Automatic body composition measurement system based on abdominal CT images and deep learning

Abstract

The present invention discloses an automatic body composition measurement system based on abdominal CT images and deep learning. The system of the present invention includes a positioning module for all axial slices corresponding to the third lumbar vertebra on the CT image, a four-category skeletal muscle segmentation module corresponding to the axial slices of the third lumbar vertebra on the CT image, and an automatic body composition measurement module based on the CT image. The deep learning model is trained using the first two modules to locate and segment the four types of skeletal muscles of the third lumbar vertebra, and automatically calculate the proportion of body components such as fat and muscle. The present invention is trained and verified on CT data of clinical patients with liver cirrhosis, and the average Dice of the four types of skeletal muscle segmentation results reaches 0.9283, and the average surface distance is 0.6779mm. The present invention can also obtain body components such as subcutaneous fat and intra-abdominal fat by threshold processing. The present invention uses 2-3 seconds to batch obtain the body components corresponding to the third lumbar vertebra, which is of great significance for the clinical diagnosis of complications in patients with liver cirrhosis.

Description

Technical Field

[0001] The invention belongs to the technical field of medical diagnostic equipment, and in particular relates to an automatic body composition measurement system based on abdominal CT images. Background Art

[0002] Body composition analysis is the quantification of muscle and fat tissue. Its measurement is usually based on two methods: biological parameters or medical imaging. In particular, measuring the mass of skeletal muscle on an axial CT slice through the center of the third lumbar vertebra (L3) has been shown to be the most accurate [1]. Measuring skeletal muscle in an L3 (or L4) axial CT slice mainly involves segmenting muscle and fat based on the HU value of the CT image. There are two ways: one method segments all muscle and fat, such as the HU value ranges of muscle, subcutaneous and intermuscular fat, and intra-abdominal fat are [-29, 150], [-190, -30], and [-150, -50], respectively [2]. The other method only measures the psoas muscle [3]. The shortcomings of these two methods are: it is difficult for these methods to distinguish tissues in certain special pathological conditions, resulting in inaccurate segmentation results; whether all skeletal muscles are analyzed as a whole or only the psoas muscle is analyzed, the CT image is not fully utilized; more importantly, using only skeletal muscle on a single axial CT slice of L3 or L4, the sample size is small and the muscle mass cannot be accurately described.

[0003] (1) Traditional segmentation method

[0004] Medical image segmentation is a key factor in determining whether medical images can provide reliable evidence in clinical diagnosis and treatment. Currently, several semi-automatic skeletal muscle area measurement software, such as ImageJ, CoreSlicer, Slice-O-matic, 3D Slicer, Horos, OsiriX, Mimics and FatSeg, are widely used for image segmentation[4][5].

[0005] In addition to software segmentation, machine learning-based methods are also widely used for muscle segmentation. For example, researchers use shape prior modeling methods to segment muscles in L3 corresponding slices or use new finite element deformation models to segment muscles [6]. However, it is challenging to segment different types of skeletal muscles on L3 axial CT slices using traditional methods because: 1) The shapes and sizes of different types of skeletal muscles vary greatly, and there is a class imbalance problem during optimization, such as Figure 3 Four skeletal muscle labels are shown; 2) The boundaries between different skeletal muscles and between skeletal muscles and surrounding tissues are unclear and have rough edges. Existing software segmentation methods cannot actively learn this useful information, so the accuracy is not high, such as Figure 3 3) The same type of skeletal muscle has morphological differences between different individuals, which affects the segmentation results.

[0006] (2) Deep Learning Segmentation Method

[0007] Deep learning can automatically and batch-segment related organs, tissues, and tumors in patient images in various existing disease databases, so it can bring low muscle quantity and quality from theoretical research to clinical practice. Related studies on deep learning in abdominal CT images include: segmenting skeletal muscle in axial CT images of the third or fourth lumbar vertebra for body composition analysis [7]; automatically locating the axial slice of the L3 center from the whole body or partial body CT image, and then segmenting skeletal muscle and other body components [8]; using the axial CT slice corresponding to the lower end plate of the third lumbar vertebra for skeletal muscle segmentation analysis [9]. However, all of the above methods segment the entire skeletal muscle on a single slice to consider the impact of physiology and pathology on the entire skeletal muscle, but do not consider the impact of physiology and pathology on different types of muscle.

