Body fat rate estimation method based on whole body CT scanning image and neural network model
By using whole-body CT scan images and U-Net neural network model, the problem of high cost and insufficient accuracy of body fat rate measurement in the prior art is solved, and accurate and low-cost different types of body fat rate measurement are achieved, providing clinical Valid reference.
Patent Information
- Application Number
- CN202410898054.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-05
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art cannot accurately measure different types of body fat rates at a lower cost, especially visceral body fat rates, subcutaneous body fat rates and interstitial body fat rates, and the existing methods have high costs or insufficient accuracy.
Using a method based on full-body CT scan image and U-Net neural network model, subcutaneous fat, visceral fat and intermuscular fat tissue were divided into subcutaneous fat, visceral fat and intermuscular fat tissue, thereby calculating the proportion of various fats and Overall fat rate.
On the premise of ensuring measurement accuracy, different types of body fat content measurements at a lower cost are achieved, providing reference for clinical treatment and prescription.
Smart Images

Figure CN118902389A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of body fat percentage measurement methods, and in particular to a body fat percentage estimation method based on whole-body CT scan images and a neural network model. Background Art
[0002] Body fat percentage can be divided into visceral body fat percentage, subcutaneous body fat percentage and intermuscular body fat percentage. They are important indicators of health both in clinical practice and in daily life. At present, people usually use bioelectrical impedance analysis, dual-energy x-ray absorptiometry and magnetic resonance imaging for measurement. All three methods have certain problems.
[0003] First of all, the principles of bioelectrical impedance analysis and dual-energy x-ray absorptiometry are similar. The former measures fat content based on the difference in electrical resistance between fat and other tissues, while the latter measures fat content based on the difference in absorption rates of two types of x-rays between fat tissue and other tissues. Although these two methods can accurately give the total body fat percentage, many interfering factors such as water content and individual differences will change the resistance value and radiation absorption rate of different types of fat tissue. Therefore, these two methods cannot accurately measure the respective proportions of visceral fat, subcutaneous fat, and intermuscular fat. Secondly, although magnetic resonance imaging combined with certain algorithms can obtain the content of various types of fat, its measurement cost is very high. Statistics show that under the same conditions, the price of an MRI scan is several times that of a CT scan.
[0004] Studies have shown that obese people have a greatly increased risk of chronic diseases such as diabetes, hypertension, fatty liver, and atherosclerosis. Even if the body fat percentage is the same, the location of fat accumulation will have different effects on people's health. For example, people with fat accumulation in the internal organs are more likely to suffer from the above-mentioned related diseases than those with fat accumulation under the skin. After the Second World Health Revolution, prevention and active intervention became the mainstream of medicine, and monitoring of physical condition is a prerequisite for preventive medicine. Therefore, the measurement of total fat percentage and different types of body fat percentage has become an increasingly important demand for people and has a good market prospect. Summary of the invention
[0005] The present invention provides a body fat percentage estimation method based on whole-body CT scan images and a neural network model to solve the problem that the existing human body fat percentage measurement method cannot measure different types of fat content at a low cost.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is:
[0007] A body fat percentage estimation method based on whole-body CT scan images and neural network models uses a U-Net neural network model to process whole-body CT scan images and estimates the whole-body subcutaneous body fat percentage, visceral body fat percentage, intermuscular body fat percentage, and total body fat percentage by category. The process is as follows:
[0008] Step 1, obtaining a data set, wherein the data set is composed of continuous cross-sectional images obtained by CT scanning of the human body, and each image in the continuous cross-sectional images has segmented visceral fat tissue, subcutaneous fat tissue and intermuscular fat tissue, and they are made into corresponding labels; then all image data in the obtained data set are divided into a training set and a test set;
[0009] Step 2: Build a U-Net neural network model, which is used to process four-classification tasks;
