institute standard brain template to unify the coordinate spaces of all the images. The main registration methods include linear registration and non-linear registration: linear registration includes linear coordinate transformation and affine transformation, and non-linear registration performs non-linear transformation on the local area.

[0071] Step S23) Smoothing processing step: use a Gaussian filter to perform smoothing processing on the second image to reduce the noise of the second image, thereby further improving the signal-to-noise ratio of the second image, and obtain a third image.

[0072] Step S24) Sliding window sampling step: sample the third image with a sliding window of a preset size and a preset step length to obtain an initial image to ensure the consistency of the input data; for different types of image samples, adopt an adaptive enhancement method to improve the detectability of the developed area.

[0073] As Figure 3 shown, step S3) Model establishment step: construct an elimination model, the elimination model includes a generator and a discriminator, the generator and the discriminator form an adversarial mechanism, input the initial image into the elimination model, perform image feature extraction on the initial image by introducing dilated convolution and attention mechanism, extract the developed features and background features of each initial image, and based on the developed features and the background features, use the adversarial mechanism to perform choroidal development elimination and repair processing on the initial image to obtain a final PET brain image with choroidal development removed.

[0074] In this embodiment, in order to effectively cope with the interference of choroidal development on PET image analysis and improve the performance of the network in processing complex medical images, this method proposes to introduce an attention mechanism and a dilated convolution layer into the basic network to enhance the extraction of effective features and suppress irrelevant or interfering features. This improvement not only improves the model's ability to process global and local features, but also enhances the model's recognition and elimination effect on the choroidal high-signal area.

[0075] The channel attention mechanism assigns different weights to each channel by analyzing the dependencies between different channels. Specifically, when the convolutional layer of a deep learning network outputs a feature map, different channels represent different levels of features. The channel attention mechanism can highlight task-related features and suppress irrelevant or noisy features by adaptively adjusting the importance of each channel; when processing PET images, the strong signal interference caused by choroid visualization often manifests in multiple channels. By introducing the channel attention mechanism, the network can adaptively assign different weights to each channel, thereby reducing the influence of channels affected by choroid interference and strengthening those channels that contain key brain information. This mechanism helps the network more accurately identify lesion areas and avoid misjudging choroid visualization as lesion signals.

[0076] The spatial attention mechanism assigns different weights to each pixel in the image by analyzing the relationship between the local features and global features of each region in the image. Specifically, the spatial attention mechanism can help the network pay more attention to important position regions in the image. Especially when the lesion area may be masked under the influence of choroid visualization, it can effectively improve the network's perception ability of the lesion area; in PET images, the high-signal region of the choroid may be confused with the low-signal regions of other brain regions, especially when the choroid is adjacent to an important lesion area. Traditional convolutional operations are difficult to effectively distinguish these regions. By introducing the spatial attention mechanism, the network can capture the global and local feature distributions in the image, assign higher weights to important lesion areas, thereby suppressing the high-signal interference of the choroid and improving the resolution and recognition effect of the lesion area.

[0077] To further enhance the synergistic effect between channels and space, the present invention proposes to add a cross-attention mechanism to the attention module. The cross-attention mechanism can capture more complex feature relationships in the image by establishing interactive dependencies in the channel and spatial dimensions; through the collaborative optimization across channels and space, the cross-attention mechanism further improves the network's processing ability for complex structures, making the model more robust and efficient in dealing with the high-signal interference of choroid visualization.

[0078] To further enhance the receptive field of the model and improve the ability to capture global features, the present invention proposes to add a dilated convolutional layer to the network. Dilated convolution can effectively expand the receptive field by inserting holes between the convolutional kernels without increasing the number of parameters or computational complexity. When processing PET images, the dilated convolutional layer can help the network better capture the global relationship between the choroid and other brain regions, avoiding misjudgment caused by blurred local features or unclear boundaries. Especially for those lesion areas located near the choroid, dilated convolution can effectively improve the network's recognition ability for these complex boundaries.

