A ct medical image processing identification method and system for sarcopenia

Abstract

The application provides a CT medical image processing and recognition method and system for sarcopenia, the method comprising: acquiring a CT image, and performing image preprocessing; establishing a fitness function of a task-driven closed loop according to a sarcopenia pathological image feature enhancement parameter, and performing sarcopenia pathological feature enhancement by using a self-adaptive particle swarm optimization algorithm; performing edge enhancement on the self-adaptively enhanced image by using a multi-scale self-adaptive Gaussian filter combination, performing pixel-level segmentation by using a segmentation model, and recognizing and locating the image regions corresponding to skeletal muscle, subcutaneous fat and visceral fat. The application combines a traditional particle swarm with a self-adaptive optimization strategy for sarcopenia features, can automatically adjust the enhancement strategy according to the image contrast and noise level, intelligently strengthens the muscle edge according to the tissue density by cooperating with the multi-scale self-adaptive Gaussian filter, clearly distinguishes the skeletal muscle from the fat and the viscera, and improves the problem of non-uniform image quality under the condition of poor original image acquisition.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and recognition technology, and in particular to a method for processing and recognizing CT medical images for sarcopenia. Background Technology

[0002] Sarcopenia is a systemic disease characterized by a progressive decline in skeletal muscle mass and strength, accompanied by a decline in physiological function. It is closely related to aging, chronic diseases, and postoperative rehabilitation after malignant tumor surgery. It often occurs concurrently with osteoporotic fractures and motor dysfunction, seriously affecting the quality of life of the elderly and patients with chronic diseases. It has become an important concern in clinical rehabilitation, geriatrics, and tumor prognosis assessment. The mainstream techniques for sarcopenia assessment currently include dual-energy X-ray absorptiometry (DXA), bioelectrical impedance analysis (BIA), CT, MRI, and anthropometry. Among these, CT imaging is widely recognized as the most accurate and objective method for sarcopenia assessment because it can accurately obtain information on the cross-sectional area, density, and structure of muscle in the human body. Clinically, transverse muscle indices at the third lumbar vertebra (L3) level are often used as the core assessment basis.

[0003] Traditional CT-based sarcopenia assessment relies heavily on manual operation: medical staff manually locate the L3 vertebral body center level, manually delineate muscle region boundaries, measure muscle area, and calculate skeletal muscle index, etc. The entire process is cumbersome, time-consuming, and affected by physician experience and differences in image interpretation, making it difficult to guarantee assessment consistency and efficiency. While there are reports of automated segmentation in existing technologies, technical bottlenecks still exist. For example, Chinese patent document CN113327235A discloses an image processing method for sarcopenia assessment based on CT images. Although it achieves automated L3 localization and sarcopenia grading, it only uses 2D Unet for single muscle region segmentation, which is prone to boundary missegmentation and the assessment results are not representative enough. Another example is Chinese patent document CN117522857A, which discloses a method for subject component analysis, label generation, and model training. It adopts a two-level strategy of coarse and fine segmentation, first distinguishing between fat and non-fat regions, and then further segmenting subcutaneous fat, visceral fat, psoas major muscle, and other abdominal muscles to achieve quantitative analysis of three-dimensional components of the entire abdomen. This method is mainly used to distinguish between fat and non-fat regions and to identify skeletal muscle components in sarcopenia images, but it is not specifically designed for sarcopenia assessment, does not automatically locate the L3 standard level, cannot directly calculate the skeletal muscle index (SMI), and cannot automatically determine the sarcopenia grade. Furthermore, the two-level segmentation process is cumbersome and computationally intensive, making it difficult to meet the needs of rapid clinical assessment.

[0004] Therefore, there is a need for a CT medical image processing and recognition method and system for sarcopenia that can accurately distinguish muscles (especially skeletal muscle) from other tissues (especially fat) while balancing segmentation accuracy and computational efficiency. Summary of the Invention

[0005] The present invention aims to solve, to a certain extent, one of the technical problems existing in the background art.

[0006] Therefore, the first objective of this invention is to provide a CT medical image processing and recognition method for sarcopenia, in order to solve the problems of low tissue differentiation, image enhancement easily destroying the quantitative consistency of CT values, and difficulty in achieving the required assessment accuracy in the process of improving CT image discrimination of sarcopenia.

[0007] The second objective of this invention is to provide a CT medical image processing and recognition system for sarcopenia.

[0008] A third objective of this invention is to provide a computer device.

[0009] A fourth object of the present invention is to provide a computer-readable storage medium.

[0010] To achieve the above objectives, the first aspect of this invention proposes a CT medical image processing and recognition method for sarcopenia. The method includes: acquiring CT images of the lumbar or abdominal region of patients with sarcopenia or suspected sarcopenia; performing image preprocessing operations on the CT images, including format conversion, size standardization, CT value normalization, and denoising; configuring contrast, brightness, sharpening intensity, noise suppression, and tissue-specific enhancement parameters based on the pathological image characteristics of sarcopenia; establishing a fitness function with a task-driven closed loop including tissue separability, key boundary clarity, segmentation stability or uncertainty, HU quantitative consistency penalty, and noise amplification penalty; using an adaptive particle swarm optimization algorithm to enhance the pathological features of sarcopenia in the formatted CT images; using a multi-scale adaptive Gaussian filter combination based on density-dynamically adjusted kernel parameters to perform regionally differentiated edge enhancement on the adaptively enhanced images, strengthening the boundary contours of skeletal muscle, fat, and visceral fat; and using a UNet segmentation model combining HViTER feature extraction and residual feature fusion to perform pixel-level segmentation on the edge-enhanced images, identifying and locating the image regions corresponding to skeletal muscle, subcutaneous fat, and visceral fat.