[0008] Existing studies have also fully automatically segmented multiple muscle groups in the third and fourth lumbar vertebrae for the detection of central sarcopenia

[10] ; there are also studies that automatically segment paraspinal muscles on axial magnetic resonance imaging of different vertebrae

[11] ; researchers have also automatically segmented the left paraspinal muscles in the axial position of the twelfth thoracic vertebra

[12] . Although these methods have focused on different types of muscle groups, they are still limited to a single slice, without analyzing the overall volume, and do not cover all muscle groups. The latest paper has demonstrated by segmenting the entire abdominal and pelvic skeletal muscle that the average difference in skeletal muscle volume is significantly lower than the average difference in the corresponding area of ​​a single CT slice

[13] .

[0009] Malnutrition in patients with cirrhosis often leads to sarcopenia. However, currently, the diagnosis of sarcopenia depends on the measurement of skeletal muscle mass and quantity on a single axial slice through the third lumbar vertebra (L3) on abdominal or abdominopelvic CT images. This measurement lacks an analysis of the association between different types of skeletal muscle and sarcopenia, and only uses skeletal muscle on a single axial CT slice, resulting in a small sample size.

[0010] Based on this, the present invention divides the skeletal muscles on all axial CT slices related to L3 into four categories, and automatically calculates the number of skeletal muscles and the average muscle attenuation; based on threshold processing, the amount of fat and the average HU value are calculated, so that the system can automatically measure the body composition of abdominal CT images. Summary of the invention

[0011] The purpose of the present invention is to provide an automatic body composition measurement system based on abdominal CT images and deep learning with fast measurement speed and high degree of automation.

[0012] The automatic body composition measurement system based on abdominal CT images and deep learning proposed in the present invention can automatically obtain the muscle, fat and other parameters of all layers corresponding to the third lumbar vertebra, including three modules: (I) a positioning module for positioning all axial slices corresponding to the third lumbar vertebra on the CT image; (II) a segmentation module for segmenting the four types of skeletal muscles in the axial slices corresponding to the third lumbar vertebra on the CT image; (III) an automatic measurement module for automatically measuring the body composition of clinical CT images. The overall architecture of the system is as follows: Figure 1 As shown. Among them:

[0013] The positioning module first performs an abdominal or abdominal-pelvic CT scan on the patient, and the obtained CT image includes the third lumbar vertebra and other spinal vertebrae; an experienced doctor locates and provides the position of the third lumbar vertebra. The slice corresponding to the third lumbar vertebra is selected as the positive sample input for positioning network training, and half of the remaining slices are selected as the negative sample input for positioning network training. Here, the positioning network uses the pre-trained network EfficientNet-b2, and sets the output of the fully connected layer to 2 categories to determine whether it is the slice corresponding to the third lumbar vertebra. After training, the optimal positioning model is obtained, and this positioning module can automatically screen out all axial slices corresponding to the third lumbar vertebra;

[0014] (1) The segmentation module extracts four types of skeletal muscles from all axial slices corresponding to the automatically screened third vertebra, wherein the four categories are divided based on the spatial distribution of the skeletal muscles and the muscle texture characteristics; the axial slices of the third lumbar segment of the abdominal CT image are used as the input of the training segmentation network, and the output is the four types of skeletal muscle labels in the segment area; after training, the optimal segmentation model is obtained; here, the segmentation network adopts the residual structure enhanced U-Net (2D Residual U-Net) framework widely used in medical image segmentation, and draws on its encoding and decoding path design.

[0015] The automatic measurement module uses the positioning model and segmentation model trained by the first two modules, and takes the new abdominal or abdominal pelvic CT image and the corresponding third lumbar vertebra position as input. On the one hand, according to the deep learning model, the number and average muscle attenuation of four types of skeletal muscles (i.e., the HU value of the CT image) can be calculated. On the other hand, according to the HU value intervals of subcutaneous fat and intra-abdominal fat, which are [-29, 150] and [-150, -50], threshold processing is performed to obtain the fat quantity and average muscle attenuation. The volume of the third lumbar vertebra of each patient is calculated, and the proportion of the third lumbar vertebra can be calculated. Finally, automatic measurement of body composition can be achieved.

[0016] The three modules are further described below.