[0010] Step 3, training and testing the U-Net neural network model, using the training set obtained in step 1 to train the U-Net neural network model constructed in step 2, so that the U-Net network model has the ability to segment subcutaneous fat tissue, visceral fat tissue and intermuscular fat tissue from CT scan images that have never been seen; and using the test set obtained in step 1 to test the trained U-Net network model, if the trained U-Net network model meets the requirements after testing, then use the network as the fat tissue prediction model;
[0011] Step 4, obtaining and processing the scanned image of the subject, the CT instrument divides the human body to be measured into a number of continuous sampling cross sections at a certain interval, thereby generating continuous cross-sectional images of the whole body of the human body to be measured obtained by CT scanning, and inputting them into the fat tissue prediction model obtained in step 3, and the model is used to segment subcutaneous fat tissue, visceral fat tissue and intermuscular fat tissue from each cross-sectional image in the above continuous cross-sectional images;
[0012] Step 5, calculate the body fat percentage, and count the proportion of subcutaneous fat pixels, visceral fat pixels and intermuscular fat pixels to the total number of pixels in the segmentation results of each cross-sectional image, so as to obtain the proportion of the three types of fat tissue areas in each sampling cross-section corresponding to each cross-sectional image, and calculate the visceral body fat percentage, subcutaneous body fat percentage, intermuscular body fat percentage and total body fat percentage of the human body to be measured based on these ratios.
[0013] Furthermore, in step 1, the dataset is divided into a training set and a test set in a ratio of 8:2.
[0014] Furthermore, in step 2, a U-Net neural network model is used as a tool for processing continuous cross-sectional images, and the output layer of the model should have four channels, each channel corresponding to a category, namely background, subcutaneous fat tissue, visceral fat tissue, and intermuscular fat tissue.
[0015] Furthermore, in step 3, when the U-Net neural network model is trained, the loss function is a pixel-level cross entropy loss function, and the optimizer is Adam.
[0016] Furthermore, in step 5, the area ratio of the three types of fat tissues in each sampling cross section is multiplied by the sampling cross section area to obtain the area of the three types of fat tissues in each sampling cross section. Then, the area of the three types of fat tissues in each sampling cross section is multiplied by the sampling interval to obtain the scan volume represented by the three types of fat tissues in each sampling cross section. Next, the scan volumes represented by the three types of fat tissues in all sampling cross sections are accumulated to obtain the total volume of the three types of fat in the human body to be measured, and then combined with the body fat density to obtain the total mass of the three types of fat in the human body to be measured, and then the total mass of the three types of fat in the human body to be measured is divided by the body weight of the human body to be measured to obtain the subcutaneous body fat rate, visceral body fat rate and intermuscular body fat rate of the human body to be measured. Finally, the three are added together to obtain the total body fat rate of the human body to be measured.
[0017] Compared with the prior art, the present invention has the following advantages:
[0018] The present invention can calculate the content of different types of body fat at a relatively low cost while ensuring measurement accuracy, thereby providing a reference for clinical treatment and prescription. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a flow chart of body fat percentage estimation of the present invention.
[0020] Figure 2 It is a schematic diagram of the U-Net neural network used in the present invention.
[0021] Figure 3 It is a flow chart of U-Net neural network training in the present invention.
[0022] Figure 4 It is a schematic diagram of the principle of generating continuous cross-sectional images in the present invention. DETAILED DESCRIPTION
[0023] The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0024] like Figure 1As shown, this embodiment discloses a body fat percentage estimation method based on a whole-body CT scan image and a neural network model, which is implemented based on an existing U-Net neural network model and a whole-body CT scan image of a subject, and includes the following steps:
[0025] Step 1: Get the dataset.
[0026] In this embodiment, the data set is composed of continuous cross-sectional images obtained by CT scanning of the whole body of the human body. Since the whole body is scanned, the continuous cross-sectional images contain cross-sectional images of multiple different parts of the human body. Next, these image data are preprocessed.
[0027] The first step of preprocessing is to normalize the dataset and resize it to 512×512 pixels.