[0011] In one embodiment of the invention, the tissue separability term J in the fitness function sep Quantify the degree of grayscale separation of different tissues; key boundary sharpness term G edge Quantify the edge identifiability of critical boundaries; segment the stability or uncertainty term U. seg Suppression of structural misclassification due to enhancement; HU quantitative consistency penalty term P HU HU statistics of the organization before and after constraint enhancement; noise amplification penalty term N ampSuppress noise amplification caused by sharpening or contrast enhancement; the above feature terms participate in the operation of the fitness function F(x) according to the weight ratio.

[0012] In one embodiment of the present invention, the tissue separability term J sep The ratio of inter-class variance to intra-class dispersion among different tissue pairs is selected, i.e., J sep = (μ m - μ f ) 2 / (σ m 2 + σ f 2 + ε), where μ m σ m The mean HU value and standard deviation of a tissue region are given by μ. f σ f , respectively, represent the mean HU value and standard deviation of another tissue region, and ε is a minimal constant to prevent the denominator from being zero.

[0013] In one embodiment of the present invention, the tissue separability term J sep We selected a weighted sum of the ratios of inter-class differences to intra-class dispersion of multiple tissue pairs.

[0014] In one embodiment of the present invention, the critical boundary sharpness term G edge Calculate the gradient energy within the boundary band B obtained from the coarse segmentation or pre-masking, i.e., , among which, I enh To enhance the image, For gradient operators, ( i,j () represents the coordinates of any point.

[0015] In one embodiment of the present invention, the segmentation stability or uncertainty term U seg This refers to inputting the enhanced image into a lightweight coarse segmenter to obtain a tissue region probability map P(k|i,j), and calculating the average entropy as the uncertainty, i.e., Where Ω represents the human body region or ROI, ( i,j () is the coordinate of any point. k Indicates the area of ​​organization.

[0016] In one embodiment of the present invention, the HU quantitative consistency penalty term P HU Calculate according to the following formula: Where T represents the set of organizations, μ t (after) and μ t(before) represents the average HU statistic of organization t before and after enhancement.

[0017] In one embodiment of the present invention, the fitness function employs a task-driven closed-loop optimization strategy, which includes: initializing the particle swarm, randomly generating several particles x, and setting a velocity vector; performing enhancement on each particle parameter x to obtain an enhanced image I. enh ; Enhanced image I enh The coarse segmenter is used to obtain the tissue region probability map P(k|i,j); the tissue separability term J is calculated. sep Key boundary clarity item G edge , segmentation of stability or uncertainty term U seg HU quantitative consistency penalty item P HU Noise amplification penalty term N amp The fitness function F(x) is obtained based on the weight ratio; the individual optimal pbest and the global optimal gbest are updated; the particle velocity and position are updated based on pbest and gbest, and parameters that exceed the boundary are truncated or bounced back to the legal range; the process terminates early when the maximum number of iterations is reached or the improvement of the global optimal fitness is less than the threshold.

[0018] In one embodiment of the present invention, the adaptive sarcopenia pathological feature enhancement dynamically adjusts the closed-loop optimization strategy based on the statistical characteristics of the input CT medical images, including: if the noise level is high, increasing the HU quantitative consistency penalty term P. HU and noise amplification penalty term N amp The weights of the components are adjusted, and the upper limit of the sharpening intensity is limited. If the tissue contrast is low, the weights of the tissue separability and key boundary sharpness are increased, and the effective range of the contrast enhancement coefficient α is widened. If the coarse segmentation uncertainty is high, a parameter combination that can reduce the entropy value is selected to suppress the over-enhancement strategy that leads to missegmentation.

[0019] In one embodiment of the present invention, the adaptive Gaussian filter forms a σ map by combining segmented range and continuous mapping, and dynamically adjusts the Gaussian kernel parameter σ according to tissue density, including: using a smaller σ value for high-density regions; using a medium σ value for medium-density regions; and using a larger σ value for low-density regions.

[0020] In one embodiment of the present invention, the adaptive Gaussian filter uses multiple Gaussian filter responses of different scales in each direction, introduces direction consistency weighting, and enhances edges that are consistent with the anatomical orientation of the tissue and have stable cross-scale responses; the Laplacian operator responses of the Gaussian filtered images at different scales are calculated, and a nonlinear mapping is used to weight and superimpose the results with the adaptive coefficient λ to obtain the enhancement result: I enh = Ioriginal +λ×f(LoG(I original ), where λ is the adaptive enhancement coefficient, I original To enhance the image, LoG represents the Laplacian Gaussian operator.