[0017] (I) The positioning module (I) includes:

[0018] (1) Abdominal or abdominal pelvic CT images contain a large abdominal lumbar area. Therefore, the first step is to locate the third lumbar vertebra. The doctor selects all axial slices of the third lumbar vertebra to form the location labels required for training the deep model, totaling 4-8 layers;

[0019] (2) Positioning preprocessing: After marking the position of the third lumbar vertebra, select the slice corresponding to the third lumbar vertebra as the positive sample input, and set the label to 1; select half of the remaining slices as the negative sample input, and set the label to 0. And normalize the HU value to the interval [0,1];

[0020] (3) The positioning network is trained using CT data of 100 patients with liver cirrhosis (216 cases in the embodiment). The positioning network uses the pre-trained network EfficientNet-b2, and the fully connected layer is set to output 2 categories to determine whether it is the slice corresponding to the third lumbar vertebra. Figure 2 As shown, all slices corresponding to the third lumbar vertebra are screened out, and the model with the highest accuracy is saved as the optimal model.

[0021] (ii) The segmentation module (II) includes:

[0022] (1) Experienced doctors divide the slices of the third lumbar vertebra obtained by the positioning module into four types of skeletal muscles, and analyze multiple types of skeletal muscles in terms of volume: the rectus abdominis, external oblique muscles, internal oblique muscles, and transverse abdominal muscles (see Figure 3 Label1); psoas major, psoas minor, and quadratus lumborum on the left and right sides of the L3 vertebra (see Figure 3 Label2 and Label3); and the erector spinae and other paraspinal muscles behind L3 (see Figure 3 Label4);

[0023] (2) Segmentation preprocessing: Each CT data is normalized to a uniform size of 8, 512, and 512 in depth, height, and width, respectively, and the HU value is normalized to the interval [0, 1];

[0024] (3) Construct a four-category skeletal muscle segmentation network based on the residual structure enhanced U-Net (2D Residual U-Net) framework widely used in medical image segmentation, and draw on its encoding and decoding path design. The input of the segmentation network is the axial slice of the third lumbar segment of the abdominal CT image, and the output is the four-category skeletal muscle labels of the segment area. Among them:

[0025] 1) 2D encoding-decoding backbone network

[0026] This is a segmentation network for axial slice images. The network has a total of four layers, consisting of an encoder branch and a decoder branch with skip connections. Each layer of the encoder branch is composed of Block2D, which is downsampled by 2 times through the max pooling layer to reduce the size of the feature map and pass it to the next layer. Each layer of the decoder branch is first upsampled by 2 times, and after performing a jump connection with the feature map of the same size in the corresponding layer of the encoder branch, Block2D processing is performed. In the last layer of the decoder branch, the features are restored to the same size as the input image, and after performing a channel-wise addition operation with the features output by the 3D encoder branch, the segmentation label is output through a 1x1 convolution kernel.

[0027] 2) 3D coding branch structure

[0028] This is a contextual feature extraction network for body regions consisting of multiple layers of axial slices associated with L3, such as Figure 4 As shown. Each layer of the three-dimensional Encoder branch is composed of Block3D, a maximum pooling layer and an attention enhancement mechanism. The feature map output by Block3D is downsampled by 2 times through the maximum pooling layer (max pooling). On the one hand, the obtained features are passed to the next layer, and on the other hand, after the fuzzy area attention enhancement mechanism, the upsampled image is restored to the original image size and fused with the output features of the two-dimensional branch through the channel addition operation. In the present invention, for the input L3 axial slice volume three-dimensional image I, it is directly input into the network in the form of channel×depth×height×width, with a size of 1×8×512×512. After three times of Block3D and downsampling, the feature maps of different sizes are enhanced by the attention mechanism and then connected to the two-dimensional features through channel addition.

[0029] 3) Block2D and Block3D

[0030] As the basic structure of the segmentation network, the residual structure in 2D ResU-Net is borrowed, but there are also differences, namely improvements. First, the convolution layer is cascaded with InstanceNorm and LeakyReLU to form a basic block; then, the three groups of basic blocks are cascaded and skipped to form Block2D or Block3D with residual structure, such as Figure 5As shown in the figure. The convolution kernel of Block2D is 3x3, and the convolution kernel of Block3D is 3×3×3. Neither of these two structural blocks changes the number of channels, but can deepen the model, help extract more refined edge information, and have a better correction effect on the refinement of skeletal muscles. In particular, the three-dimensional encoding branch contains more trainable spatial and texture information, and more convolutional layers are needed to extract spatial information. Therefore, the positions of the residual connections of Block3D and Block2D are different, and Block3D has more convolutional layers, which can fully train the extracted information.

