A brown fat recognition method and recognition device based on plain scan CT images

Abstract

The application provides a brown fat recognition method and recognition device based on plain CT images, and solves the technical problem that the recognition cost and recognition accuracy of the existing brown fat recognition cannot be matched. The method comprises the following steps: establishing a tissue feature recognition model of brown fat by using the image features of the brown fat in the plain CT image resources; identifying the brown fat in the to-be-detected plain CT image according to the tissue feature recognition model, and quantifying the physical properties of the brown fat. The tissue feature recognition model is mainly trained or optimized by using low-cost plain CT image data, so that the brown fat and white fat can be effectively distinguished. The application provides a basic tool for target point recognition and tissue quantification, so as to explore how to activate the brown fat tissue in the human body through drugs, nutrition and other means, and to expect to find new strategies for treating obesity and diabetes.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and specifically to a method and device for identifying brown fat based on CT images. Background Technology

[0002] Brown adipose tissue (BAT) is a special type of adipose tissue found in mammals. This tissue is rich in mitochondria, which contain a special protein called UCP1, capable of generating heat by oxidizing fatty acids. Brown adipose tissue is considered a potential therapeutic target for obesity and diabetes. Basic research indicates that brown adipose tissue is indirectly related to serum cholesterol and triglyceride levels, potentially slowing the progression of atherosclerosis. Individuals with brown adipose tissue have a significantly lower probability of developing diabetes, hypertension, coronary heart disease, and cerebrovascular disease than those without, and brown adipose tissue is an independent protective factor against these common chronic metabolic diseases. Currently, methods for identifying the distribution and activity of brown adipose tissue mainly include color MRI (magnetic resonance imaging), PET-CT (positron emission tomography), scanning thermography, and blood biomarkers. However, color MRI and PET-CT are costly in terms of scanning and manpower, and other identification methods are less effective at distinguishing brown adipose tissue from other types of adipose tissue due to individual differences. Summary of the Invention

[0003] In view of the above problems, embodiments of the present invention provide a method and device for identifying brown fat based on plain CT images, which solves the technical problem that the cost and accuracy of existing brown fat identification methods cannot be matched.

[0004] The brown fat identification method based on plain CT images according to embodiments of the present invention includes:

[0005] A tissue feature recognition model for brown adipose tissue was established using the imaging features of brown adipose tissue in plain CT images.

[0006] Brown fat in plain CT images is identified using a tissue feature recognition model, and its physical properties are quantified.

[0007] In one embodiment of the present invention, the step of establishing a tissue feature recognition model for brown adipose tissue using the image features of brown adipose tissue in plain CT image resources includes:

[0008] Image feature data of brown fat was obtained using an image feature extraction tool. The image feature data includes feature points and feature descriptions.

[0009] Filter or reduce the dimensionality of image feature data to establish an image training dataset;

[0010] The first tissue feature recognition model is established by adjusting the classifier using the image training dataset.

[0011] In one embodiment of the present invention, the step of establishing a tissue feature recognition model for brown adipose tissue using the image features of brown adipose tissue in plain CT image resources includes:

[0012] Omics feature data of brown adipose tissue were obtained using radiomics feature extraction tools;

[0013] Filter or reduce the dimensionality of omics feature data to establish an omics training dataset;

[0014] A second tissue feature recognition model was established by adjusting the classifier using an omics training dataset.

[0015] In one embodiment of the present invention, the step of establishing a tissue feature recognition model for brown adipose tissue using the image features of brown adipose tissue in plain CT image resources includes:

[0016] The initial state of the tissue feature recognition model is formed by selecting the basic network, loss function and gradient backpropagation algorithm. The tissue feature recognition model is trained using plain CT image resources to determine the detection state of the tissue feature recognition model.