[0028] After size normalization, the visceral adipose tissue, subcutaneous adipose tissue, and intermuscular adipose tissue in each cross-sectional image were manually segmented and made into label maps corresponding to the cross-sectional images.
[0029] Finally, 80% of the cross-sectional images and their corresponding label maps in the dataset are randomly selected as the training set, and the remaining 20% of the cross-sectional images and their corresponding label maps are used as the test set.
[0030] Step 2: Build the U-Net network model.
[0031] The U-Net neural network model in this embodiment is a fully convolutional neural network widely used in the field of image processing, and it plays an important role in the field of medical image segmentation.
[0032] like Figure 2 As shown, the U-Net network model used in this embodiment mainly includes four key operations: downsampling, upsampling, skip connection and output. The downsampling part contains four layers. Starting from the upper left corner, each layer uses two 3×3 convolution kernels to perform two convolution operations in succession, and each convolution operation is followed by a ReLU activation function. Every two layers in the downsampling part are connected by a maximum pooling operation, and the pooling kernel size used is 2×2 with a step size of 2. Through the above operations, the model realizes the feature extraction of the input CT cross-sectional image.
[0033] The upsampling part also includes four layers, which correspond to the downsampling part one by one. Starting from the lower right corner, the feature matrix obtained by downsampling is deconvolved, and the deconvolution kernel used is 2×2, and the method is bilinear interpolation. After the deconvolution is completed, it enters the next layer. First, the image obtained by the corresponding layer on the left during downsampling is cropped to the same size as the right through the skip layer connection and then spliced. Next, two convolution operations are performed successively through two 3×3 convolution kernels. Each convolution operation is followed by a ReLU activation function. The above process is repeated to complete the upsampling operation.
[0034] After the above steps, the output operation is performed, and a 1×1×4 convolution kernel is selected to process the final feature matrix, so that the features are mapped to the category space, and the image segmentation results including four categories of background, visceral fat, subcutaneous fat and intramuscular fat are obtained.
[0035] Therefore, in this embodiment, the U-Net network model is used to segment the three types of fat tissue from the continuous cross-sectional images to be measured obtained by CT scanning of the whole body of the human body to be measured. Specifically, in this embodiment, the U-Net network model is generated using Python language or other computer languages.
[0036] Step 3: Train and test the U-Net neural network model.
[0037] After the U-Net network model is constructed in this embodiment, the U-Net network model is trained using the training set obtained in step 1. Figure 3 As shown, the training and testing process is as follows:
[0038] After starting training, the data in the training set is sent to the U-Net network model in batches. The model outputs the predicted value through forward propagation, and then the difference between the model prediction map and the training set label map, that is, the loss value, is calculated according to the loss function. In this embodiment, a pixel-level cross entropy loss function is used, and its principle is: each pixel in the training set label map can be represented as a category distribution vector, each element of which represents a category. For example, in this embodiment, the image has four categories: background, subcutaneous fat, visceral fat, and intermuscular fat. Then the category distribution vector of the background pixel point is (1,0,0,0), and the category distribution vector of the subcutaneous fat pixel point is (0,1,0,0), and so on. Similarly, each pixel in the prediction result of the model is also represented as a probability distribution vector, in which each element represents the probability of belonging to the corresponding category. For example, if the probability distribution vector of a pixel point in the predicted image is (0.1,0.1,0.7,0.1), it means that the model predicts that the pixel point belongs to visceral fat. The pixel-level cross entropy loss function measures the difference between the model prediction map and the training set label map by comparing these two distribution vectors. The calculation formula is as follows:
[0039]
[0040] Among them, y i and represents the pixels in the training set label image and the model prediction image respectively, N represents the total number of pixels in the prediction image or the test set label image, in this embodiment N = 512 × 512; K represents the number of categories, in this embodiment K = 3, which are visceral fat, subcutaneous fat and intermuscular fat respectively; y ij Indicates whether the i-th pixel in the label map of the training set belongs to the j-th category. If so, y ij =1, if it does not belong to y ij =0; Represents the probability that the i-th pixel in the model's prediction graph belongs to the j-th category.
