Body composition analysis and comparison system based on medical image

Abstract

The invention provides a body composition analysis and comparison system based on a medical image. The body composition analysis and comparison system comprises a composition statistics module and an analysis and comparison module, the component statistics module performs preliminary statistics on data related to body components based on the segmented medical image; the component statistics module comprises a pixel statistics module for determining the proportional relation of different body components by counting the number of different types of pixel points in the segmented medical image, an area measurement module for converting the number of pixels into a real physical area on the basis of pixel statistics, a density evaluation module for evaluating the density of the physical area, and an area calculation module for calculating the density of the physical area. On the basis of pixel statistics and area measurement, calculating an average gray value of the segmented regions, and estimating physical characteristics of body tissues corresponding to the medical image; and the analysis and comparison module carries out variable quantity calculation, visualization processing of a change area and generation of an analysis and comparison report based on different data obtained by the component statistics module. According to the invention, the comprehensiveness, accuracy and consistency of medical image recognition and analysis can be improved.

Description

Technical Field

[0001] The present invention relates to a medical image analysis technology, and in particular to a system for segmenting, refining, analyzing and comparing feature information in medical images. Background Art

[0002] Medical images are one of the basic materials for doctors to analyze patients' bodies. The recognition and interpretation of medical images require very professional medical knowledge. Manual recognition and interpretation are subjective, time-consuming and labor-intensive.

[0003] With the development of image recognition and analysis technology, many existing technologies have proposed methods for interpreting medical images through machine recognition and analysis. For example, Chinese patent document CN119151967A discloses a medical image analysis method and system based on plain scan CT data. By using a deep learning model, a three-dimensional model of a healthy lung is automatically reconstructed and compared with the patient's actual lung model. The automated comparison reduces human error and improves analysis efficiency, allowing doctors to quickly and accurately identify the lesion area and the extent of the disease. For example, Chinese patent document CN118429632A discloses an image processing method, device, electronic device, and computer-readable storage medium. By segmenting a CT image, a target bone segmentation image is obtained, thereby solving the technical problem of difficulty in ensuring the accuracy of bone segmentation in CT images due to the close proximity of CT images of cartilage and soft tissue. The above technical documents disclose methods for analyzing and comparing CT images from different angles, solving some recognition problems in medical images. However, these technical documents are all for image analysis of a certain organ or a certain layer, and do not provide a complete or comprehensive method for recognizing and analyzing CT images.

[0004] With the development of medical image recognition technology, medical images can not only identify lesions in a certain part or aspect, but can also be used to evaluate human body structure, especially body composition analysis. Existing body composition analysis methods mainly include the following methods: (1) Traditional manual measurement method: Doctors manually draw the region of interest (ROI) in the CT image to calculate indicators such as muscle area, fat area and bone density. This method is time-consuming, highly subjective and has differences between operators. (2) Semi-automated software tools: General medical image analysis software such as SliceOmatic and ImageJ provide semi-automated measurement functions based on threshold and edge detection, but still require a lot of manual intervention and adjustment. (3) Commercial single analysis system: GE and Siemens both have software that can automatically analyze the body composition of a single CT scan, but the function is mainly limited to static analysis. In this regard, there are also relevant patent documents disclosed in the prior art. For example, Chinese patent document CN118505615A discloses a CT image body composition analysis system and method based on deep learning. The document proposes to perform image segmentation based on the UNet image segmentation method and the hierarchical visual enhancement ResNet, and then perform body composition analysis. However, image segmentation processing cannot obtain analysis results. It also contains very complex image information and data transformation processes. Moreover, in clinical practice, doctors also need to compare medical images of the same patient at different time points to evaluate disease progression or treatment effects. This comparison is called "longitudinal comparative analysis". Currently, longitudinal comparative analysis is mainly achieved in the following ways: (1) Manual comparison method: doctors manually record and compare measurement data at different time points, and calculate the change amount and change rate. (2) Table recording method: enter data into a spreadsheet and calculate the change situation through formulas. (3) Independent software combination: use independent image processing software for segmentation, and then use statistical software for comparative analysis.

[0005] Therefore, there is a lack of a system that can completely and comprehensively analyze, identify and output medical images. Summary of the Invention

[0006] The technical purpose of the present invention is to provide a body composition analysis and comparison system based on medical images to address the problems existing in the background technology, so as to improve the comprehensiveness, accuracy and consistency of medical image recognition and analysis.

[0007] The technical solution adopted by the present invention to solve the above technical problems is:

[0008] A body composition analysis and comparison system based on medical images, including a composition statistics module and an analysis and comparison module;

[0009] The component statistics module is configured to perform preliminary statistics on data related to body composition based on the segmented medical image;

[0010] The component statistics module includes:

[0011] The pixel statistics module is configured to determine the proportional relationship of different body components by counting the number of pixels of different categories in the segmented medical image.

[0012] The area measurement module is configured to convert the number of pixels into the real physical area based on pixel statistics.