[0031] 4) Texture Attention Enhancement Mechanism

[0032] In order to better integrate 3D features with 2D features, this system builds a 3D texture attention enhancement mechanism (Texture Attention Enhancement Mechanism) based on the traditional standard enhancement block (SE Block) to perform channel-wise feature compression on a 3D network and enhance the fuzzy edge area. Figure 6 As shown. The feature map of size channel×depth×height×width extracted by Block3D is input into the attention mechanism. First, it is subjected to a global average pooling operation to obtain a feature map of size C×1×1×1. Then it passes through two fully connected layers, where the number of neurons in the first fully connected layer is C / 16 (similar to SE Block, the ratio is set to 16), and the second fully connected layer restores the original number of neurons C. This operation adds nonlinear processing and can fit the complex correlations between channels. Then the Sigmoid function is used to generate a probability map that can be input into the Texture Enhance Block. The proposed Texture Enhance Block can increase the scope of attention, as shown in formula (1),

[0033] TextureEnhance(x)=x(1-x) (1)

[0034] Among them, x represents the probability map of the input. This formula gives higher weights to the edge areas where the probability is close to 0.5, and reduces the weights to the areas far away from 0.5. By adding this part, the network no longer focuses only on the middle part of the skeletal muscle, but also strengthens the refinement of the edges, especially the shape of the skeletal muscle fiber bundles. The generated information is weighted by multiplying each pixel of the input feature map. Finally, the channel is compressed by Squeeze and Upsample, and the feature map is restored to its original size of 512×512 in height and width for fusion with the two-dimensional network.

[0035] 5) Construct loss function

[0036] For a 2D abdominal CT image I of size H×W, the commonly used multi-class cross entropy loss function in medical image segmentation tasks is combined with the Dice loss function to segment four types of skeletal muscles, as shown in formula (2):

[0037]

[0038] Among them, C = 5, representing four types of skeletal muscles and background. c represents the weight of the c-th skeletal muscle, Indicates the gold standard value of pixel i belonging to the c-th class label, It indicates the prediction result of pixel i as the c-th class label, H and W represent the height and width of the two-dimensional axial image respectively.

[0039] The sizes of the four types of skeletal muscles and the background are very different, which means that there is a class imbalance problem that causes the segmentation framework to be unstable. Therefore, during the training process, in the loss function, it is necessary to penalize low-confidence predictions (such as Label2 and Label3) by setting weights. Specifically, first, in the training image, the pixel ratios of the four types of skeletal muscles and the background are counted, and then the skeletal muscles with a small ratio are set to a small weight, and the skeletal muscles with a large ratio are set to a large weight, as shown in formula (3).

[0040]

[0041] Among them, H, W and D represent the height, width and depth of the two-dimensional image, N c represents the pixel count of the c-th label. c The prior statistics of ensure the class-balanced optimization of the loss function.

[0042] (4) Using CT data of several patients with liver cirrhosis (216 in the embodiment), the segmentation network is trained to automatically segment four types of skeletal muscles. The model with the highest Dice index is saved as the optimal segmentation model.

[0043] (III) The automatic measurement module (III) includes:

[0044] (1) After positioning preprocessing, the new abdominal or abdominal-pelvic CT image is input into the trained optimal positioning model to obtain 4-8 slices corresponding to the third lumbar vertebra;

[0045] (2) The 4-8 slices corresponding to the third lumbar vertebra were preprocessed and input into the optimal segmentation model to obtain four types of skeletal muscles;

[0046] (3) At the same time, the 4-8 slices corresponding to the third lumbar vertebra were threshold processed to obtain the volume of the third lumbar vertebra and the corresponding areas of intra-abdominal fat and subcutaneous fat;

[0047] (4) The quantity and average muscle attenuation of the four types of skeletal muscles, intra-abdominal fat and subcutaneous fat are calculated to achieve automatic measurement of body composition.