[0079] Step S4) Model training step: The initial image set is input into the elimination model. The generator is used to generate a reconstructed image with choroid development removed. The discriminator compares the reconstructed image with the real image, and through the adversarial mechanism, improves the image quality of the reconstructed image output by the generator. When the image quality reaches the standard, the training is stopped, and the trained elimination model is obtained. Among them, the real image is the image corresponding to the initial image without choroid development; the input step is to input the initial image into the trained elimination model to obtain the final PET brain image.

[0080] In this embodiment, the generator in the model is used to automatically extract multi-scale features and adaptively segment the choroid development area through an unsupervised learning method; the features extracted by the generator are input into the discriminator, and the generator is continuously optimized through adversarial learning so that it can accurately segment the choroid area; the adversarial training between the generator and the discriminator enables the generator to learn the features of removing choroid development, and the discriminator is responsible for evaluating the difference between the generated image and the real image; this adversarial mechanism enables the model to continuously optimize its segmentation effect during the training process, ensuring that abnormal development areas that cannot be accurately recognized can be automatically eliminated during the test stage, improving the accuracy of development removal, and ensuring that the model only effectively processes images with consistent modalities.

[0081] As Figure 4 shown, step S4) the model training step specifically includes step S41) dilated convolution step, step S42) weight update step, and step S43) development elimination step.

[0082] Step S41) Dilated convolution step: By setting the dilation factor to expand the receptive field of the convolution kernel, it can effectively capture the global features and local features in the initial image with complex geometric shapes while keeping the number of convolution kernel parameters unchanged. Its formula is

[0083]

[0084] where Y(t) represents a feature map with a larger receptive field, t represents the position of the image pixel, k represents the size of the convolution kernel, i represents the index of the convolution kernel, X represents the initial image, and M represents the convolution kernel.

[0085] In this embodiment, by introducing the dilation factor d, the convolution kernel samples at a larger interval when extracting features, thereby expanding the receptive field without increasing the computational complexity, especially suitable for processing structures with complex geometric shapes such as choroid development. Compared with traditional convolution operations, dilated convolution can better retain global context information and improve the recognition ability of complex image structures.

[0086] Step S42) Weight Update Step: Construct an adaptive attention weight distribution to optimize the feature extraction process, obtain a final feature map, enhance the features of the developing area and suppress the features of the background interference area through the final feature map, so as to obtain the developing features and the background features, enabling the elimination model to accurately segment and remove the developing area.

[0087] As Figure 5 shown, Step S42) Weight Update Step specifically includes Step S51) Feature Extraction Step, Step S52) First Weight Distribution Calculation Step, Step S53) Second Weight Distribution Calculation Step, and Step S54) Feature Weight Adjustment Step.

[0088] Step S51) Feature Extraction Step: Input the initial image into the elimination model to extract a multi-channel feature map F, where F ∈ R C×H×W , where C represents the number of feature channels, H represents the height of the feature map, W represents the width of the feature map, R is the set of real numbers, and each element in the feature map is a real number.

[0089] Step S52) First Weight Distribution Calculation Step: For the feature map F, calculate the first global average pooling and the first global max pooling on each channel respectively to obtain the global information describing the channel characteristics, and obtain the importance weight of the channel dimension. The formula is

[0090]

[0091]

[0092]

[0093] where, represents the first global average pooling, represents the first global max pooling, MLP represents the multi-layer perceptron, σ represents the sigmoid activation function, α c ∈[0,1], and α c represents the first weight of the channel dimension.

[0094] Step S53) Second Weight Distribution Calculation Step: Calculate the second global average pooling and the second global max pooling on the spatial dimension of the feature map F to generate a two-dimensional feature map and obtain the importance weight of the space. The formula is

[0095]

[0096]

[0097]

[0098] where, represents the second global average pooling, represents the second global max pooling, and CONV represents the convolution operation, represents the feature stacking operation, where β(i,j) ∈ [0,1], and β(i,j) represents the second weight at the spatial position (i,j).