[0017] In one embodiment of the present invention, the image feature extraction tool includes at least one of the following:

[0018] Scale-invariant feature transformation, oriented gradient histogram, grayscale histogram, Gabor feature extraction, edge operators;

[0019] The method for filtering or dimensionality reduction of image feature data shall employ at least one of the following:

[0020] Principal component analysis, linear discriminant analysis, L1 regularization, L2 regularization, and adaptive enhancement.

[0021] In one embodiment of the present invention, the radiomics feature extraction tool includes at least one of the following:

[0022] Histogram features, texture features, spatial grayscale information of objects or shapes, frequency domain features without spatial positioning, frequency domain features with spatial positioning, frequency domain features of image intensity and texture features, coarse scanning features of grayscale variation regions, and geometric features of shapes.

[0023] The method for filtering or dimensionality reduction of image feature data shall employ at least one of the following:

[0024] Principal component analysis, linear discriminant analysis, L1 regularization, L2 regularization, and adaptive enhancement.

[0025] In one embodiment of the present invention, the classifier employs at least one of the following:

[0026] Support Vector Machine, Logistic Regression, Random Forest, Multilayer Perceptron.

[0027] In one embodiment of the present invention, the selected base network includes ResNet, DenseNet, VisionTransformer, Unet, or Faster RCNN; the selected loss function includes the cross-entropy loss function or the Dice loss function; and the selected gradient backpropagation algorithm includes SGD, Adam, or Rmsprop.

[0028] The brown fat recognition device based on plain CT images according to an embodiment of the present invention includes:

[0029] The memory is used to store the program code for the processing of the brown fat recognition method based on plain CT images described above.

[0030] A processor for executing the program code.

[0031] The brown fat recognition device based on plain CT images according to an embodiment of the present invention includes:

[0032] The model training module is used to establish a tissue feature recognition model for brown adipose tissue by utilizing the image features of brown adipose tissue in plain CT image resources.

[0033] The detection output quantification module is used to identify brown fat in the plain CT image to be detected based on the tissue feature recognition model, and to quantify the physical properties of brown fat.

[0034] The brown adipose tissue identification method and device based on plain CT images in this invention uses low-cost plain CT image data as the primary source to train or optimize the tissue feature recognition model, achieving effective differentiation between brown and white adipose tissue. This provides a fundamental tool for target identification and tissue quantification to explore how to activate brown adipose tissue in the human body through drugs, nutrition, and other means, with the aim of discovering new strategies for treating obesity and diabetes. Attached Figure Description

[0035] Figure 1 The diagram shown is a flowchart of a brown fat identification method based on plain CT images according to an embodiment of the present invention.

[0036] Figure 2 The diagram shown is a schematic representation of the architecture of a brown fat recognition device based on plain CT images according to an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0038] An embodiment of the present invention provides a method for identifying brown fat based on plain CT images, as follows: Figure 1 As shown. In Figure 1 In this invention, embodiments include:

[0039] Step 100: Establish a tissue feature recognition model for brown adipose tissue using the imaging features of brown adipose tissue in plain CT images.

[0040] Plain CT image resources mainly include CT images formed by ordinary intravenous scans without iodine-containing contrast agents. The imaging features of brown fat include, but are not limited to, region grayscale vectors, boundaries, local invariance, texture, medical omics features, and higher-order occult features. Using computer images to represent the image features of identical, similar, and analogous tissues as training or optimization input data, a general recognition or classification model is trained to form a tissue feature recognition model. This tissue feature recognition model is then used to identify and quantify brown fat objects in the test images. This forms the training or optimization input data for establishing a regularity recognition model of physical properties.

[0041] Step 200: Identify brown fat in the plain CT image to be detected based on the tissue feature recognition model, and quantify the physical properties of brown fat.

[0042] The physical properties of brown fat include, but are not limited to, entity identification, entity location, entity outline, entity size, and entity volume.