[0048] The present invention was trained and verified on CT data of 216 clinical patients with liver cirrhosis, and the average Dice of the four types of skeletal muscle segmentation results reached 0.9283, and the average surface distance was 0.6779 mm. Further generalization verification was performed on CT data of 101 non-cirrhotic patients, and the average Dice was 0.9277, and the average surface distance was 0.8279 mm. The present invention can not only measure the four types of skeletal muscles corresponding to the third lumbar vertebra in the CT image, but also use threshold processing to obtain other body components such as subcutaneous fat and intra-abdominal fat in this area. The present invention obtains the body components corresponding to the third lumbar vertebra in batches in 2-3 seconds, which is of great significance for clinically assisting in the diagnosis of complications of patients with liver cirrhosis.

[0049] Advantages of the present invention:

[0050] (1) The automatic body composition measurement system based on abdominal CT images and deep learning proposed in the present invention can automatically measure the four types of skeletal muscles corresponding to the third lumbar vertebra and the intra-abdominal fat and subcutaneous fat, which is helpful for the diagnosis and treatment of complications of cirrhosis (such as sarcopenia).

[0051] (2) The positioning method of the present invention uses a pre-trained network to quickly and accurately locate the third lumbar vertebra. The segmentation method is based on the advanced U-Net architecture and enhances it; considering the characteristics of different types of skeletal muscles such as size and fiber bundle shape, the segmentation method also combines the texture attention mechanism for enhancing the fuzzy area of ​​skeletal muscle edges; extracting the 3D encoding branch of fiber bundle features; and reducing the loss function of class imbalance. The system of the present invention accurately segmented multiple types of skeletal muscles in all axial slices related to L3 on more than 300 abdominal or abdominal and pelvic CT images.

[0052] The automatic measurement system uses the optimal model of positioning and segmentation network. Compared with other traditional methods, it has the advantages of fast measurement speed, high degree of automation, and does not require a large amount of memory. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the overall framework diagram of the system of the present invention.

[0054] Figure 2 Schematic diagram of screening out all slices corresponding to the third lumbar vertebra.

[0055] Figure 3 Schematic diagram of the four types of skeletal muscle.

[0056] Figure 4 Schematic diagram of the skeletal muscle segmentation network.

[0057] Figure 5 Schematic diagram of the residual module in the skeletal muscle segmentation network.

[0058] Figure 6 Texture feature enhancement module in skeletal muscle segmentation network.

[0059] Figure 7 The following are the muscle and fat segmentation results of four patients with liver cirrhosis.

[0060] Figure 8 It is a flowchart of the application of the system of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described below with reference to examples and a system overall framework diagram.

[0062] Taking the clinical original images collected from the clinic as an example, the application process of the present invention for automatically measuring body components is as follows: Figure 8 shown.

[0063] Module 1 screens the axial slices corresponding to the third lumbar vertebra. First, the clinical CT image set of 101 non-cirrhotic patients was pre-processed for positioning, that is, the three-dimensional data set was saved as a picture with the extension "png format", and then randomly flipped vertically and affine transformed. All slices were input into the positioning model trained with the 216 cirrhosis data sets to obtain the prediction results of the third lumbar vertebra slices, as shown in Table 1. In Table 1, "0" indicates an axial slice that is not corresponding to the third lumbar vertebra, and "1" indicates an axial slice corresponding to the third lumbar vertebra. Accuracy indicates accuracy.

[0064] Module 2 is the skeletal muscle segmentation process. First, the slice corresponding to the third lumbar vertebra located in module 1 is segmented and preprocessed, and the size is padded to 8, 512, and 512 respectively in depth, height, and width, and the value is normalized to the interval [0,1]. Then the image is input into the segmentation model trained with the 216 liver cirrhosis dataset to obtain the segmentation results of four types of skeletal muscles.

[0065] Tables 1 and 2 show the experimental results of 101 cases of non-cirrhosis data on the localization model and segmentation model. In the localization model, the slice corresponding to the third lumbar vertebra is used as the gold standard for training. The higher the accuracy obtained in Table 1, the better the localization effect. In the segmentation model, four types of skeletal muscles are used as the gold standard for training. The higher the segmentation indexes Dice, Sensitivity, Precision, and Specificity, the better, and the smaller the ASSD (mm), the better. Table 2 calculates the segmentation indexes of the four skeletal muscles and their means. Here, Dice is a commonly used indicator for evaluating image segmentation accuracy, Sensitivity represents sensitivity, Precision represents accuracy, Specificity represents specificity, and ASSD represents average surface distance.