[0099] Step S54) Feature weight adjustment step: Using the obtained first weight and second weight, adjust the channel dimension and spatial dimension of the feature map F to obtain the adjusted final feature map.

[0100] As Figure 6 shown, step S54) The feature weight adjustment step specifically includes step S61) Channel dimension adjustment step and step S62) Spatial dimension adjustment step.

[0101] Step S61) Channel dimension adjustment step: Based on the obtained first weight, adjust the channel dimension of the feature map F, and the formula is

[0102] F′ C = α C ·F C

[0103] where F C represents the feature map of the c-th channel, and F' C represents the adjusted feature map of the c-th channel;

[0104] Step S62) Spatial dimension adjustment step: Based on the obtained second weight, adjust the spatial dimension of the feature map F, and the formula is

[0105] F′ ij = β(i,j)·F ij

[0106] where F ij represents the value at the position (i,j) in the feature map F, and F' ij represents the adjusted value at the position (i,j) in the feature map F; After the feature map F performs the channel dimension adjustment step and the spatial dimension adjustment step, the final feature map is obtained.

[0107] In this embodiment, by constructing an attention module, the feature weights are dynamically adjusted in the channel dimension and the spatial dimension of the feature map respectively to obtain the final feature map, so as to enhance the saliency of the development region. In the channel dimension, the global feature information of each channel is extracted by using the global pooling operation to generate weight coefficients, and higher weight values are given to the feature channels related to the development region, and the channel features irrelevant to the segmentation are suppressed. In the spatial dimension, by calculating the spatial attention weights pixel by pixel, the response values of the pixels in the development region are enhanced, and the response values of the background or other interference regions are reduced at the same time, so as to improve the accuracy and robustness of the segmentation.

[0108] Step S43) Development elimination step: Calculate and obtain the third weight of the elimination model for specific voxels of the image, update the adversarial mechanism through the third weight to obtain the final adversarial mechanism, and based on the development features and the background features, use the final adversarial mechanism to perform choroid development elimination and repair processing on the initial image to obtain the final PET brain image with choroid development removed.

[0109] As Figure 7 shown, step S43) The development elimination step specifically includes step S55) Matrix calculation step, step S56) Score calculation step, step S57) Normalization step, and step S58) Weighted summation step.

[0110] Step S55) Matrix calculation step: Calculate and obtain the query matrix, key matrix, and value matrix. The formula is

[0111] Q = XW q ;

[0112] K = XW k ;

[0113] V = XW v ;

[0114] where Q represents the query matrix; K represents the key matrix; V represents the value matrix; W q represents the query weight matrix, which is used to project the input data matrix X into the query matrix Q; W k represents the key weight matrix, which is used to generate the key matrix K; W v represents the value weight matrix, which is used to generate the value matrix V, and X represents the data matrix.

[0115] In this embodiment, the input data matrix X usually represents the feature map after several previous convolutional or feature extraction operations in the attention mechanism, and the input data matrix X will be mapped into the query matrix Q.

[0116] Both are parameters that are continuously updated during the training process of the model. Through this linear transformation, the input data matrix X is projected into different feature spaces, providing a basis for subsequent attention calculation.

[0117] Step S56) Score calculation step, calculating the dot product score between the query matrix and the key matrix, and scaling the dot product score, with the formula being

[0118]

[0119] where scaled score(Q,K) represents the dot product score between the query matrix and the key matrix d, and d k represents the dimension of the query matrix and the key matrix.

[0120] Step S57) Normalization step, performing Softmax normalization on the obtained dot product score to obtain the attention weight, with the formula being

[0121]

[0122] where a ij represents the attention weight.