[0043] The brown adipose tissue identification method based on plain CT images in this invention uses low-cost plain CT image data as the primary source to train or optimize the tissue feature recognition model, achieving effective differentiation between brown and white adipose tissue. This provides a fundamental tool for target identification and tissue quantification to explore how to activate brown adipose tissue in the human body through drugs, nutrition, and other means, with the aim of discovering new strategies for treating obesity and diabetes.

[0044] like Figure 1 As shown, in one embodiment of the present invention, a tissue feature recognition model for brown adipose tissue is established using the image features of brown adipose tissue in plain CT image resources, including:

[0045] Step 110: Obtain image feature data of brown fat using an image feature extraction tool. The image feature data includes feature points and feature descriptions.

[0046] Computer image feature extraction tools include, but are not limited to, scale-invariant feature transform, histogram of oriented gradients, grayscale histograms, Gabor feature extraction, edge operators, and other image feature quantization or statistical methods. These tools can extract and effectively quantify image information to obtain local or overall image features of corresponding tissues or organs. Feature points characterize the correlation of local image information of tissues and organs, while feature descriptions characterize the specificity of local image information.

[0047] Step 120: Filter or reduce the dimensionality of the image feature data to establish an image training dataset.

[0048] Filtering or dimensionality reduction tools include, but are not limited to, principal component analysis (PCA), linear discriminant analysis (CDA), L1 and L2 regularization, and adaptive boosting. Supervised or unsupervised learning is used to filter image feature data, ensuring the diversity of training samples while retaining as much sample information as possible.

[0049] Step 130: Use the image training dataset to adjust the classifier and establish the first tissue feature recognition model.

[0050] Classifiers include, but are not limited to, Support Vector Machines (SVM), Logistic Regression, Random Forest, and Multilayer Perceptron. By using training data to train, adjust, or optimize the classifier, a set of parameters can be obtained to effectively quantify and distinguish the unique image features of brown fat.

[0051] The first tissue feature recognition model is constructed by classifying brown adipose tissue using computer image features. The classification result has the ability to recognize the image features of brown adipose tissue. Based on the characteristics of the classifier and the classification result, the model makes a judgment and quantification of brown fat in the plain CT image to be detected.

[0052] In one embodiment of the present invention, a classifier is used to form a pixel, region, and image object recognition process based on gray-level threshold segmentation: by setting a threshold, the gray-level values ​​of pixels in the CT image are compared with the threshold, and pixels with values ​​higher than the threshold are marked as brown adipose tissue. This method is simple and easy to implement, but its effect is not ideal when the gray-level value distribution in the image is uneven or there is noise.

[0053] In one embodiment of the present invention, a classifier is used to form a region and image object recognition process based on graphic morphology: the image is processed through operations such as dilation, erosion, and opening / closing to extract brown adipose tissue. For example, the opening operation is used to remove noise and small vascular structures from the image while retaining the brown adipose tissue.

[0054] In one embodiment of the present invention, a classifier is used to form a region and image object recognition process based on graphic texture features: texture feature extraction algorithms, such as gray-level co-occurrence matrix and wavelet transform, are used to process the classification results to extract the texture features of brown adipose tissue. For example, a gray-level co-occurrence matrix can be used to calculate the statistical features of the gray-level distribution between pixels in the image, which can be used to identify brown adipose tissue.

[0055] like Figure 1 As shown, in one embodiment of the present invention, a tissue feature recognition model for brown adipose tissue is established using the image features of brown adipose tissue in plain CT image resources, including:

[0056] Step 140: Obtain omics feature data of brown fat using radiomics feature extraction tools.

[0057] Plain CT image features include typical computer-generated image features. Plain CT images also include radiomics features specific to radiographic imaging. Radiomics features extracted from radiographic images using radiomics feature extraction tools include, but are not limited to:

[0058] Histogram features and texture features based on statistics;

[0059] Based on object or shape features represented by spatial grayscale information, such as object or shape features obtained using pixel autoregressive models;

[0060] Transformation-based frequency domain features without spatial localization, frequency domain features with spatial localization, frequency domain features of image intensity and texture features, and coarse scanning features of gray-level variation regions, etc.