[0021] To achieve the above objectives, a second aspect of the present invention proposes a CT medical image processing and recognition system for sarcopenia. The system includes: a medical image acquisition module for acquiring CT medical images of patients with sarcopenia or suspected sarcopenia; an adaptive enhancement module for configuring enhancement parameters specific to sarcopenia characteristics, establishing a task-driven closed-loop fitness function, and using a task-driven closed-loop adaptive particle swarm optimization algorithm to adaptively enhance the formatted CT image, thereby improving the contrast between skeletal muscle, subcutaneous fat, and visceral fat; an edge enhancement module for performing edge enhancement processing on the adaptively enhanced image using a combination of multi-scale adaptive Gaussian filters; and an image segmentation module for segmenting the edge-enhanced CT image using residual feature fusion to identify the image regions corresponding to skeletal muscle, subcutaneous fat, and visceral fat. The system is configured to implement the method described in the first aspect.

[0022] To achieve the above objectives, a third aspect of the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in the first aspect.

[0023] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0024] The beneficial effects of this invention are: This invention constructs a task-driven closed-loop fitness function for sarcopenia, uniformly optimizing tissue separability, boundary clarity, segmentation stability, CT value quantitative consistency, and noise suppression. While enhancing sarcopenia-related pathological features, it strictly maintains unchanged CT values, solving the problem that traditional enhancement algorithms easily violate clinical quantitative standards and improving the reliability and reproducibility of sarcopenia assessment. This invention combines traditional particle swarm optimization with an adaptive optimization strategy targeting sarcopenia features. It can automatically adjust the enhancement strategy based on image contrast and noise level, and with multi-scale adaptive Gaussian filtering, intelligently enhances muscle edges according to tissue density, clearly distinguishing skeletal muscle from fat and internal organs. This effectively improves the problem of inconsistent image quality in scenarios with different image acquisition environments and age groups.

[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method provided in Implementation Example 1.

[0027] Figure 2 This is the original image of the CT image obtained as provided in Example 1.

[0028] Figure 3 The image provided in Example 1 is a preprocessed, formatted CT image.

[0029] Figure 4 The image provided in Example 1 is an adaptively enhanced CT image.

[0030] Figure 5 This is the CT image with edge enhancement provided in Example 1.

[0031] Figure 6 This is the result image output after image segmentation provided in Example 1.

[0032] Figure 7 This is a schematic diagram of the system structure provided in Embodiment 2.

[0033] Figure 8 This is a structural block diagram of the computer device provided in Embodiment 3. Detailed Implementation

[0034] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Furthermore, based on the embodiments of the present invention, any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art to all other embodiments obtained without creative effort should be included within the protection scope of the present invention.

[0035] Sarcopenia, also known as muscle wasting syndrome, is a systemic degenerative disease characterized by the continuous loss of skeletal muscle mass, decreased muscle strength, and reduced muscle function. Its pathogenesis is related to multiple factors, including age-related hormonal changes, chronic low-grade inflammation, malnutrition, lack of exercise, and neuromuscular degeneration. It has a very high incidence in the elderly, patients with chronic kidney disease, chronic obstructive pulmonary disease, heart failure, and malignant tumors. Sarcopenia can be classified into four levels according to its course and severity: normal, pre-sarcopenia, sarcopenia, and severe sarcopenia. Its core pathological changes are manifested in skeletal muscle atrophy, reduced muscle fiber volume, increased intramuscular fat infiltration, decreased muscle density, blurred muscle boundaries, and a reduction in muscle functional units. These changes directly lead to decreased motor ability, weakened balance, and an increased risk of falls and fractures. They also significantly reduce patients' tolerance to surgery and radiotherapy / chemotherapy, making them important risk factors affecting long-term prognosis and quality of life. In clinical practice, the diagnosis of sarcopenia relies heavily on quantitative assessment of skeletal muscle. Among them, the transverse CT image of the third lumbar vertebra is the standard imaging level for sarcopenia assessment because it can fully display the core muscle groups such as the psoas major, erector spinae, transverse abdominis, and rectus abdominis. The assessment indicators mainly include skeletal muscle cross-sectional area, muscle density, and skeletal muscle index. These indicators all need to be accurately measured on the basis of clear, stable, and high-fidelity images.

[0036] Before providing a detailed description of the specific implementation methods of this invention, in order to enable those skilled in the art to better understand the contributions of this invention, it is necessary to explain the essential differences between sarcopenia CT images and other medical images such as conventional abdominal CT images, tumor images, fatty liver images, and obesity images, so as to highlight the specificity and necessity of this invention in the field of sarcopenia.