[0013] a density assessment module configured to calculate an average grayscale value of the segmented region based on pixel statistics and area measurement, and estimate physical properties of body tissue corresponding to the medical image;

[0014] The analysis and comparison module is configured to calculate the variation, visualize the variation area, and generate an analysis and comparison report based on the different data obtained by the component statistics module.

[0015] Preferably, the specific steps of performing pixel statistics by the pixel statistics module include:

[0016] (1) Classification labeling: assigning different category labels (such as subcutaneous fat, muscle, visceral fat, background, etc.) to each pixel in the segmented medical image;

[0017] (2) Color mapping: Use different colors to mark different categories of areas; for example:

[0018] Subcutaneous fat (Bone): white (0, 0, 255), muscle (Muscle): red (255, 0, 0), visceral fat (Fat): yellow (255, 255, 0), background (Background): black (0, 0, 0);

[0019] (3) Pixel counting: Traverse all pixels in the segmented medical image and count the number of pixels corresponding to each category:

[0020]

[0021] in, Representation category The total number of pixels; Representing coordinates Whether the pixel at belongs to the category , if yes, take 1, otherwise take 0; and are the height and width of the image respectively;

[0022] (4) Proportion calculation:

[0023] Count the pixel ratios of different categories:

[0024]

[0025] in, For category Scale in the image; is the total number of pixels in the entire image.

[0026] Preferably, the specific steps of the area measurement module for performing pixel area conversion are:

[0027] Medical images usually have a known pixel spatial resolution (Pixel Spacing), which is the actual physical size corresponding to each pixel, usually in millimeters (mm) or centimeters (cm).

[0028] Set the pixel spacing of medical images to , unit: mm², then the area of ​​each category is:

[0029]

[0030] in, for The area of ​​the category (in mm²), and are the pixel pitches in the horizontal and vertical directions, respectively.

[0031] Preferably, the estimation step of the density assessment module specifically includes:

[0032] (1) Obtain the grayscale value of each coordinate position in the medical image and calculate the average grayscale density as follows:

[0033]

[0034] in, For category The average grayscale density, unit HU (Hounsfield, Hounsfield unit), For coordinates Gray value at ;

[0035] (2) According to the grayscale values ​​at different coordinate positions, the standard deviation of tissue density at the corresponding parts is calculated as follows:

[0036]

[0037] (3) Determine body composition based on the grayscale density and tissue density standard deviation calculated in the first two steps.

[0038] Preferably, the analysis and comparison module includes a change calculation module, a visualization generation module and a report generation module; the change calculation module includes a volume change calculation module, a HU value change calculation module and a clinical significance judgment module, and the clinical significance judgment module judges the significance of the change based on the volume change calculation module and the HU value change calculation module using a preset threshold; the visualization generation module uses a heat map and / or a difference map for visual display; the report generation module generates an analysis report based on the results of the change calculation module and the visualization generation module.

[0039] Preferably, the specific calculation method of the variation calculation module is:

[0040] For the volume of tissue c at two different time points (Baseline and Follow-up), calculate its change and rate of change :

[0041]

[0042]

[0043] in, is the tissue area at baseline (initial scan), It is the tissue area at the time of follow-up. The above tissue area can be obtained by an automatic delineation algorithm. The automatic delineation algorithm here can be implemented in a manner disclosed in the prior art.

[0044] according to Compare the size of the data and push corresponding prompts, where Threshold increase Represents a preset threshold value, which is a standard threshold value preset based on known data such as medical experience, big data models, etc., and is used for comparison in the present invention.

[0045] Preferably, the specific calculation contents of the HU value change calculation module include:

[0046] HU mean change:

[0047]

[0048] HU change rate:

[0049]

[0050] in, Indicates the HU value at follow-up (reexamination), represents the HU value at baseline (first scan), Indicates the rate of change of HU value.

[0051] Preferably, the visualization generation module generates a difference map by calculating the pixel-level difference between the two scans, and highlights the changed area. The specific calculation method is:

[0052]

[0053] in, Indicates the pixel value at the follow-up (review) scan, Indicates the pixel value at the baseline (first scan), is the difference between the two;

[0054] like , the volume or density of the area increases, marked in blue,

[0055] like , the volume or density of the area decreases and is marked in red.

[0056] like , then there is no obvious change in the area and no color is marked.

[0057] Preferably, the algorithm steps executed by the report generation module include:

[0058] (1) Obtain report information and suggestions based on the automatic delineation algorithm; and make adjustments based on manual modification input to regain the area and opinions corresponding to the manual modification;

[0059] (2) Calculation of skeletal muscle index:

[0060]

[0061] Among them, SMI represents skeletal muscle index, unit is cm² / m², SMA represents skeletal muscle area, unit is cm², calculated by CT images;

[0062] (3) Calculate the SMRA value:

[0063]

[0064] Among them, SMRA represents the mean radiation attenuation value of skeletal muscle, the numerator represents the sum of all HU values ​​in the muscle area, and the denominator represents the total number of pixels in the muscle area, that is, the total pixel area.