[0061] Shape-based geometric features.

[0062] Step 150: Filter or reduce the dimensionality of the omics feature data to establish an omics training dataset.

[0063] Screening or dimensionality reduction tools include, but are not limited to, principal component analysis (PCA), linear discriminant analysis (CDA), L1 and L2 regularization, and adaptive boosting. Supervised or unsupervised learning is used to screen omics feature data, ensuring the diversity of training samples while preserving as much sample information as possible.

[0064] Step 160: Use the omics training dataset to adjust the classifier and build a second tissue feature recognition model.

[0065] Classifiers include, but are not limited to, Support Vector Machines (SVM), Logistic Regression, Random Forest, and Multilayer Perceptron. By utilizing training data to perform necessary training, adjustment, or optimization of the classifier, a set of parameters can be obtained to effectively quantify and distinguish the unique tissue features of brown adipose tissue.

[0066] The second tissue feature recognition model is constructed by classifying brown adipose tissue using radiomics features. The classification results have the ability to identify the radiometric features of brown adipose tissue. Based on the characteristics of the classifier and the classification results, the model makes a judgment and quantification of brown fat in the plain CT image to be detected.

[0067] In one embodiment of the present invention, a classifier is used to form morphological analysis, morphological labeling, and morphological segmentation based on morphology of the plain CT image to be detected, so as to extract the morphological features of brown adipose tissue, such as shape, size, and form.

[0068] In one embodiment of the present invention, a classifier is used to form a texture feature extraction algorithm based on texture features, such as gray-level co-occurrence matrix and wavelet transform, on the plain CT image to be detected. The image is then processed to extract the texture features of brown adipose tissue, such as texture complexity and texture uniformity.

[0069] In one embodiment of the present invention, a classifier is used to generate a machine learning-based pixel classification of the plain CT image to be detected in order to identify brown adipose tissue. For example, a random forest algorithm can be used, which combines morphological and texture features in the image to classify the pixels in the image.

[0070] like Figure 1 As shown, in one embodiment of the present invention, a tissue feature recognition model for brown adipose tissue is established using the image features of brown adipose tissue in plain CT image resources, including:

[0071] Step 170: Select the basic network, loss function and gradient backpropagation algorithm to form the initial state of the tissue feature recognition model, and use plain CT image resources to train the tissue feature recognition model to determine the detection state of the tissue feature recognition model.

[0072] Deep learning-based processing requires first selecting a foundational network, such as, but not limited to, ResNet, DenseNet, Vision Transformer, UNet, and Faster R-CNN. Then, a suitable loss function is chosen, such as cross-entropy loss or Dice loss. Finally, an appropriate gradient backpropagation method is selected, such as SGD, Adam, or RMSprop. Effective training using plain CT image resources can optimize the parameters of the tissue feature recognition model to meet the detection requirements of the application.

[0073] In one embodiment of the present invention, a tissue feature recognition model based on a convolutional neural network (CNN) is presented: CNNs can be used to perform end-to-end learning on plain CT images to extract features of brown adipose tissue. For example, classic CNN models such as ResNet and Inception, as well as some improved models specifically for brown adipose tissue, can be used.

[0074] In one embodiment of the present invention, an autoencoder-based tissue feature recognition model is proposed: an autoencoder can be used to compress and reconstruct images to extract useful features from the images. For example, a convolutional autoencoder (CAE) can be used to learn plain CT images to extract texture and morphological features from the images.

[0075] In one embodiment of the present invention, a tissue feature recognition model based on Generative Adversarial Networks (GANs) is proposed: GANs can be used to generate images of brown adipose tissue to improve the recognition accuracy of brown adipose tissue. For example, a conditional GAN ​​can be used, taking plain CT images as conditional inputs to generate corresponding brown adipose tissue images.