[0037] In CT images, different tissues exhibit significantly different grayscale, density, texture, and boundary features. These differences are the core basis for image analysis and tissue segmentation. Skeletal muscle appears as a medium-density tissue in CT images, with a HU value typically stable between 30 and 60. It has a uniform grayscale distribution, a dense and continuous bundle-like texture, smooth and clear edges, and a well-defined anatomical morphology. In patients with sarcopenia, skeletal muscle shows overall reduced grayscale, sparse texture, widened gaps, blurred boundaries, and is accompanied by scattered low-grayscale fat infiltration. Adipose tissue appears as a typical low-density tissue in CT images. The HU values ​​of subcutaneous fat and visceral fat are mostly between -190 and -30, with significantly lower grayscale values ​​than skeletal muscle. Its texture is loose and uniform, without a fixed shape, and its boundaries are relatively soft. There is a natural grayscale difference between fat and skeletal muscle. Compared to CT images of healthy individuals, the most prominent features of sarcopenia patients on CT images are a significant reduction in skeletal muscle cross-sectional area, muscle atrophy, low muscle density, and marked fat infiltration. Muscle margins are blurred, intermuscular spaces are enlarged with fat infiltration, and the contrast between muscle and surrounding fat is extremely low. These changes are often accompanied by osteoporosis and ectopic visceral fat deposition, among other complex alterations. These subtle changes are difficult to discern directly in conventional CT images, which is a key reason why conventional body composition analysis methods cannot accurately identify these changes. For example, CT images of obesity-related diseases primarily focus on increased fat volume, differentiating between fat and non-fat components and the distribution of fat bodies. Muscle is typically shown as normal and not a primary focus. In contrast, sarcopenia is characterized by muscle atrophy and decreased density; the pathological changes are completely opposite. Conventional fat segmentation algorithms are highly prone to misclassifying intramuscular fat infiltration in sarcopenia patients as normal tissue. Compared to images of abdominal lesions such as fatty liver and abdominal tumors, which primarily focus on the solid structure of organs and do not involve fine textural changes such as skeletal muscle volume, intermuscular spaces, muscle density, and muscle boundaries, sarcopenia assessment requires precise capture of subtle pathological features such as skeletal muscle atrophy, blurred boundaries, and fatty infiltration. The precision and fidelity requirements for image enhancement and segmentation are far higher than in conventional organ imaging analysis. Compared to images of patients after tumor surgery or with cachexia, sarcopenia involves progressive degenerative changes, with muscle atrophy exhibiting symmetrical, diffuse, and uniform decline rather than localized defects or destruction. Therefore, image processing must preserve the overall muscle morphology and grayscale distribution patterns, avoiding excessive enhancement that could distort the structure. Thus, sarcopenia CT image recognition needs to focus on key features such as skeletal muscle area ratio, muscle boundary clarity, muscle grayscale uniformity, degree of intermuscular fat infiltration, HU value differentiation between muscle and adipose tissue, and bilateral muscle symmetry.From an image analysis perspective, muscles have a stable HU range, uniform grayscale, directional texture, and regular anatomical structure; fat has extremely low HU value, non-directional texture, and variable shape; bone tissue has extremely high HU value and sharp edges; and internal organs have large grayscale fluctuations, irregular shapes, and unstable boundaries. The core difficulty in sarcopenia identification lies in the fact that atrophied muscles and infiltrated fat have similar grayscale, mixed textures, and blurred boundaries, making it difficult for conventional enhancement and segmentation methods to reliably distinguish them. This invention is based on the inherent differences of the above-mentioned tissues such as fat, muscle, and bone in CT images and the special imaging characteristics of sarcopenia, and constructs an image enhancement and segmentation system specifically for sarcopenia.

[0038] Example 1 This embodiment provides a CT medical image processing and recognition method for sarcopenia, as shown in the attached figure. Figures 1-5 The method includes steps S101 to S104.

[0039] Step S101, acquire the CT image to be processed ( Figure 2 Image preprocessing is performed to obtain a formatted image. Figure 3 ).

[0040] In this step, plain CT scan images of the waist or abdomen are used. The original CT images are in DICOM format with a resolution of 768×768 pixels and a slice thickness of 2.5mm. The CT values ​​are uneven and contain noise.

[0041] In the image preprocessing stage, regarding image format, the DICOM image format is taken as input and converted into an internal processing format. In this embodiment, the PNG format is used as an example. Regarding image size, the CT image is resampled to a resolution range of 512×512 to 1024×1024 pixels with a uniform interval of 1-5mm slice thickness. The CT value (Hounsfield Unit, HU) is remapped to a standard range (e.g., muscle tissue -29 to 150 HU, adipose tissue -190 to -30 HU, bone tissue >150 HU) through normalization processing.

[0042] In this embodiment, after image preprocessing, a standard size of 512×512 pixels is obtained, the slice thickness is reconstructed to a uniform interval of 2.0 mm, and the CT values ​​are remapped to a formatted CT image within the standard range.

[0043] Step S102: Configure enhancement parameters for sarcopenia characteristics, establish a task-driven closed-loop fitness function, and perform adaptive sarcopenia pathological feature enhancement on the formatted CT image.

[0044] In the adaptive image enhancement stage, an improved task-driven closed-loop adaptive particle swarm optimization (PSO) algorithm is used to enhance the formatted CT images. This differs from optimizations that only use general image quality metrics (such as contrast and sharpness) as optimization targets. This improved PSO algorithm incorporates the core requirements for sarcopenia image feature recognition (tissue separability, key boundary discernibility, segmentation stability, and HU quantitative consistency) into the optimization loop, making the enhancement parameters automatically selected by the system more suitable for subsequent body composition segmentation and longitudinal quantitative comparison tasks. The specific closed-loop adaptive PSO algorithm is as follows: (1) Configure enhancement parameters for sarcopenia characteristics, and encode the enhancement parameters as particle position vectors in the PSO algorithm, as follows: x = [α, β, γ, δ, ω], where: Contrast enhancement factor α: range [0.8, 1.5]; Brightness adjustment value β: range [-10, 10]; Sharpening intensity γ: Value range [0, 2.0]; Noise suppression threshold δ: value range [0, 5.0]; Tissue-specific enhancement weight ω: ranges from [0, 1.0].