[0065] Preferably, the system further includes a medical image processing module configured to preprocess and segment the acquired initial medical images. The preprocessing includes image formatting, image adaptive enhancement, image edge enhancement, and data augmentation. The image segmentation employs a UNet encoding module, an H-ViTER processing module, a multi-layer ViT processing module, a UNet decoding module, and an output convolutional layer;

[0066] The UNet encoding module extracts features from the initial medical images and improves the representational ability of the features through successive downsampling. The UNet encoding module includes an input convolutional layer and a pooling layer;

[0067] The H-ViTER module includes a multi-level convolutional module, a residual module, a vision Transformer module, and a fusion module;

[0068] The multi-level convolutional module extracts features using multi-scale convolutional kernels of 3×3, 4×4, and 5×5 respectively. Among them, the 3×3 convolution is responsible for extracting local detail information, the 4×4 convolution expands the receptive field to improve the understanding of the spatial structure, and the 5×5 convolution further enhances the global feature capture ability;

[0069] The residual module uses residual connections to maintain gradient flow and prevent deep network degradation. The expression of the residual connection is as follows:

[0070]

[0071] where Input represents the input feature map of the residual module; Output represents the output feature map of the residual module, that is, the transformed features are superimposed on the basis of the input features to achieve feature enhancement and gradient transmission; f(Input) represents the features obtained after operations such as convolution, normalization, and activation on the input feature map;

[0072] The vision Transformer module captures long-range dependencies to improve the global information modeling ability. The calculation process of the vision Transformer module is as follows:

[0073] (1) Input feature transformation: Convert the convolutional features extracted by H-ViTER into a fixed-length Patch sequence;

[0074] (2) Self-attention calculation: Use a multi-head self-attention model to calculate the global relationship between features:

[0075]

[0076] where Q, K, and V are the query, key, and value matrices after mapping the input features respectively, is the dimension of the key, T represents the matrix;

[0077] (3) Feature fusion: Integrate the information extracted by multiple layers of ViT and enhance the expression ability through residual feature fusion (RFF);

[0078] The fusion module includes residual information fusion and global feature fusion. The residual information fusion combines the features extracted by ViT and CNN to improve the network's expressiveness. The global feature fusion performs feature fusion through a multi-layer Transformer to ensure information flow between different scales.

[0079] The multi-layer ViT processing module embeds the ViT structure at different levels and combines the features extracted by CNN to improve the global feature learning ability. The multi-layer ViT processing module includes a first ViT module configured to perform Transformer processing on local convolutional features to supplement long-range information; a second ViT module configured to further enhance semantic features and improve global information modeling capabilities;

[0080] The UNet decoding module is used to restore the original resolution; the UNet decoding module uses skip connections for information recovery and feature fusion to avoid information loss caused by multi-layer downsampling, and combines shallow and deep features to improve segmentation accuracy; the decoding path of the UNet decoding module includes a three-layer decoder, and each layer uses upsampling to restore features;

[0081] The output convolution layer performs feature fusion through a 3×3 convolution layer to generate the final segmentation result.

[0082] The beneficial effects of the present invention are:

[0083] (1) The present invention provides a system for comprehensive identification and analysis of medical images. The system can not only automatically analyze the body composition in a single CT image, but also accurately compare CT images at different time points, thereby achieving a quantitative assessment of the dynamic changes in the patient's body composition. Although the existing technology has made certain progress in the field of CT image body composition analysis, it lacks an integrated solution that organically combines automatic segmentation and longitudinal comparison functions. The present invention addresses this technical gap and realizes the automation of the entire process from image preprocessing, automatic segmentation to longitudinal comparison analysis, providing clinicians with a more efficient and accurate tool for evaluating changes in body composition.

[0084] (2) The system of the present invention realizes an integrated solution for medical images. Unlike the existing method that requires the use of multiple independent software, the present invention provides a full-process automated solution from image preprocessing, automatic segmentation to longitudinal comparative analysis, which greatly improves work efficiency.

[0085] (3) The system of the present invention provides a visual comparison tool that can generate heat maps and difference maps that intuitively display changes in body composition, allowing doctors to quickly understand and evaluate changes in patients' body composition. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is an overall framework diagram of embodiment 1 of the present invention.

[0087] Figure 2 This is a CT medical image before image preprocessing in Example 2 of the present invention.

[0088] Figure 3 This is a CT medical image after image preprocessing in Example 2 of the present invention.

[0089] Figure 4 This is a schematic diagram of a preliminary report of the results obtained by the present invention in Example 4 of the present invention.

[0090] Figure 5 Schematic diagram of manually correcting a report in Example 4 of the present invention.

[0091] Figure 6 This is a revised report diagram in Example 4 of the present invention. DETAILED DESCRIPTION

[0092] In order to enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention is described in detail, clearly, and completely in the following embodiments in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Moreover, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts, any modifications, equivalent replacements, improvements, etc., should be included in the scope of protection of the present invention.

[0093] To help those skilled in the art better understand the technical solutions of the present invention, we first explain some of the technical terms of the present invention. In the context of the present invention, these explanations are primarily intended to help those skilled in the art understand and implement the technical solutions of the present invention, and are not intended to limit the present invention. Furthermore, other disclosures outside of the present invention may provide different technical interpretations of these technical terms. To avoid technical ambiguity, these explanations should be considered in conjunction with the present invention when understanding the present invention.