[0066] Table 1. Statistical results of the localization network experiment on CT images of 101 patients without liver cirrhosis

[0067]

[0068] Table 2. Statistical results of the segmentation network experiment on 101 non-cirrhotic CT images

[0069]

[0070] Table 3. Average body composition of 101 non-cirrhotic CT images

[0071]

[0072] References

[0073] [1]Schweitzer L, Geisler C, Pourhassan M, et al. What is the best reference site for a single MRI slice to assess whole-body skeletal muscle and adipose tissue volumes in healthy adults? [J].The American journal of clinical nutrition,2015,102(1):58-65.

[0074] [2]Chang C D,Wu J S,Mhuircheartaigh J N,et al.Effect of sarcopenia onclinical and surgical outcome in elderly patients with proximal femurfractures[J].Skeletal radiology,2018,47(6):771-777.

[0075] [3]Fuchs G,Chretien Y R,Mario J,et al.Quantifying the effect of slicethickness,intravenous contrast and tube current on muscle segmentation:Implications for body composition analysis[J].European radiology,2018,28(6):2455-2463.

[0076] [4]Janne W,Thobias R,Sofia T,et al.Precision of MRI-based bodycomposition measurements of postmenopausal women[J].Plos One,2018,13(2):e0192495.

[0077] [5]Lenchik L,Boutin R.Sarcopenia:Beyond Muscle Atrophy and into theNew Frontiers of Opportunistic Imaging,Precision Medicine,and MachineLearning[J].Seminars in musculoskeletal radiology,2018,22(03):307-322.

[0078] [6]Popuri K,Cobzas D,Esfandiari N,et al.Body Composition Assessmentin Axial CT Images Using FEM-Based Automatic Segmentation of Skeletal Muscle[J].IEEE Transactions on Medical Imaging,2016,35(2):512-520.

[0079] [7]Dabiri S,Popuri K,Feliciano E M C,et al.Muscle segmentation inaxial computed tomography(CT)images at the lumbar(L3)and thoracic(T4)levelsfor body composition analysis[J].Computerized Medical Imaging and Graphics,2019,75:47-55.

[0080] [8]Castiglione J,Somasundaram E,Gilligan L A,et al.Automatedsegmentation of abdominal skeletal muscle on pediatric CT scans using deeplearning[J].Radiology:Artificial Intelligence,2021,3(2):e200130.

[0081] [9]Park H J,Shin Y,Park J,et al.Development and validation of a deeplearning system for segmentation of abdominal muscle and fat on computedtomography[J].Korean journal of radiology,2020,21(1):88-100.

[0082]

[10] Burns J E,Yao J,Chalhoub D,et al.Amachine learning algorithm toestimate sarcopenia on abdominal CT[J].Academic radiology,2020,27(3):311-320.

[0083]

[11] Jiawei H,Shen H,Chen B,et al.Segmentation of Paraspinal Musclesat Varied Lumbar Spinal Levels by Explicit Saliency-Aware Learning[C] / / 23thInternational Conference on Medical Image Computing and Computer AssistedIntervention(MICCAI).2020.

[0084]

[12] Barnard R,Tan J,Roller B,et al.Machine learning for automaticparaspinous muscle area and attenuation measures on low-dose chest CT scans[J].Academic radiology,2019,26(12):1686-1694.

[0085]

[13] Borrelli P,Kaboteh R,Enqvist O,et al.Artificial intelligence-aided CT segmentation for body composition analysis:a validation study[J].European Radiology Experimental,2021,5(1):1-6。