[0045] In some implementations, ω is in vector form ω=[ω m , ω f , ω b These values ​​correspond to the enhancement weights for muscle, subcutaneous fat, and visceral fat, respectively. They are used to achieve differentiated enhancement of different tissues within the same enhancement framework, avoiding over-enhancement or under-enhancement of certain tissues caused by a single global enhancement parameter.

[0046] (2) Construct a task-driven closed-loop fitness function. The fitness function not only evaluates whether the image looks clearer, but also enables the targeted enhancement parameters to directly serve the segmentation and quantitative stability of CT images through a closed loop of "enhancement → coarse segmentation / pre-masking → task index calculation → feedback optimization".

[0047] In some implementations, the fitness function is: F = w1·J sep + w2·G edge - w3·U seg - w4·P HU - w5·N amp , The meanings of each item are as follows: Organizational divisibility term J sepThis is used to quantify the degree of grayscale separation between different tissues. The ratio of inter-class variation to intra-class dispersion is preferred, for example, using muscle and fat as an example: J sep = (μ m - μ f ) 2 / (σ m 2 + σ f 2 + ε), Where, μ m σ m The mean HU value and standard deviation for the muscle region are shown in μ. f σ f , , represent the mean HU value and standard deviation of the fat region, respectively, and ε is a minimal constant to prevent the denominator from being zero.

[0048] In some implementations, J sep It can be expanded to a weighted summation of multiple tissue pairs such as (muscle-subcutaneous fat, muscle-visceral fat, bone-muscle).

[0049] Key boundary clarity term G edge Used to quantify the edge discernibility of key boundaries (muscle-fat, bone-muscle, etc.).

[0050] In some implementations, a strip region B is constructed near the boundary obtained by coarse segmentation / pre-masking (e.g., a boundary band with width d is obtained by dilating the boundary), and the gradient energy or contrast enhancement within the enhanced boundary band is calculated, for example: , Among them, I enh To enhance the image, For gradient operators, ( i,j ) represents the coordinates of any point within the strip region B. This term can suppress situations where "the overall area becomes sharper but the boundaries remain blurry" or "oversharpening leads to false boundaries".

[0051] Segmentation stability / uncertainty term U seg Used to suppress structural misclassification caused by enhancement, making parameter selection more favorable for segmentation.

[0052] Preferably, the enhanced image is input into a lightweight coarse segmenter to obtain the probability map P(k|i,j) of each tissue, and the average entropy is calculated as the uncertainty: ; Where Ω represents the human body region or ROI, ( i,j () is the coordinate of any point. k U represents the number of organizational regions. segThe smaller the value, the more stable the segmentation and the more defined the boundary.

[0053] In some implementations, the lightweight coarse segmenter may be at least one of a low-resolution UNet, a simplified encoder, or a threshold pre-mask + morphological correction module. This lightweight coarse segmenter or pre-mask module is only used for closed-loop evaluation and optimization feedback, and does not require achieving the final segmentation accuracy. Its purpose is to quickly provide tissue regions and boundary bands for calculating task metrics, thereby enabling direct optimization of sarcopenia characteristic tasks through enhanced parameter selection.

[0054] HU Quantitative Consistency Penalty Item P HU This is used to constrain the relative stability of tissue HU statistics before and after enhancement, avoiding systematic shifts in body composition quantification caused by the enhancement process. The formula is as follows: ; Where T represents the set of organizations, μ t (after) and μ t (before) represents the mean HU statistic for organization t before and after enhancement. This term ensures the separability of the enhancement, rather than rewriting the HU.

[0055] In some embodiments, the tissue set is a predefined tissue set, including at least muscle and fat. In some embodiments, bone tissue may also be included. Generally, bone tissue has a significantly different HU value from muscle and fat, making the boundary easier to distinguish. Therefore, the core problem addressed by this invention is the relationship between skeletal muscle and fat, especially the distinction between muscle and fat in situations where there is a possibility of fat infiltration.

[0056] w1, w2, w3, w4, and w5 are the weights for each item.

[0057] (3) Closed-loop optimization process: (3.1) Initialize the particle swarm: randomly generate several particles x and set the velocity vector v; (3.2) Closed-loop evaluation: Enhancement is performed on each particle parameter x to obtain the enhanced image I. enh ; will I enh Input the lightweight coarse segmenter or pre-mask module described in section (2) to obtain the tissue region and probability map P(k|i,j); calculate the tissue separability term J. sep Key boundary clarity item G edge , segmentation of stability or uncertainty term U seg HU quantitative consistency penalty item P HU Noise amplification penalty term N amp The fitness function F(x) is obtained by combining the weight ratio; (3.3) Update the best: Update the individual best pbest and the global best gbest; (3.4) Update position: Update particle velocity and position according to pbest and gbest. Parameters that exceed the boundary are truncated or bounced back to the legal range. (3.5) Termination condition: Terminate early when the maximum number of iterations is reached or the improvement of the global optimal fitness is less than the threshold, so as to reduce invalid computation.

[0058] (4) Adaptive mechanism: The optimization strategy is dynamically adjusted based on the statistical characteristics of the input CT images to improve robustness across devices / scanning conditions. If the noise level is high, increase the HU quantitative consistency penalty term P. HU and noise amplification penalty term N amp The weights are determined, and the upper limit of sharpening intensity is limited; If the tissue contrast is low, increase the weights of the tissue separability and key boundary sharpness terms, and widen the effective range of the contrast enhancement coefficient α. If the uncertainty of coarse segmentation is high, then select a parameter combination that can reduce the entropy value and suppress over-enhancement strategies that lead to misclassification.