[0094] Medical imaging

[0095] Medical imaging is a crucial tool in the medical field. Using tools, equipment, and instruments, they image and analyze human tissues and organs, helping doctors determine disease diagnoses and treatment plans. Medical images primarily include X-rays, CT scans, MRIs, and other types. With technological advancements, the scope of medical imaging is expanding. While CT scan images are used as an example in this disclosure, this does not limit the application of CT images to this technology.

[0096] HU value

[0097] In this application, HU refers to the Hounsfield Unit, a numerical value used to represent tissue density in CT images. The HU value is a standardized value. In the prior art, the HU value standard uses water as the "baseline," with the CT HU value of water defined as 0 and the CT HU value of air as approximately -1000. The larger the HU value, the harder and denser the object. For example, tissues like bone can have HU values ​​as high as several hundred or even thousands.

[0098] Body composition analysis

[0099] Body composition analysis, also known as human body composition analysis, involves measuring various body components through various methods, such as the density and distribution of body fat, muscle, and bone, as well as body water content, muscle mass, fat mass, and bone mineral content. Existing body composition tests can be performed using methods such as sebum measurement, bioelectrical impedance, and X-ray absorptiometry. These analyses are primarily used to infer a subject's nutritional status, basal metabolic rate, and obesity level.

[0100] Example 1: A body composition analysis and comparison system based on medical images.

[0101] like Figure 1 As shown, this embodiment provides a body composition analysis and comparison system based on medical images, including a composition statistics module 100 and an analysis and comparison module 200.

[0102] The component statistics module includes:

[0103] The pixel statistics module 101 is configured to determine the proportional relationship of different body components by counting the number of pixels of different categories in the segmented medical image.

[0104] The area measurement module 102 is configured to convert the number of pixels into a real physical area based on pixel statistics.

[0105] The density assessment module 103 is configured to calculate the average grayscale value of the segmented region based on pixel statistics and area measurement, and estimate the physical properties of the body tissue corresponding to the medical image;

[0106] The area measurement module 102 receives the statistical results of the pixel statistics module 101 and performs further processing, and the density assessment module 103 receives the calculation results of the pixel statistics module 101 and the area measurement module 102 and performs further processing.

[0107] The analysis and comparison module 200 performs longitudinal analysis and comparison based on the calculation results of the pixel statistics module 101, the area measurement module 102 and the density assessment module 103, and respectively calculates the volume change and the HU value change through the change calculation module 201 to determine parameters such as clinical significance, and generates a visual heat map and difference map through the visualization generation module 202, and can also modify and generate an analysis report through the report generation module 203.

[0108] Embodiment 2, image processing module.

[0109] In order to enable those skilled in the art to better understand and implement the technical solution of the present invention, this embodiment first provides an image processing module to process the medical images that need to be analyzed. At the same time, this embodiment also provides the specific conditions of the patients corresponding to the medical images to verify the technical solution of the present invention.

[0110] The image processing module of the present invention is used to preprocess and segment the collected initial medical images, and includes an image preprocessing module and an image segmentation module respectively, wherein the image preprocessing module performs image formatting, image adaptive enhancement, image edge enhancement and data enhancement operations on the original CT image.

[0111] This embodiment takes a patient undergoing abdominal CT examination as an example. The original CT image is in DICOM format, with a resolution of 768×768 pixels, a slice thickness of 2.5 mm, and an uneven CT value range and noise.

[0112] During the image formatting phase, the image preprocessing module takes DICOM image format as input and converts it into an internal processing format, such as PNG. For size standardization, the image preprocessing module resamples CT images to a resolution range of 512×512 to 1024×1024 pixels, with uniform slice thicknesses of 1–5 mm. Furthermore, the image preprocessing module normalizes the images, remapping CT values ​​(Hounsfield Units, HU) to a standard range (e.g., -29 to 150 HU for muscle tissue, -190 to -30 HU for adipose tissue, and >150 HU for bone tissue). In this example, after image formatting, the original CT images were converted to a standard size of 512×512 pixels, slice thicknesses were reconstructed to uniform intervals of 2.0 mm, and CT values ​​were remapped to a standard range and normalized, effectively mitigating the impact of device variability and inconsistent scanning parameters. Comparison images before and after processing show that the formatted images preserve the original anatomical structure information while significantly improving the contrast and boundary definition between different tissues, laying the foundation for subsequent automatic segmentation.