[0076] In one embodiment of the present invention, a tissue feature recognition model based on transfer learning can be used: a pre-trained deep learning model, such as a CNN model trained on a large-scale image dataset, can be used to perform transfer learning on plain CT images to extract useful features from the images.

[0077] In one embodiment of the present invention, a tissue feature recognition model based on a self-attention mechanism is proposed: the self-attention mechanism in Transformer can be used to process plain CT images to extract the texture features of brown adipose tissue. For example, self-attention mechanisms such as SE-Net and CBAM can be used to enhance useful features in the images.

[0078] In one embodiment of the present invention, a tissue feature recognition model based on multimodal fusion is proposed: the tissue feature recognition model can be fused with other image modal data, such as MRI, PET-CT, and other image resources, to improve the recognition accuracy of brown adipose tissue. For example, Transformer-based multimodal fusion methods, such as TRIM and ViT, can be used to jointly process multimodal images to extract the features of brown adipose tissue.

[0079] like Figure 1 As shown, in one embodiment of the present invention, a parallel recognition process for the plain CT image to be detected is formed using different tissue feature recognition models.

[0080] like Figure 1 As shown, in one embodiment of the present invention, a serial recognition process for the plain CT image to be detected is formed by using different tissue feature recognition models.

[0081] An embodiment of the present invention provides a brown fat recognition device based on plain CT images, comprising:

[0082] The memory is used to store the program code of the processing procedure of the brown fat recognition method based on plain CT images in the above embodiments;

[0083] A processor is used to execute program code for the processing of the brown fat identification method based on plain CT images described in the above embodiments.

[0084] The processor can be a DSP (Digital Signal Processor), an FPGA (Field-Programmable Gate Array), an MCU (Microcontroller Unit) system board, a SoC (System on a Chip) system board, or a PLC (Programmable Logic Controller) minimum system including I / O. Memory includes, but is not limited to, any type of disk (including floppy disks, hard disks, optical disks, CD-ROMs, and magneto-optical disks), ROM (Read-Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic cards, or optical cards. In other words, readable storage media includes any medium by which a device (e.g., a computer) stores or transmits information in a readable form.

[0085] An embodiment of the present invention provides a brown fat recognition device based on plain CT images, such as... Figure 2 As shown. In Figure 2 In this invention, embodiments include:

[0086] Model training module 10 is used to establish a tissue feature recognition model for brown fat by utilizing the image features of brown fat in plain CT image resources.

[0087] The detection output quantization module 20 is used to identify brown fat in the plain CT image to be detected based on the tissue feature recognition model and to quantify the physical properties of brown fat.

[0088] like Figure 2 As shown, in one embodiment of the present invention, the model training construction module 10 includes:

[0089] The image feature extraction unit 11 is used to obtain image feature data of brown fat through an image feature extraction tool. The image feature data includes feature points and feature descriptions.

[0090] Training data unit 12 is used to filter or reduce the dimensionality of image feature data and establish an image training data set.

[0091] The first model building unit 13 is used to adjust the classifier using the image training data set to establish a first tissue feature recognition model.

[0092] like Figure 2 As shown, in one embodiment of the present invention, the model training construction module 10 includes:

[0093] The omics feature extraction unit 14 is used to obtain omics feature data of brown fat through an image omics feature extraction tool;

[0094] Training data extraction unit 15 is used to filter or reduce the dimensionality of omics feature data and establish an omics training data set.

[0095] The second model component unit 16 is used to adjust the classifier using the omics training dataset to establish a second tissue feature recognition model.

[0096] like Figure 2 As shown, in one embodiment of the present invention, the model training construction module 10 includes:

[0097] The neural network construction unit 17 is used to select the basic network, loss function and gradient backpropagation algorithm to form the initial state of the tissue feature recognition model, and to train the tissue feature recognition model using plain CT image resources to determine the detection state of the tissue feature recognition model.