[0123] Step S58) Weighted summation step, based on the obtained attention weight, performing weighted summation on the value matrix to obtain the third weight, with the formula being

[0124]

[0125] where Attention(Q,K,V) represents the third weight, and the third weight is the final attention weight.

[0126] As Figure 8 shown, the present application also proposes a storage medium 100, storing computer-readable instructions, which, when read by at least one processor 110, cause at least one processor 120 to execute at least one step in the multi-device multi-channel power data synthesis method.

[0127] The present application provides a method for eliminating choroid visualization in PET brain images and a storage medium. This method is mainly applied to the problem of high-signal visualization of the choroid in positron emission tomography (PET). By collecting TSPO (18F-DPA714) and Tau protein (18F-Florzolotau) PET negative and positive visualization images, combining with clinical imaging data of neurodegenerative diseases such as Alzheimer's disease (AD) and progressive supranuclear palsy (PSP), using a GAN model to accurately distinguish the hypermetabolic signal of the choroid; adopting a generative adversarial network of deep learning to optimize the PET image processing flow and reduce the interference of the high-visualization signal of the choroid; by comparing the positive and negative visualization images, enhancing the visualization accuracy of the lesion tissue, the automatic elimination of the choroid signal in the PET image is realized, and the accuracy of image analysis is improved.

[0128] The above has provided a detailed introduction to a portable head-mounted display device of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A method for eliminating choroidal development in PET brain images, characterized in that: include: An image acquisition step of acquiring at least one set of PET brain images, wherein the PET brain images include a choroidal development area; A preprocessing step, converting the PET brain image from a DCM format to an NII format, and then preprocessing the PET brain image in the NII format to obtain an initial image set, wherein the initial image set includes at least one initial image; A model building step, constructing an elimination model, wherein the elimination model includes a generator and a discriminator, wherein the generator and the discriminator form an adversarial mechanism, inputting the initial image into the elimination model, extracting image features of the initial image by introducing dilated convolution and attention mechanisms, extracting development features and background features of each initial image, and performing choroidal development elimination and repair processing on the initial image based on the development features and the background features using the adversarial mechanism to obtain a final PET brain image with choroidal development removed; A model training step, inputting the initial image set into the elimination model, the generator is used to generate a reconstructed image with choroidal development removed, the discriminator compares the reconstructed image with the real image, and improves the image quality of the reconstructed image output by the generator through the adversarial mechanism. When the image quality meets the standard, the training is stopped to obtain a trained elimination model, wherein the real image is an image corresponding to the initial image without choroidal development; as well as The input step is to input the initial image into the trained elimination model to obtain the final PET brain image.

2. The method for eliminating choroidal development in PET brain images according to claim 1, characterized in that: The pre-processing step specifically comprises the following steps: an origin correction step, performing AC-PC correction on the brain part in the PET brain image, unifying the spatial origins of different PET brain images, setting the origin position of the PET brain image at AC, and obtaining a first image; a spatial standardization step, registering the first images to the standard brain template space of the Montreal Neurological Institute in sequence, unifying the coordinate space of all the first images, and obtaining a second image; a smoothing step of using a Gaussian filter to smooth the second image to obtain a third image; as well as The sliding window sampling step samples the third image with a sliding window of a preset size and a preset step length to obtain an initial image.

3. The method for eliminating choroidal development in PET brain images according to claim 1, characterized in that: The model training step specifically includes the following steps: The dilated convolution step expands the receptive field of the convolution kernel by setting the dilation factor. While keeping the convolution kernel parameters unchanged, it can effectively capture the global and local features of the initial image with complex geometric shapes. The formula is: Among them, Y(t) represents a feature map with a larger receptive field, t represents the position of the image pixel, k represents the size of the convolution kernel, i represents the index of the convolution kernel, X represents the initial image, and M represents the convolution kernel; A weight updating step, by constructing an adaptive attention weight distribution, optimizing the feature extraction process, obtaining a final feature map, enhancing the features of the development area through the final feature map, suppressing the features of the background interference area, and obtaining the development features and the background features; and The development elimination step calculates and obtains the third weight of the elimination model for the specific voxel of the image, updates the adversarial mechanism by the third weight, obtains the final adversarial mechanism, and performs choroidal development elimination and repair processing on the initial image based on the development characteristics and the background characteristics by using the final adversarial mechanism to obtain the final PET brain image with the choroidal development removed.