[0113] In the image adaptive enhancement stage, the image preprocessing module uses an improved particle swarm optimization (PSO) algorithm to perform adaptive enhancement on the formatted CT images. By intelligently searching for the optimal image enhancement parameter combination, personalized enhancement of CT images under different scanning conditions is achieved. The specific intelligent search algorithm is as follows:

[0114] (1) Parameter encoding: Encode the key enhancement parameters into the particle position vector in the PSO algorithm, including:

[0115] Contrast enhancement coefficient (α): value range [0.8-1.5]

[0116] Brightness adjustment value (β): value range [-10, 10]

[0117] Sharpening strength (γ): value range [0-2.0]

[0118] Noise suppression threshold (δ): value range [0-5.0]

[0119] Tissue-specific enhancement weight (ω): range [0-1.0]

[0120] (2) Fitness function design: Evaluate the quality of parameter combinations based on multi-objective fitness functions:

[0121] F(p) = w1×SNR(p) + w2×Contrast Index(p) + w3×Edge Definition(p) –w4×Information Entropy Change(p)

[0122] Among them, p represents the parameter combination, w1, w2, w3, and w4 are weight coefficients;

[0123] Specifically, the signal-to-noise ratio (p) = f(α, δ, ω) = α×(original signal strength) / (δ×(1-ω)×(noise level)); the contrast index (p) = g(α, β) = α×(maximum grayscale value - minimum grayscale value) + β; the edge clarity (p) = h(γ, α) = γ×α×(edge ​​gradient strength); and the information entropy change (p) = j(α, β, δ, ω) = |original entropy - processed entropy|.

[0124] (3) Search process:

[0125] (3.1) Initialize 20-30 random particles (parameter combinations),

[0126] (3.2) Perform parameter combination iteration. In each iteration, the particle updates its speed and position based on its own optimal position and the global optimal position.

[0127] (3.3) Use inertia weight strategy to accelerate convergence;

[0128] (3.4) Iterate 25-50 times or until the fitness change is less than the preset threshold.

[0129] (4) Implementation of adaptive mechanism:

[0130] The weights in the fitness function are automatically adjusted according to the histogram characteristics of the input CT image:

[0131] (4.1) For images with large noise, increase the weight of the noise suppression threshold;

[0132] (4.2) For images with lower contrast, increase the weight of the contrast enhancement coefficient.

[0133] Taking the horizontal abdominal CT image above as an example, the formatted image still suffers from insufficient contrast between muscle and fat tissue. After applying the adaptive PSO algorithm in this embodiment, the system determined the optimal parameter combination: α = 1.25 (contrast enhancement), β = 3 (brightness adjustment), γ = 1.2 (sharpening intensity), δ = 2.5 (noise suppression), and ω = 0.7, 0.8, 0.5.

[0134] The enhanced results are as follows: the CT value distribution of muscle tissue is more concentrated, and the contrast with surrounding tissues is increased by about 27%. The boundary of adipose tissue is clearer. While the details of bone structure are preserved, artifacts are suppressed, and the overall image signal-to-noise ratio is increased by about 18%. The comparison before and after image enhancement shows that the image after adaptive enhancement significantly improves the distinguishability between different tissues while maintaining the authenticity of anatomical structures, providing a better input for subsequent automatic segmentation.

[0135] In the image edge enhancement stage, the image and processing module adopt a multi-scale adaptive Gaussian filter combination strategy to perform edge enhancement on CT images, specifically as follows:

[0136] (1)Selection of density-related adaptive kernel parameters:

[0137] (1.1)Dynamically adjust the Gaussian kernel parameters according to the HU value characteristics of different tissues in the CT image:

[0138] For high-density regions (such as bone, HU>150), a smaller σ value (0.5 - 1.0) is used to retain fine structures;

[0139] For medium-density regions (such as muscle, -29 < HU < 150), a medium σ value (1.0 - 2.0) is used;

[0140] For low-density regions (such as adipose, -190 < HU < -30), a larger σ value (2.0 - 3.0) is used to reduce the influence of noise;

[0141] (2)Use Gaussian filters with 3 different scales (σ values) in each direction respectively, and fuse the filtering results in each direction through the tensor voting mechanism to enhance the edges consistent with the tissue anatomical orientation;

[0142] (3)Edge enhancement based on Laplacian difference:

[0143] (3.1)Calculate the Laplacian operator (LoG) response of the image after Gaussian filtering at different scales;

[0144] (3.2)Nonlinearly weight and combine the original image and the LoG response:

[0145] I_enhanced = I_original + λ×f(LoG(I_original))

[0146] Among them, λ is the adaptive enhancement coefficient, and f() is a nonlinear mapping function, which dynamically adjusts the enhancement degree according to the edge intensity.

[0147] Taking the above-mentioned horizontal abdominal CT image as an example, after applying the edge enhancement method of this embodiment, the edge clarity of the muscle-fat interface is improved by about 35%, and the edge clarity of the bone-muscle interface is improved by about 40%. At the same time, the contours of the internal organs are also significantly enhanced. Compared with conventional Gaussian filtering, the method of this embodiment maintains a significant increase in the noise level while enhancing the edge (the signal-to-noise ratio decreases by no more than 5%), and edge sharpening does not lead to a significant change in tissue density values ​​(the average HU value changes by <3HU), making it more suitable for subsequent precise segmentation tasks. This Gaussian filtering edge enhancement method, which is specifically designed for CT image characteristics, fully considers the density characteristics and anatomical structure characteristics of different tissues, and is significantly superior to the standard Gaussian filtering method in general image processing, laying a solid foundation for subsequent tissue segmentation.