Claims

1. An automatic body composition measurement system based on abdominal CT images and deep learning, characterized in that: Can automatically obtain Take the muscle and fat parameters corresponding to all levels of the third lumbar vertebra, including three modules: A positioning module, used for positioning all axial slices corresponding to the third lumbar vertebra on the CT image; A segmentation module is used to segment the four types of skeletal muscles in the axial slices corresponding to the third lumbar vertebra on the CT image; Automatic measurement module for automatic measurement of body composition in clinical CT images; in: The positioning module first performs a CT scan of the abdomen or abdomen and pelvis of the patient, and the obtained CT image includes the third lumbar vertebra and other spinal vertebrae; an experienced doctor locates and provides the position of the third lumbar vertebra; selects the slice corresponding to the third lumbar vertebra as the positive sample input for positioning network training, and selects half of the remaining slices as the negative sample input for positioning network training; here, the positioning network uses the pre-trained network EfficientNet-b2, sets the output of the fully connected layer to 2 categories, and judges whether it is the slice corresponding to the third lumbar vertebra; after training, the optimal positioning model is obtained; The segmentation module extracts four types of skeletal muscles from all axial slices corresponding to the automatically screened third vertebra, wherein the four categories are divided based on the spatial distribution of skeletal muscles and muscle texture characteristics; the axial slices of the third lumbar segment of the abdominal CT image are used as input for training the segmentation network, and the output is the four types of skeletal muscle labels in the segment area; after training, the optimal segmentation model is obtained; here, the segmentation network adopts the residual structure enhanced U-Net framework for medical image segmentation, and draws on its encoding and decoding path design; The automatic measurement module uses the positioning model and segmentation model trained by the first two modules, takes the new abdominal or abdominal pelvic CT image and the corresponding third lumbar vertebra position as input, and calculates the number and average muscle attenuation of four types of skeletal muscles, that is, the HU value of the CT image, based on the deep learning model. On the other hand, threshold processing is performed based on the HU value intervals of subcutaneous fat and intra-abdominal fat, which are [-29, 150] and [-150, -50] respectively, to obtain the fat quantity and average muscle attenuation; the volume of the third lumbar vertebra of each patient is calculated, and the proportion of the third lumbar vertebra is calculated, so as to finally realize automatic measurement of body composition.

2. The automatic body composition measurement system according to claim 1, characterized in that: The positioning module includes: (1) Since the abdominal or abdominal pelvic CT images contain a large abdominal lumbar area, the third lumbar vertebra must be located first. The doctor selects all axial slices of the third lumbar vertebra to form the location labels required for training the deep model, totaling 4-8 layers; (2) Positioning preprocessing: After marking the position of the third lumbar vertebra, select the slice corresponding to the third lumbar vertebra as the positive sample input, and set the label to 1; select half of the remaining slices as the negative sample input, and set the label to 0; and normalize the HU value to the interval [0, 1]; (3) The CT data of a patient with liver cirrhosis is used to train the positioning network. The positioning network uses the pre-trained network EfficientNet-b2. The fully connected layer is set to output 2 categories to determine whether it is the slice corresponding to the third lumbar vertebra; all slices corresponding to the third lumbar vertebra are screened out, and the model with the highest accuracy is saved as the optimal model.

3. The automatic body composition measurement system according to claim 2, characterized in that: The segmentation module includes: (1) Experienced doctors divide the slices of the third lumbar vertebra obtained by the positioning module into four types of skeletal muscles, so as to analyze multiple types of skeletal muscles in terms of volume: the rectus abdominis, external oblique abdominal muscles, internal oblique abdominal muscles, and transverse abdominal muscles in the anterior periphery of the L3 abdomen; the psoas major, psoas minor, and quadratus lumborum muscles on the left and right sides of the L3 vertebra; and the erector spinae and other paraspinal muscles in the posterior part of the L3; (2) Segmentation preprocessing: Each CT data is normalized to a uniform size of 8, 512, and 512 in depth, height, and width, respectively, and the HU value is normalized to the interval [0, 1]; (3) Constructing a four-category skeletal muscle segmentation network based on the residual structure enhanced U-Net framework widely used in medical image segmentation, drawing on its encoding and decoding path design; the input of the segmentation network is the axial slice of the third lumbar segment of the abdominal CT image, and the output is the four-category skeletal muscle labels of the segment area; (4) The segmentation network was trained using CT data from several patients with liver cirrhosis to automatically segment four types of skeletal muscle. The model with the highest Dice index was saved as the optimal segmentation model.