[0059] After applying the task-driven closed-loop adaptive PSO algorithm of this embodiment, the system output parameters can be: α=1.25, β=3, γ=1.2, δ=2.5, and the tissue-specific enhancement weights can be ω=[0.70, 0.80, 0.50] (corresponding to the enhancement weights of skeletal muscle, subcutaneous fat, and visceral fat, respectively). These parameters are not directly set by human experience, but are automatically optimized by the system with tissue separability, boundary clarity, segmentation stability, and HU quantitative consistency as joint objectives. This improves the identifiability of key tissue boundaries while reducing the systematic disturbance of enhancement to tissue HU statistics, making the enhancement results more suitable for subsequent quantitative analysis of body composition and longitudinal comparison.

[0060] The enhancement results show that the CT values ​​of muscle tissue are more concentrated and the contrast with surrounding tissues is improved by approximately 27%, the boundaries of adipose tissue are clearer, and the overall image signal-to-noise ratio is improved by approximately 18%. A comparison of images before and after enhancement shows that the image after adaptive enhancement (…) Figure 4 While maintaining the authenticity of anatomical structures, it significantly improves the distinguishability between different tissues, providing higher quality input for subsequent automatic segmentation.

[0061] Step S103: Use a combination of multi-scale adaptive Gaussian filters to perform edge enhancement processing on the adaptively enhanced image.

[0062] In the image edge enhancement stage, a multi-scale adaptive Gaussian filter combination strategy of "density-guided scale-space + directional consistency fusion" is adopted to enhance the edges of CT images. Since the HU values of different tissues in CT images reflect density and attenuation differences, the inherent edge gradient intensity, texture spectrum, and noise sensitivity of tissues with different densities are significantly different; and the Gaussian kernel parameter σ determines the smoothing scale and high-frequency suppression intensity, directly affecting the edge localization accuracy, the probability of generating false edges, and the quality of the subsequent Laplacian of Gaussian (LoG) response. To avoid problems such as over-smoothing of bone structures, tearing of muscle boundaries by noise, and amplification of noise in fat regions to generate false textures in traditional fixed-σ filtering, this embodiment introduces a HU-driven σ adaptive selection and multi-scale fusion mechanism to make the smoothing and enhancement present a closed-loop consistency that is self-adaptive to different density tissues.

[0063] In some embodiments, the density-guided scale-space mentioned above refers to: An adaptive mapping relationship of σ is established based on the pixel / voxel HU value. Preferably, a combination of segmented ranges and continuous mapping is used to form a σ map (σ-map) to achieve dynamic adjustment of the smoothing scale according to tissue density. The specific strategy is as follows: For high-density regions (such as bone tissue, HU > 150), a smaller σ (0.5 - 1.0) is used to retain high-frequency details such as bone cortex / trabeculae and maintain the localization accuracy of the bone-muscle interface; For medium-density regions (such as muscle tissue, - 29 < HU < 150), a medium σ (1.0 - 2.0) is used to balance noise reduction and boundary fidelity, avoid the formation of false edges by noise in subsequent enhancement, and avoid over-smoothing from erasing the true boundary between muscle and fat; For low-density regions (such as fat tissue, - 190 < HU < - 30), a larger σ (2.0 - 3.0) is used to stronger suppress high-frequency noise and avoid phenomena such as granulated edges and false texture enhancement after sharpening or LoG enhancement, which may affect fat boundary and quantitative stability.

[0064] In some embodiments, the HU threshold can be adaptively adjusted according to scanning parameters, reconstruction kernels, and target populations, and the σ range is updated with the HU threshold to improve cross-device robustness.

[0065] In some implementations, to simultaneously ensure the continuity of edges at different scales and different anatomical orientations, multiple Gaussian filter responses (including first-order gradients and / or second-order LoG responses) of different scales (σ) are used in each direction, and a direction consistency fusion mechanism (such as tensor voting or structural tensor consistency weighting) is introduced to enhance edges that are consistent with the tissue anatomical orientation and have stable cross-scale responses, while suppressing cross-scale unstable responses and inconsistent directions, thereby reducing noise-induced false edges and improving edge coherence.

[0066] In some implementations, the directional consistency fusion may assign higher weights near the critical muscle-fat boundary to highlight tissue interfaces of clinical concern.

[0067] The Laplacian operator (LoG) responses of images after Gaussian filtering at different scales are calculated, and the enhanced results are obtained by weighting and superimposing the nonlinear mapping with adaptive coefficients λ. I enh = I original +λ×f(LoG(I original )), Among them, I original To enhance the image, LoG represents the Laplacian Gaussian operator.

[0068] λ is an adaptive enhancement coefficient that is dynamically adjusted according to the local edge confidence and noise level: λ is increased in areas with high edge confidence and controlled noise to improve boundary discernibility, and λ is decreased in areas with high noise or unstable response to suppress artifacts; f(·) is a nonlinear mapping function, preferably with saturation suppression characteristics (e.g., compressing excessively large responses) to avoid halos, ringing or structural shifts caused by over-enhancement.