4. The method for eliminating choroidal development in PET brain images according to claim 3, characterized in that: The weight updating step specifically includes the following steps: The feature extraction step is to input the initial image into the elimination model to extract the multi-channel feature map F, where F∈R C ×H×W , where C represents the number of feature channels, H represents the height of the feature map, W represents the width of the feature map, and R is a real number set; In the first weight distribution calculation step, for the feature map F, the first global average pooling and the first global maximum pooling are calculated on each channel to obtain the global information describing the channel characteristics and obtain the importance weight of the channel dimension. The formula is: in, represents the first global average pooling, represents the first global maximum pooling, MLP represents multi-layer perceptron, σ represents the sigmoid activation function, α c ∈[0,1], and α c Represents the first weight of the channel dimension; In the second weight distribution calculation step, the second global average pooling and the second global maximum pooling are calculated on the spatial dimension of the feature map F to generate a two-dimensional feature map and obtain the importance weight of the space. The formula is: in, represents the second global average pooling, represents the second global maximum pooling, CONV represents a convolution operation, ⊕ represents a feature stacking operation, β(i,j)∈[0,1], and β(i,j) represents the second weight of the spatial position (i,j); and The feature weight adjustment step adjusts the channel dimension and the spatial dimension of the feature map F based on the obtained first weight and the second weight to obtain an adjusted final feature map.

5. The method for eliminating choroidal development in PET brain images according to claim 3, characterized in that: The development and elimination step specifically comprises the following steps: Matrix calculation steps, calculate and obtain the query matrix, key matrix and value matrix, the formula is: Q=XW q ; K=XW k ; V=XW v ; Where Q represents the query matrix; K represents the key matrix; V represents the value matrix; W q W represents the query weight matrix, which is used to project the input data matrix X into the query matrix Q; k represents the key weight matrix, which is used to generate the key matrix K; W v represents the value weight matrix, which is used to generate the value matrix V, and X represents the data matrix; The score calculation step calculates the dot product score of the query matrix and the key matrix, and scales the dot product score. The formula is: Wherein, scaled score (Q, K) represents the dot product score of the query matrix and the key matrix d, d k represents the dimensions of the query matrix and the key matrix; Normalization step: Softmax normalization is performed on the obtained dot product score to obtain the attention weight, which is formulated as follows: Among them, a ij represents the attention weight; and In the weighted summation step, based on the obtained attention weight, the value matrix is ​​weighted summed to obtain the third weight, which is formulated as follows: Among them, Attention(Q,K,V) represents the third weight.

6. The method for eliminating choroidal development in PET brain images according to claim 4, characterized in that: The feature weight adjustment step specifically includes the following steps: The channel dimension adjustment step adjusts the channel dimension of the feature map F based on the first weight obtained, and the formula is: F' C =a C ·F C Among them, F C represents the feature map of the cth channel, F' C represents the adjusted feature map of the cth channel; and The spatial dimension adjustment step is to adjust the spatial dimension of the feature map F based on the second weight obtained, and the formula is: F’ ij =β(i,j ) ·F ij Among them, F ij Represents the value of position (i, j) in feature map F, F' ij Represents the adjusted value of position (i, j) in feature map F; After the feature map F performs the channel dimension adjustment step and the space dimension adjustment step, the final feature map is obtained.

7. A storage medium storing computer-readable instructions, which, when read by at least one processor, enables at least one processor to execute at least one step in the multi-device multi-channel power data synthesis method according to any one of claims 1 to 6.

Images