[0148] Figure 2 and Figure 3 The comparison pictures before and after image preprocessing show that the preprocessed image retains the original anatomical structure information, while significantly improving the contrast and boundary clarity between different tissues, laying the foundation for subsequent automatic segmentation.

[0149] After the image preprocessing is completed, this embodiment further performs segmentation processing on the preprocessed image through the image segmentation module. The specific process of the segmentation processing is as follows:

[0150] (1) The preprocessed 512×512 CT image is fed into the UNet architecture;

[0151] (2) The HViTER module in the encoder extracts features at multiple levels;

[0152] (3) The first layer HViTER module (64 channels) extracts shallow edge and texture features;

[0153] (4) The second layer HViTER module (128 channels) extracts the morphological features of the middle layer tissue;

[0154] (5) Enhance feature representation capabilities through residual feature fusion and global feature fusion;

[0155] (6) The decoder gradually upsamples through transposed convolution and combines the feature maps of skip connections;

[0156] (7) The final output is a 3-channel probability map (muscle, fat, bone).

[0157] During the image segmentation processing of this embodiment, after applying residual feature fusion, the boundaries of muscle and fat are more accurately represented. Compared with the traditional UNet, the blurred boundary area is reduced by about 32%; global feature fusion enables the model to consider the overall anatomical structure, reducing misclassification due to local similarities, and improving classification accuracy by about 3.5%; in the identification of muscle-fat mixed areas and small-area muscle bundles, the accuracy rates are improved by 5.2% and 6.8%, respectively.

[0158] The results of the segmentation processing are as follows: the abdominal CT image is segmented into muscle tissue (red marker), subcutaneous fat (blue marker), visceral fat (green marker), and bone (yellow marker); the boundary of the segmentation result is smooth and conforms to the anatomical characteristics, with Dice coefficients reaching 0.958 (muscle), 0.946 (subcutaneous fat), 0.937 (visceral fat), and 0.975 (bone). Compared with the manual segmentation results of doctors, the overlap rate reached 93.5%, and the average boundary difference was less than 2 pixels.

[0159] Example 3, component statistics module.

[0160] This embodiment provides a component statistics module that can be used in the body composition analysis and comparison system of the present invention. In order to enable those skilled in the art to better understand and implement the technical solution of the present invention, this embodiment provides the overall structure and specific statistical methods of a component statistics module.

[0161] The specific steps of the pixel statistics module performing pixel statistics include:

[0162] (1) Classification labeling: assigning different category labels to each pixel in the segmented medical image;

[0163] (2) Color mapping: Use different colors to mark different categories of areas;

[0164] (3) Pixel counting: Traverse all pixels in the segmented medical image and count the number of pixels corresponding to each category:

[0165]

[0166] in, Representation category The total number of pixels; Representing coordinates Whether the pixel at belongs to the category , if yes, take 1, otherwise take 0; and are the height and width of the image respectively;

[0167] (4) Proportion calculation:

[0168] Count the pixel ratios of different categories:

[0169]

[0170] in, For category Scale in the image; is the total number of pixels in the entire image.

[0171] According to the above method, a specific implementation method for pixel extraction of the CT image processed in Example 2 is as follows:

[0172] 1. Identification of total tissue volume:

[0173] (1) Distinguishing human tissue from background based on HU value range;

[0174] (2) Apply a threshold (-250HU to +1500HU) to filter all human tissue pixels;

[0175] (3) Calculate the total number of pixels that meet the conditions as a reference benchmark.

[0176] 2. Classification statistics of segmentation results:

[0177] (1) A multi-channel labeled map (one channel for each body component) generated using a segmentation model;

[0178] (2) Binarize each channel (pixel values ​​> 0.5 are considered to be of this category);

[0179] (3) Count the total number of pixels in each channel and their coordinate distribution;

[0180] (4) Handling overlapping areas: Apply priority rules to resolve classification conflicts.

[0181] 3. HU value comprehensive correction:

[0182] (1) Secondary confirmation of the specific HU value range based on existing anatomical knowledge;

[0183] ①Muscle tissue: -29HU to +150HU

[0184] ②Subcutaneous fat: -190HU to -30HU

[0185] ③Visceral fat: -175HU to -25HU

[0186] (2) Correct the misclassified pixels that do not conform to the specific tissue HU value range.

[0187] 4. Calculation of proportional relationship:

[0188] (1) Calculate the percentage of each body component pixel to the total tissue pixels;

[0189] (2) Calculate the ratio of muscle to total fat (subcutaneous + visceral);

[0190] (3) Calculate the ratio of subcutaneous fat to visceral fat.