[0069] By combining the above image edge enhancement strategies, differentiated smoothing control and stable edge enhancement can be achieved between tissues of different densities. This preserves the details and boundary positioning accuracy of high-density structures while suppressing noise amplification and false textures in low-density areas. At the same time, it improves the continuity and separability of key interfaces such as muscle-fat, thus providing a more stable and consistent input for subsequent segmentation and longitudinal quantitative analysis of body components.

[0070] After applying the edge enhancement method of this embodiment, as follows: Figure 5 As shown, the edge sharpness of the muscle-subcutaneous fat interface is improved by approximately 35%, and the edge sharpness of the muscle-visceral fat interface is improved by approximately 40%, while the contours of visceral organs are also significantly enhanced. Compared with conventional Gaussian filtering, the method in this embodiment maintains a non-significant increase in noise level (signal-to-noise ratio decrease of no more than 5%) while enhancing edges, and edge sharpening does not lead to significant changes in tissue density values ​​(average HU value change < 3 HU), making it more suitable for subsequent precise segmentation tasks.

[0071] This Gaussian filtering edge enhancement method, specifically designed for the characteristics of CT images used in sarcopenia identification, fully considers the density characteristics and anatomical features of different tissues. It is significantly superior to the standard Gaussian filtering method in general image processing, laying a solid foundation for subsequent tissue segmentation.

[0072] Step S104: Apply residual feature fusion to segment the enhanced image and identify the image regions corresponding to skeletal muscle, subcutaneous fat and visceral fat.

[0073] The specific process of the segmentation is as follows: The preprocessed 512×512 CT images are fed into the UNet architecture; The HViTER module in the encoder extracts features at multiple levels. The first layer HViTER module (64 channels) extracts shallow edge and texture features; The second-layer HViTER module (128 channels) extracts the morphological features of the middle layer tissue; Enhance feature representation capabilities through residual feature fusion and global feature fusion; The decoder upsamples progressively through transposed convolutions and combines the feature maps with skip connections; The final output includes a 3-channel probability map of muscle, subcutaneous fat, and visceral fat. Figure 6 ).

[0074] In the image segmentation process of this embodiment, the boundary representation of muscle and fat is more accurate after applying residual feature fusion. 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 caused by local similarity and improving classification accuracy by about 3.5%. The accuracy is improved by 5.2% and 6.8% in the identification of muscle-fat mixed regions and small muscle bundles, respectively.

[0075] The segmentation results are as follows: the abdominal CT image is segmented into muscle tissue (red marker), subcutaneous fat (purple marker), and visceral fat (yellow marker); the segmentation results have smooth boundaries and conform to anatomical characteristics, with Dice coefficients of 0.958 (muscle), 0.946 (subcutaneous fat), and 0.937 (visceral fat). Compared with the results of manual segmentation by doctors, the overlap rate reaches 93.5%, and the boundary difference is less than 2 pixels on average.

[0076] Example 2 This embodiment provides a CT medical image processing and recognition system for sarcopenia, such as... Figure 7 As shown, the system includes: The medical image acquisition module 201 is used to acquire CT medical images of patients with sarcopenia or suspected sarcopenia; the CT medical images are plain CT scan images of the waist or abdomen. In some embodiments, the medical image acquisition module 201 can perform image preprocessing on the original CT images to obtain formatted CT images within a standard range.

[0077] The adaptive enhancement module 202 is used to configure enhancement parameters for sarcopenia characteristics, establish a task-driven closed-loop fitness function, and use an adaptive particle swarm optimization algorithm with task-driven closed loop to perform adaptive enhancement processing on the formatted CT image to obtain an adaptively enhanced CT image.

[0078] The edge enhancement module 203 uses a combination of multi-scale adaptive Gaussian filters to perform edge enhancement processing on the adaptively enhanced image, thereby obtaining an edge-enhanced CT image.

[0079] The image segmentation module 204 applies residual feature fusion to segment the edge-enhanced CT image and identify the image regions corresponding to skeletal muscle, subcutaneous fat and visceral fat.

[0080] The processing and identification system is configured to implement the method described in Embodiment 1.

[0081] Example 3 This embodiment provides a computer device 300, the structure of which is as follows: Figure 8 As shown, it includes a memory 301 and a processor 302. The memory 301 stores a computer program 303. When the processor 302 executes the computer program 301, it implements the processing and identification method described in Embodiment 1.