[0098] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying brown fat based on plain CT images, characterized in that, include: A tissue feature recognition model for brown adipose tissue was established using the imaging features of brown adipose tissue in plain CT images. Brown fat in plain CT images to be detected is identified based on tissue feature recognition models, and the physical properties of brown fat are quantified. The establishment of the tissue feature recognition model for brown adipose tissue includes: Image feature data of brown fat was obtained using an image feature extraction tool. The image feature data includes feature points and feature descriptions. Filter or reduce the dimensionality of image feature data to establish an image training dataset; A first tissue feature recognition model was established by adjusting the classifier using an image training dataset. Or include: Omics feature data of brown adipose tissue were obtained using radiomics feature extraction tools; Filter or reduce the dimensionality of omics feature data to establish an omics training dataset; A second tissue feature recognition model was established by adjusting the classifier using an omics training dataset.

2. The brown fat identification method based on plain CT images as described in claim 1, characterized in that, The method for establishing a tissue feature recognition model for brown adipose tissue using the image features of brown adipose tissue in plain CT image resources includes: The initial state of the tissue feature recognition model is formed by selecting the basic network, loss function and gradient backpropagation algorithm. The tissue feature recognition model is trained using plain CT image resources to determine the detection state of the tissue feature recognition model.

3. The brown fat identification method based on plain CT images as described in claim 1, characterized in that, The image feature extraction tool includes at least one of the following: Scale-invariant feature transformation, oriented gradient histogram, grayscale histogram, Gabor feature extraction, edge operators; The method for filtering or dimensionality reduction of image feature data shall employ at least one of the following: Principal component analysis, linear discriminant analysis, L1 regularization, L2 regularization, and adaptive enhancement.

4. The brown fat identification method based on plain CT images as described in claim 1, characterized in that, The radiomics feature extraction tool includes at least one of the following: Histogram features, texture features, spatial grayscale information of objects or shapes, frequency domain features without spatial positioning, frequency domain features with spatial positioning, frequency domain features of image intensity and texture features, coarse scanning features of grayscale variation regions, and geometric features of shapes; The method for filtering or dimensionality reduction of image feature data shall employ at least one of the following: Principal component analysis, linear discriminant analysis, L1 regularization, L2 regularization, and adaptive enhancement.

5. The brown fat identification method based on plain CT images as described in claim 1, characterized in that, The classifier shall employ at least one of the following: Support Vector Machine, Logistic Regression, Random Forest, Multilayer Perceptron.

6. The brown fat identification method based on plain CT images as described in claim 2, characterized in that, The selected base network includes ResNet, DenseNet, VisionTransformer, Unet, or Faster RCNN; the selected loss function includes the cross-entropy loss function or the Dice loss function; and the selected gradient backpropagation algorithm includes SGD, Adam, or Rmsprop.

7. A brown fat recognition device based on plain CT images, characterized in that, include: The memory is used to store the program code of the processing procedure of the brown fat identification method based on plain CT images as described in any one of claims 1 to 6; A processor for executing the program code.

8. A brown fat recognition device based on plain CT images, characterized in that, include: The model training module is used to establish a tissue feature recognition model for brown adipose tissue by utilizing the image features of brown adipose tissue in plain CT image resources. The detection output quantification module is used to identify brown fat in the plain CT image to be detected based on the tissue feature recognition model and quantify the physical properties of brown fat. The model training construction module includes: The image feature extraction unit is used to obtain image feature data of brown fat through an image feature extraction tool. The image feature data includes feature points and feature descriptions. Training data units are used to filter or reduce the dimensionality of image feature data and establish an image training data set. The first model building unit is used to adjust the classifier using the image training dataset to establish the first tissue feature recognition model; Or include: The omics feature extraction unit is used to obtain omics feature data of brown fat through image omics feature extraction tools; The training data extraction unit is used to filter or reduce the dimensionality of omics feature data and establish an omics training data set. The second model building unit is used to adjust the classifier using the omics training dataset to establish a second tissue feature recognition model.

Images