[0191] Statistical results:

[0192] Taking the above horizontal abdominal CT image as an example, the pixel statistics after preprocessing and segmentation are as follows:

[0193]

[0194] Key ratio indicators:

[0195] Muscle / total fat ratio: 0.78

[0196] Subcutaneous fat / visceral fat ratio: 1.59

[0197] Muscle mass index (SMI): 54.2 cm² / m² (assuming the patient is 1.68m tall)

[0198] The images and results are shown below:

[0199] In the color-coded segmentation result image:

[0200] Red area: muscle tissue (24,682 pixels)

[0201] Blue area: subcutaneous fat (19,428 pixels)

[0202] Green area: visceral fat (12,176 pixels)

[0203] The distribution characteristics of different tissues can be clearly observed in the image:

[0204] Muscle tissues such as the psoas and erector spinae are located around the spine;

[0205] Subcutaneous fat forms a continuous annular band located on the outer side of the abdominal wall;

[0206] Visceral fat is mainly distributed around the organs in the abdominal cavity;

[0207] Clinical significance analysis:

[0208] The patient's SMI value was 54.2 cm² / m², which is higher than the diagnostic threshold for sarcopenia (<55 cm² / m² for men and <39 cm² / m² for women);

[0209] The visceral fat / subcutaneous fat ratio was 0.63, which is below the high-risk threshold for metabolic syndrome (>1.0);

[0210] The average muscle HU value was 56.8, which was above the threshold for muscle mass loss (<30).

[0211] Then, the corresponding results are calculated according to the methods of the pixel area conversion module and the density evaluation module.

[0212] Example 4, analysis and comparison module.

[0213] This embodiment provides an analysis and comparison module of the present invention. The analysis and comparison module includes a change calculation module, a visualization generation module, and a report generation module; the change calculation module includes a volume change calculation module, a HU value change calculation module, and a clinical significance judgment module, and the clinical significance judgment module judges the significance of the change based on a preset threshold value of the volume change calculation module and the HU value change calculation module; the visualization generation module uses a heat map and / or a difference map for visualization; and the report generation module generates an analysis report based on the results of the change calculation module and the visualization generation module, as follows:

[0214] 1. Change Quantification

[0215] (1) Calculation of volume change

[0216] For the volume of tissue c at two different time points (baseline and follow-up), calculate its change and rate of change :

[0217]

[0218]

[0219] in: is the tissue volume at baseline (initial scan), is the tissue volume at follow-up.

[0220] Clinical interpretation:

[0221] Will and preset thresholds For comparison,

[0222] like , the system can issue prompts of abnormal growth (such as edema, tumor proliferation).

[0223] like , the system can issue a reminder of tissue atrophy (such as muscle loss and fat loss).

[0224] (2) Calculation of HU value change

[0225] The HU value in medical imaging (CT) reflects tissue density. For example, a higher HU value for muscle tissue indicates greater density and better muscle quality. Therefore, this example calculates:

[0226] HU mean change:

[0227]

[0228] HU change rate:

[0229]

[0230] in, Indicates the HU value at follow-up (reexamination), represents the HU value at baseline (initial scan), Indicates the change in HU value.

[0231] Clinical interpretation:

[0232] Decreased muscle HU : The system can issue reminders of muscle degeneration and fat infiltration.

[0233] Increased fat HU: This can send a signal through the system that may be related to inflammation.

[0234] (3) Judgment of clinical significance

[0235] The clinical significance of the changes is determined by pre-defined thresholds, such as:

[0236] Muscle atrophy:

[0237] Decreased bone density:

[0238] Abnormal fat changes:

[0239] 2. Change Visualization

[0240] In order to intuitively present the changes in tissues such as muscle, subcutaneous fat, and visceral fat, this embodiment uses heatmaps and difference maps for visualization.

[0241] (1) Heatmap

[0242] Objective: To visually display the volume and density changes of different tissues.

[0243] Implementation:

[0244] Red (decreased): Indicates a decrease in the volume or HU value of the area (e.g., muscle atrophy).

[0245] Blue (increase): indicates that the volume or HU value of the area has increased (such as tumor proliferation).

[0246] Green (no significant change): indicates that the change in the area is within the normal range.

[0247] Example:

[0248] Visceral fat reduction: bony areas appear red.

[0249] Muscle atrophy: The muscle area appears red and the HU value decreases.

[0250] Subcutaneous fat hyperplasia: Fatty areas appear blue.

[0251] (2) Difference Map

[0252] Objective: To generate a difference map by calculating the pixel-level difference between two scans, highlighting the changed areas. Calculation method:

[0253]

[0254] in, represents the pixel value at the follow-up scan, represents the pixel value at the baseline, is the difference between the two;

[0255] like , the volume or density of the region increases (marked in blue).

[0256] like , the volume or density of the region decreases (marked in red).

[0257] like , there is no obvious change in this area.

[0258] Visualization: Overlays the original CT image to highlight the changed areas. Generates a difference map with adjustable transparency to facilitate clinicians to observe subtle changes.

[0259] 3. Report Generation

[0260] Figure 4 This is a preliminary report of the results obtained according to various embodiments of the present invention, combining the automatic delineation algorithm to obtain report information and suggestions.

[0261] exist Figure 4 Based on the manual modification of the area, the muscles, subcutaneous fat and visceral fat were corrected for the wrong outlines, such as Figure 5 shown.

[0262] After correction, regenerate the area and diagnostic opinions, such as Figure 6 shown.