[0041] After obtaining the loss value, the U-Net network model performs back propagation, and calculates the gradient of the loss value to each parameter in the network (such as the value in the convolution kernel and the deconvolution kernel) layer by layer according to the chain rule. The gradient gives the direction of parameter optimization, and then the parameters are updated through the optimizer algorithm so that the model can gradually optimize the value of the loss function, thereby reducing the difference between the image predicted by the model and the labeled image of the training set. In this embodiment, the Adam optimizer algorithm is used, and the following is the parameter update formula of the optimizer algorithm:
[0042]
[0043] Among them, θ t is a parameter vector in the model at the tth iteration; α is the learning rate, which is used to control the step size of parameter update; is the loss function J for the parameter θ t The gradient of the gradient; β1 and β2 are the exponential decay rates used to control the momentum and adaptive learning rate; m t and v t are the first-order moment estimate and the second-order moment estimate of the parameter respectively; m' t+1 and v' t+1 They are the first-order moment estimate and the second-order moment estimate after correcting the deviation; ∈ is a very small number used to prevent division by zero errors.
[0044] Through multiple training iterations, the difference between the training set label map and the U-Net network model prediction map is continuously reduced, so that the prediction value generated by the U-Net network model is closer to the label value to achieve the purpose of learning. Finally, when the training accuracy reaches the requirement, the adjustment is stopped, the U-Net network model training is completed, and a trained U-Net network model is obtained.
[0045] After obtaining the trained U-Net network model in this embodiment, the generalization ability of the trained U-Net network model to segment the three types of fat tissues is tested using the test set obtained in step 1. The cross-sectional images in the test set are sent as input to the model for segmentation. After processing, the predicted value is obtained and compared with the label image in the test set. If the error between the predicted value and the label image of the test set is found to be within an acceptable range after comparison, it means that the trained U-Net network model has the ability to segment the three types of fat tissues.
[0046] Therefore, in this embodiment, after testing with the test set, if the trained U-Net network model meets the requirements after testing, the corresponding U-Net network model is used as the fat tissue prediction model.
[0047] In accordance with the above principles, this embodiment uses computer languages such as Python to implement the training and testing of the above U-Net network model.
[0048] Step 4: Acquire and process the continuous cross-sectional images to be measured.
[0049] like Figure 4 As shown, the continuous cross-sectional images to be measured are obtained by CT scanning the whole body of the human body to be measured. During the scanning, the CT instrument divides the human body to be measured from head to toe into a number of continuous sampling cross sections at certain intervals, thereby generating continuous cross-sectional images to be measured corresponding to the continuous sampling cross sections one by one.
[0050] The continuous cross-sectional images to be measured are input into the fat tissue prediction model obtained in step 2, and the fat tissue prediction model is used to segment three types of fat tissue from each of the continuous cross-sectional images to be measured.
[0051] Step 5: Calculate the body fat percentage of the person to be measured.
[0052] In this embodiment, the final output of the model is a continuous prediction image that corresponds one-to-one to the continuous cross-sectional images of the input model, and the size and number of pixels of the continuous prediction image are exactly the same as those of the continuous cross-sectional images. In each model prediction image, the background is black pixels, the subcutaneous fat tissue is red pixels, the visceral fat tissue is blue pixels, and the intermuscular fat tissue is yellow pixels.
[0053] Calculate the ratio of the red, blue and yellow pixels to the total number of pixels in each model prediction image. This ratio is the proportion of the three types of fat tissue in each cross-sectional image, and is also the area ratio of the three types of fat tissue in each sampling cross section. Next, multiply the area ratio of the three types of fat tissue in each sampling cross section by the area of the sampling cross section to obtain the area of the three types of fat tissue in each sampling cross section of the human body to be measured from head to toe. The area of the sampling cross section can be obtained by checking the FOV parameters of the CT instrument.