[0082] Embodiments of this specification also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0083] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0084] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A CT medical image processing and recognition method for sarcopenia, characterized in that, The method includes: Acquire CT images of the lumbar or abdominal region of patients with sarcopenia or suspected sarcopenia, and perform image preprocessing operations such as format conversion, size standardization, CT value normalization, and noise reduction on the CT images to be processed. Based on the pathological image characteristics of sarcopenia, parameters such as contrast, brightness, sharpening intensity, noise suppression, and tissue-specific enhancement are configured. A task-driven closed-loop fitness function is established, which includes tissue separability, key boundary clarity, segmentation stability or uncertainty, HU quantitative consistency penalty, and noise amplification penalty. An adaptive particle swarm optimization algorithm is used to enhance the pathological features of sarcopenia in the formatted CT images. In the fitness function, Organizational divisibility term J sep Quantify the degree of grayscale separation of different tissues; the tissue separability term J sep We selected a weighted sum of the ratios of inter-class differences to intra-class dispersion of multiple tissue pairs. That is, J sep =(μ m -μ f ) 2 / (σ m 2 +σ f 2 +ε), Where, μ m σ m The mean HU value and standard deviation of a tissue region are given by μ. f σ f These are the mean HU value and standard deviation of another tissue region, respectively, and ε is a minimal constant to prevent the denominator from being zero; Key boundary clarity term G edge Quantify the edge discernibility of key boundaries; the key boundary sharpness term G edge Calculate the gradient energy within the boundary band B obtained from the coarse segmentation or pre-masking, i.e., ， Among them, I enh For the enhanced image, ▽ is the gradient operator, and (i,j) is the coordinate of any point; Segmenting stability or uncertainty term U seg Suppressing structural misclassification caused by enhancement; the segmentation stability or uncertainty term U seg This refers to inputting the enhanced image into a lightweight coarse segmenter to obtain a tissue region probability map P(k|i,j), and calculating the average entropy as the uncertainty, i.e., ； Where Ω represents the human body region or ROI, (i,j) is the coordinate of any point, and k represents the tissue region; HU Quantitative Consistency Penalty Item P HU Organizational HU statistics before and after constraint enhancement; the HU quantitative consistency penalty term P HU Calculate according to the following formula: ； Where T represents the organization set, μ t (after) and μ t (before) represents the average HU statistic of tissue t before and after enhancement; Noise amplification penalty term N amp Suppress noise amplification caused by sharpening or contrast enhancement; The fitness function employs a task-driven closed-loop optimization strategy, including: Initialize the particle swarm, randomly generate several particles x, and set their velocity vectors; Enhancement is performed on each particle parameter x to obtain the enhanced image I. enh ; Enhanced image I enh The coarse segmenter is used to obtain the tissue region probability map P(k|i,j); the tissue separability term J is calculated. sep Key boundary clarity item G edge , segmentation of stability or uncertainty term U seg HU quantitative consistency penalty item P HU Noise amplification penalty term N amp The fitness function F(x) is obtained based on the weighting ratio; Update individual optimal p best With the global optimal g best ; According to p best With g best Update particle velocity and position; parameters that exceed the boundary are truncated or bounced back to the legal range. The process terminates early when the maximum number of iterations is reached or the improvement in the global optimal fitness is less than a threshold. A multi-scale adaptive Gaussian filter combination based on density-dynamically adjusted kernel parameters is used to perform regional differential edge enhancement on the adaptively enhanced image, thereby strengthening the boundary contours of skeletal muscle, fat, and internal organs. The UNet segmentation model, which combines HViTER feature extraction with residual feature fusion, is used to perform pixel-level segmentation on the edge-enhanced image, identifying and locating the image regions corresponding to skeletal muscle, subcutaneous fat, and visceral fat.

2. The CT medical image processing and recognition method for sarcopenia according to claim 1, characterized in that, The adaptive sarcopenia pathological feature enhancement dynamically adjusts the closed-loop optimization strategy based on the statistical characteristics of the input CT medical images, including: If the noise level is high, increase the HU quantitative consistency penalty term P. HU and noise amplification penalty term N amp The weights are determined, and the upper limit of sharpening intensity is limited; If the noise level is high, increase the relevant weights of the tissue separability and key boundary sharpness terms, and widen the effective range of the contrast enhancement coefficient α. If the uncertainty of coarse segmentation is high, then select a parameter combination that can reduce the entropy value and suppress over-enhancement strategies that lead to misclassification.

3. The CT medical image processing and recognition method for sarcopenia according to claim 1, characterized in that, The adaptive Gaussian filter uses a combination of piecewise range and continuous mapping to form a σ-map, and dynamically adjusts the Gaussian kernel parameter σ according to tissue density, including: For high-density areas, a smaller σ value is used, i.e., σ = 0.5-1.0; For medium-density regions, a medium σ value is used, i.e., σ = 1.0-2.0; For low-density areas, a larger σ value is used, i.e., σ = 2.0-3.

0.

4. The CT medical image processing and recognition method for sarcopenia according to claim 3, characterized in that, The adaptive Gaussian filter uses multiple Gaussian filter responses of different scales in each direction, and introduces directional consistency weighting to enhance edges that are consistent with the tissue anatomy and have stable cross-scale responses. The Laplacian operator response of images after Gaussian filtering at different scales is calculated, and the enhanced result is obtained by weighted superposition of nonlinear mapping and adaptive coefficient λ. I enh =I original +λ×f(LoG(I original )), Where λ is the adaptive enhancement coefficient, I original To enhance the image, LoG stands for the Laplacian of Gaussian operator.

5. A CT medical image processing and recognition system for sarcopenia, characterized in that, The system includes: The medical image acquisition module is used to acquire CT medical images of patients with sarcopenia or suspected sarcopenia. The adaptive enhancement module configures enhancement parameters for sarcopenia characteristics, establishes a task-driven closed-loop fitness function, and uses an adaptive particle swarm optimization algorithm with task-driven closed loop to perform adaptive enhancement processing on the formatted CT images, thereby improving the contrast between skeletal muscle, subcutaneous fat and visceral fat. The edge enhancement module uses a combination of multi-scale adaptive Gaussian filters to perform edge enhancement processing on the adaptively enhanced image; The image segmentation module applies residual feature fusion to segment the edge-enhanced CT image and identify the image regions corresponding to skeletal muscle, subcutaneous fat and visceral fat. The system is configured to implement the method of any one of claims 1 to 4.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images