4. The automatic body composition measurement system according to claim 3, characterized in that: In the segmentation module: 1) 2D encoding-decoding backbone network A segmentation network for axial slice images; the network has a total of four layers, consisting of an encoder branch and a decoder branch with jump links; each layer of the encoder branch is composed of Block2D, and is downsampled by 2 times through the maximum pooling layer to reduce the size of the feature map and pass it to the next layer; each layer of the decoder branch is first upsampled by 2 times, and after performing a channel-wise jump connection with the feature map of the same size of the corresponding layer of the encoder branch, Block2D processing is performed; in the last layer of the decoder branch, the features are restored to the same size as the input image, and after performing a channel-wise addition operation with the features output by the 3D encoding branch, the segmentation label is output through a 1x1 convolution kernel; 2) 3D coding branch structure It is a context feature extraction network for body regions composed of multiple layers of axial slices related to L3. Each layer of this 3D encoding branch consists of Block3D, maximum pooling layer and attention enhancement mechanism. The feature map output by Block3D is downsampled by 2 times through the maximum pooling layer. The obtained features are passed to the next layer on the one hand, and on the other hand, they are upsampled to the original image size after passing through the fuzzy region attention enhancement mechanism, and fused with the output features of the 2D branch through channel-wise addition operation. 3) Block2D and Block3D As the basic structure of the segmentation network, the residual structure in 2D ResU-Net is used as a reference. The convolution layer is cascaded with InstanceNorm and LeakyReLU to form a basic block. The three groups of basic blocks are cascaded and skipped to form Block2D or Block3D with residual structures. The convolution kernel of Block2D is 3x3, and the convolution kernel of Block3D is 3×3×3. 4) Texture Attention Enhancement Mechanism In order to better integrate 3D features with 2D features, based on the traditional standard enhancement block (SE Block), a channel-wise feature compression is constructed on a 3D network to enhance the 3D texture attention enhancement mechanism in the blurred edge area. The feature map of size channel×depth×height×width extracted by Block3D is input into the attention mechanism. First, it is subjected to a global average pooling operation to obtain a feature map of size C×1×1×1. Then it passes through two fully connected layers, where the number of neurons in the first fully connected layer is C / 16, and the second fully connected layer restores the original number of neurons C. This operation increases nonlinear processing and can fit the complex correlations between channels. Then, a probability map that can be input into the Texture Enhance Block is generated through the Sigmoid function. The Texture Enhance Block can increase the range of attention, as shown in formula (1). TextureEnhance(x)=x(1-x) (1) Among them, x represents the input probability map; this formula gives higher weights to the edge areas with probabilities close to 0.5 and reduces the weights to areas far away from 0.5; by adding this part, the network no longer focuses only on the middle part of the skeletal muscle, but also strengthens the refinement of the edges, especially the shape of the skeletal muscle fiber bundles; the generated information is weighted by multiplying each pixel of the input feature map; finally, the channel is compressed by Squeeze and Upsample, and the feature map is restored to its original size of 512×512 in height and width so that it can be fused with the two-dimensional network.

5. The automatic body composition measurement system according to claim 4, characterized in that: Construct the loss function for segmentation module training: For a 2D abdominal CT image I of size H×W, the multi-class cross entropy loss function is combined with the Dice loss function in the medical image segmentation task to segment four types of skeletal muscles. The loss function formula is shown in (2): Among them, C = 5, representing four types of skeletal muscles and background; ω c represents the weight of the c-th skeletal muscle, Indicates the gold standard value of pixel i belonging to the c-th class label, represents the prediction result of pixel i as the c-th class label, H and W represent the height and width of the two-dimensional axial image respectively; weight ω c Determine as follows: First, in the training image, the pixel ratios of the four types of skeletal muscles and the background are counted, and then the skeletal muscles with a small ratio are given a small weight, and the skeletal muscles with a large ratio are given a large weight, as shown in formula (3): Among them, H, W and D represent the height, width and depth of the two-dimensional image, N c Represents the pixel count statistics of the c-th label; The segmentation network was trained using CT data of several patients with liver cirrhosis to automatically segment four types of skeletal muscle; the model with the highest Dice index was saved as the optimal segmentation model.

6. The automatic body composition measurement system according to claim 5, characterized in that: The automatic measurement module includes: (1) After positioning preprocessing, the new abdominal or abdominal-pelvic CT image is input into the trained optimal positioning model to obtain 4-8 slices corresponding to the third lumbar vertebra; (2) The 4-8 slices corresponding to the third lumbar vertebra were preprocessed and input into the optimal segmentation model to obtain four types of skeletal muscles; (3) At the same time, the 4-8 slices corresponding to the third lumbar vertebra were threshold processed to obtain the volume of the third lumbar vertebra and the corresponding areas of intra-abdominal fat and subcutaneous fat; (4) The quantity and average muscle attenuation of the four types of skeletal muscles, intra-abdominal fat and subcutaneous fat are calculated to achieve automatic measurement of body composition.

Images