[0263] The SMI and SMRA in the report are calculated as follows: Based on the skeletal muscle area of ​​the muscle obtained from the original CT image and the uploaded height, the SMI value is calculated using the following formula:

[0264]

[0265] Among them, SMI (Skeletal Muscle Index, cm² / m²) represents the skeletal muscle index, and SMA (Skeletal Muscle Area, cm²) represents the skeletal muscle area, which is generally calculated by CT images (commonly used as Figure 2 The cross-sectional area of ​​the body at the position shown is Figure 2 ), Height represents height in meters (m).

[0266] SMRA is calculated by the following formula:

[0267]

[0268] Among them, the numerator represents the sum of all HU values ​​in the muscle area, and the denominator represents the total number of pixels in the muscle area (that is, the area corresponding to the muscle area).

[0269] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A body composition analysis and comparison system based on medical images, characterized in that: Including component statistics module and analysis and comparison module, The component statistics module is configured to perform preliminary statistics on data related to body composition based on the segmented medical image; The component statistics module includes: The pixel statistics module is configured to determine the proportional relationship of different body components by counting the number of pixels of different categories in the segmented medical image. The area measurement module is configured to convert the number of pixels into the real physical area based on pixel statistics. a density assessment module configured to calculate an average grayscale value of the segmented region based on pixel statistics and area measurement, and estimate physical properties of body tissue corresponding to the medical image; The analysis and comparison module is configured to calculate the change amount, visualize the change area, and generate an analysis and comparison report based on the different data obtained by the component statistics module.

2. The body composition analysis and comparison system based on medical images according to claim 1, characterized in that: The specific steps of the pixel statistics module performing pixel statistics include: (1) Classification labeling: assigning different category labels to each pixel in the segmented medical image; (2) Color mapping: Use different colors to mark different categories of areas; (3) Pixel counting: Traverse all pixels in the segmented medical image and count the number of pixels corresponding to each category: ， in, Representation category The total number of pixels; Representing coordinates Whether the pixel at belongs to the category , if yes, take 1, otherwise take 0; and are the height and width of the image respectively; (4) Proportion calculation: Count the pixel ratios of different categories: ， in, For category Scale in the image; is the total number of pixels in the entire image.

3. The body composition analysis and comparison system based on medical images according to claim 1, characterized in that: The specific steps of the area measurement module for pixel area conversion are: Set the pixel spacing of medical images to , unit: mm², then the area of ​​each category is: ， in, for The area of ​​the category in mm², and are the pixel pitches in the horizontal and vertical directions, respectively.

4. The body composition analysis and comparison system based on medical images according to claim 1, characterized in that: The estimation steps of the density assessment module specifically include: (1) Obtain the grayscale value of each coordinate position in the medical image and calculate the average grayscale density as follows: ， in, For category The average grayscale density, unit HU, For coordinates Gray value at ; (2) According to the grayscale values ​​at different coordinate positions, the standard deviation of tissue density at the corresponding parts is calculated as follows: ， (3) Determine body composition based on the grayscale density and tissue density standard deviation calculated in the first two steps.

5. The body composition analysis and comparison system based on medical images according to claim 1, characterized in that: The analysis and comparison module includes a change calculation module, a visualization generation module, and a report generation module; the change calculation module includes a volume change calculation module, a HU value change calculation module, and a clinical significance judgment module; the clinical significance judgment module judges the significance of the change based on a preset threshold value of the volume change calculation module and the HU value change calculation module; the visualization generation module uses a heat map and / or difference map for visual display; and the report generation module generates an analysis report based on the results of the change calculation module and the visualization generation module.

6. The body composition analysis and comparison system based on medical images according to claim 5, characterized in that: The specific calculation method of the variation calculation module is: For the volume of tissue c at two different time points, calculate its change and rate of change : ， ， in, is the tissue area at baseline; is the tissue area at follow-up; according to Compare the size of the data and push corresponding prompts, where Threshold increase Indicates the preset threshold.

7. The body composition analysis and comparison system based on medical images according to claim 5, characterized in that: The specific calculation contents of the HU value change calculation module include: HU mean change: ， HU change rate: ， in, represents the HU value at follow-up, represents the HU value at baseline, Indicates the rate of change of HU value.

8. The body composition analysis and comparison system based on medical images according to claim 5, characterized in that: The visualization generation module generates a difference map by calculating the pixel-level difference between the two scans, highlighting the changed areas. The specific calculation method is as follows: ， in, represents the pixel value at the follow-up scan, represents the pixel value at the baseline, is the difference between the two; like , the volume or density of the area increases, marked in blue, like , the volume or density of the area decreases and is marked in red. like , then there is no obvious change in the area and no color is marked.

9. The body composition analysis and comparison system based on medical images according to claim 5, characterized in that: The algorithm steps performed by the report generation module include: (1) Obtain report information and suggestions based on the automatic delineation algorithm; and make adjustments based on manual modification input to regain the area and opinions corresponding to the manual modification; (2) Calculation of skeletal muscle index: ， Among them, SMI represents skeletal muscle index, unit is cm² / m², SMA represents skeletal muscle area, unit is cm², calculated by CT images; (3) Calculate the SMRA value: ， The numerator represents the sum of all HU values ​​in the muscle area, and the denominator represents the total number of pixels in the muscle area, that is, the total pixel area.

Images