[0054] Since the sampling interval is fixed and the value is very small when the CT scan is performed on the human body to be measured in this embodiment, each sampling cross section of the human body to be measured from head to toe can be equivalent to a part of the scanning volume, and the scanning volume equivalent to each sampling cross section of the human body to be measured is equal to the area of the sampling cross section multiplied by the sampling interval. Similarly, the scanning volume represented by the three types of fat tissue in each sampling cross section is also equal to the area of the three types of fat tissue in each sampling cross section multiplied by the sampling interval, and the calculation formula is as follows:
[0055]
[0056] S ij =S×r ij (8)
[0057] V ij =S ij ×d (9)
[0058] In formula 7, r ij represents the proportion of the jth type of fat tissue in the i-th cross-sectional image to be measured; n ij It represents the number of pixels representing the jth type of fat tissue in the ith prediction image. In the prediction image of this embodiment, there are three types of fat pixels: red pixels are subcutaneous fat tissue, blue pixels are visceral fat tissue, and yellow pixels are intermuscular fat tissue; N represents the total number of pixels in the model prediction image or the cross-sectional image to be measured. In this embodiment, N=512×512.
[0059] In formula 8, S ij It represents the area occupied by the jth type of fat tissue in the ith sampling cross section; S represents the area of the sampling cross section when the CT apparatus performs continuous cross-sectional scanning on the human body to be measured.
[0060] In formula 9, V ij represents the scan volume represented by the j-th type of adipose tissue in the ith sampling cross section; d represents the sampling interval.
[0061] Since the sampling during CT scanning imaging is continuous, the total scanning volume can be obtained by accumulating all scanning volumes within the sampling range. Similarly, the total volume of the three types of fat tissues in each equivalent scanning volume can be obtained by accumulating the volumes of the three types of fat tissues in each equivalent scanning volume.
[0062] Therefore, in this embodiment, the volume V of the three types of fat tissue in all sampling cross sections obtained in step 5 is ij The total volumes of the three types of fat tissues in the human body to be measured are accumulated respectively, and the calculation formula is as follows:
[0063]
[0064] V=V1+V2+V3 (13)
[0065] Wherein, V1, V2, and V3 represent the total volume of visceral fat tissue, subcutaneous fat tissue, and intermuscular fat tissue in the human body to be measured, respectively; m represents the total number of scanned volumes, that is, the total number of continuous cross-sectional images or model prediction images to be measured; V i1 、V i2 、V i3 They respectively represent the volumes of visceral adipose tissue, subcutaneous adipose tissue and intermuscular adipose tissue in the i-th scanning volume.
[0066] Since the fat tissue density of the human body is ρ = 0.74g / ml, that is, 740kg / m 3 Therefore, after obtaining the total volume of the three types of fat in the human body to be measured, the total mass of the three types of fat in the human body to be measured can be further calculated in this embodiment, and the formula is as follows:
[0067] M1=ρ×V1 (14)
[0068] M2=ρ×V2 (15)
[0069] M3=ρ×V3 (16)
[0070] Among them, M1, M2, and M3 are the total mass of the visceral fat tissue, subcutaneous fat tissue, and intermuscular fat tissue in the human body to be measured, respectively.
[0071] The weight M of the person to be measured is a known quantity, so the visceral body fat rate A1=M1 / M of the person to be measured that is finally calculated in this embodiment; the subcutaneous body fat rate A2=M2 / M of the person to be measured; and the intermuscular body fat rate A3=M3 / M of the person to be measured.
[0072] In addition, in this embodiment, the total body fat percentage A of the human body to be measured can be obtained by A1+A2+A3.
[0073] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.
[0074] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.
Claims
1. Body fat percentage estimation method based on whole-body CT scan images and neural network model, It is characterized in that The U-Net neural network model is used to process the whole-body CT scan images and estimate the subcutaneous body fat percentage, visceral body fat percentage, intermuscular body fat percentage, and total body fat percentage by category. The process is as follows: Step 1, obtaining a data set, wherein the data set is composed of continuous cross-sectional images obtained by CT scanning of the human body, and each image in the continuous cross-sectional images has segmented visceral fat tissue, subcutaneous fat tissue and intermuscular fat tissue, and they are made into corresponding labels, and then all image data in the obtained data set are divided into a training set and a test set; Step 2: construct a U-Net neural network model, where the U-Net neural network model is used to process four-category image segmentation tasks; Step 3, training and testing the U-Net neural network model, using the training set obtained in step 1 to train the U-Net neural network model constructed in step 2, so that the U-Net network model has the ability to segment subcutaneous fat tissue, visceral fat tissue and intermuscular fat tissue from cross-sectional images that have never been seen; and using the test set obtained in step 1 to test the trained U-Net network model, if the trained U-Net network model meets the requirements after testing, then use the network as the fat tissue prediction model; Step 4, obtaining and processing the scanned image of the subject, the CT instrument divides the human body to be measured into a number of continuous sampling cross sections at a certain interval, thereby generating continuous cross-sectional images of the whole body of the human body to be measured obtained by CT scanning, and inputting them into the fat tissue prediction model obtained in step 3, and the model is used to segment subcutaneous fat tissue, visceral fat tissue and intermuscular fat tissue from each cross-sectional image in the above continuous cross-sectional images; Step 5, calculate the body fat percentage, and count the proportion of subcutaneous fat pixels, visceral fat pixels and intermuscular fat pixels to the total number of pixels in the segmentation results of each cross-sectional image, so as to obtain the proportion of the three types of fat tissue areas in each sampling cross-section corresponding to each cross-sectional image, and calculate the visceral body fat percentage, subcutaneous body fat percentage, intermuscular body fat percentage and total body fat percentage of the human body to be measured based on these ratios.
2. The method for estimating body fat percentage based on whole-body CT scan images and a neural network model according to claim 1, It is characterized in that In step 1, the dataset is divided into training set and test set in a ratio of 8:
2.
3. The method for estimating body fat percentage based on whole-body CT scan images and a neural network model according to claim 1, It is characterized in that In step 2, a U-Net neural network model is used as a tool for processing continuous cross-sectional images. The output layer of the model should have four channels, each channel corresponding to a category, namely background, subcutaneous fat tissue, visceral fat tissue, and intermuscular fat tissue.
4. The method for estimating body fat percentage based on whole-body CT scan images and a neural network model according to claim 1, It is characterized in that In step 3, when training the U-Net neural network model, the loss function selects the pixel-level cross entropy loss function, and the optimizer selects Adam.
5. The method for estimating body fat percentage based on whole-body CT scan images and a neural network model according to claim 1, It is characterized in that In step 5, the area ratio of the three types of fat tissues in each sampling cross section is multiplied by the sampling cross section area to obtain the area of the three types of fat tissues in each sampling cross section. Then, the area of the three types of fat tissues in each sampling cross section is multiplied by the sampling interval to obtain the scan volume represented by the three types of fat tissues in each sampling cross section. Next, the scan volumes represented by the three types of fat tissues in all sampling cross sections are accumulated to obtain the total volume of the three types of fat in the human body to be measured, and then combined with the body fat density to obtain the total mass of the three types of fat in the human body to be measured. Then, the total mass of the three types of fat in the human body to be measured is divided by the weight of the human body to be measured to obtain the subcutaneous body fat rate, visceral body fat rate and intermuscular body fat rate of the human body to be measured. Finally, the three are added together to obtain the total body fat rate of the human body to be measured.
Citation Information
Patent Citations
System and method for performing automatic quantitative evaluation on whole-body fat by using magnetic resonance (MR) technology
CN112168168A
Abdominal fat segmentation method and device and computer equipment
CN117422874A
Inspected object component analysis method and device, label generation method and device and model training method and device
CN117522857A
Three-dimensional full-automatic human body component analysis method and system based on CT image and medium
CN118052783A
System and method for image-based quantification of white and brown adipose tissue at the whole-body, organ and body-region levels
US20180165808A1
