A slicer-based deep learning fat muscle segmentation method and system

Abstract

This invention provides a deep learning-based fat and muscle segmentation method and system based on Slicer, applied in the field of data processing technology. This application preprocesses preoperative CT image data of the target patient to generate skeletal muscle image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features; preprocesses the training sample set, adjusting window width and level and normalizing it to generate target training image data; trains a pre-defined fat and muscle segmentation model based on a deep learning network using the target training image data to generate a target fat and muscle segmentation model; processes the skeletal muscle image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features based on the target fat and muscle segmentation model to generate target image segmentation information; and processes the target image segmentation information and the target patient's gene detection information to generate preoperative warning information for the target patient.

Description

[0001] This invention claims priority to Chinese Patent Application CN202510709602.2, filed on May 29, 2025, the contents of which are incorporated herein by reference in their entirety and form part of the original description of this invention. The applicant further declares that it has the right to amend the description and claims of this invention based on this priority document. Technical Field This invention relates to the field of data processing technology, and in particular to a deep learning method and system for fat and muscle segmentation based on Slicer. Background Technology

[0002] In the medical field, accurate preoperative assessment of a patient's fat and muscle condition is crucial for surgical planning and prognosis. Currently, there are several shortcomings in processing preoperative patient data and performing fat and muscle analysis. From an image data processing perspective, raw CT images are susceptible to noise interference, resulting in poor image quality and increasing the difficulty of subsequent observation and analysis of fat and muscle tissue. For example, noise may blur the boundary between muscle and fat, affecting the surgeon's accurate judgment of tissue morphology and extent. Furthermore, the lack of unified standards for image data acquired by different devices and with different parameters, coupled with the absence of mature and universally applicable guidelines for operations such as window width and level adjustments, leads to significant differences in processing results, making it difficult to guarantee the accuracy and consistency of the analysis.

[0003] Regarding fat and muscle segmentation models, existing models lack sufficient segmentation accuracy when dealing with the complex structure of human fat and muscle. For example, some models cannot accurately distinguish the boundary between subcutaneous fat and visceral fat, and they also have shortcomings in the detailed segmentation of skeletal muscles, failing to meet the clinical demand for precise segmentation. Furthermore, the data samples used for model training often suffer from limited features and insufficient quantity, resulting in poor generalization ability of the trained models and difficulty in adapting to individual differences among patients.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a deep learning-based fat and muscle segmentation method and system based on Slicer, which at least to some extent overcomes the problems existing in the prior art. By acquiring data such as preoperative CT images and gene testing information from patients, and performing denoising, window width and level adjustment, and normalization preprocessing, tissue image features and target training image data are generated. The trained model is then used to process the image features to obtain target image segmentation information. Combined with gene testing information, analysis is performed from multiple aspects such as tissue morphology and image quantification to generate risk level information and preoperative warnings, assisting doctors in comprehensively understanding the patient's condition and aiming to achieve accurate preoperative assessment.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of this application, a deep learning-based fat and muscle segmentation method based on Slicer is provided, comprising: acquiring preoperative CT image data of a target patient, gene detection information of the target patient, a preset fat and muscle segmentation model based on a deep learning network, and a training sample set; preprocessing the preoperative CT image data of the target patient to generate skeletal muscle image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features; preprocessing the training sample set, adjusting the window width and window level, and normalizing it to generate target training image data; training the preset fat and muscle segmentation model based on the target training image data to generate a target fat and muscle segmentation model; processing the skeletal muscle image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features based on the target fat and muscle segmentation model to generate target image segmentation information, wherein the target image segmentation information includes the segmentation results, region summation, and average density of each of the skeletal muscle image, subcutaneous adipose tissue image, and visceral adipose tissue image; and processing the target image segmentation information and the gene detection information of the target patient to generate preoperative warning information for the target patient.

[0008] Another aspect of this application discloses a deep learning-based fat and muscle segmentation device based on a slicer, characterized by comprising: an acquisition module for acquiring preoperative CT image data of a target patient, gene detection information of the target patient, a preset fat and muscle segmentation model based on a deep learning network, and a training sample set; a processing module for preprocessing the preoperative CT image data of the target patient to generate skeletal muscle image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features; preprocessing the training sample set by adjusting the window width and window level and normalizing it to generate target training image data; training the preset fat and muscle segmentation model based on the target training image data to generate a target fat and muscle segmentation model; processing the skeletal muscle image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features based on the target fat and muscle segmentation model to generate target image segmentation information; and processing the target image segmentation information and the gene detection information of the target patient to generate preoperative warning information for the target patient.

[0009] According to another aspect of this application, an electronic device is characterized by comprising: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described Slicer-based deep learning fat and muscle segmentation method by executing the executable instructions.

[0010] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a second processor, implements the above-described Slicer-based deep learning fat and muscle segmentation method.

[0011] This application provides a deep learning-based fat and muscle segmentation method and system based on Slicer. The system acquires preoperative CT images and genetic testing information from the patient via a server. After denoising, window width and level adjustment, and normalization preprocessing, tissue image features and target training image data are generated. The trained model is then used to process the image features to obtain target image segmentation information. Finally, combined with genetic testing information, analysis is performed from multiple aspects such as tissue morphology and image quantification to generate risk level information and preoperative warnings. This assists doctors in comprehensively understanding the patient's condition, formulating more reasonable surgical plans, and improving the accuracy and effectiveness of medical decisions, aiming to achieve precise preoperative assessment.

[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0013] Figure 1 The flowchart illustrates a deep learning fat and muscle segmentation method based on Slicer provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the structure of a target fat-muscle segmentation model provided in an embodiment of this application; Figure 3 This illustration shows a schematic diagram of a deep learning fat and muscle segmentation device based on Slicer provided in one embodiment of this application. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] The following is combined with Figure 1 This application describes a Slicer-based deep learning method for fat and muscle segmentation according to exemplary embodiments thereof. In one embodiment, this application also proposes a Slicer-based deep learning method and system for fat and muscle segmentation, such as... Figure 1 As shown, this method is applied to a server and includes: S101: Acquire preoperative CT image data of the target patient, gene testing information of the target patient, a pre-set fat and muscle segmentation model based on a deep learning network, and a training sample set.

[0016] In one implementation, preoperative chest CT images of the target patient are retrieved from the hospital's imaging database. This data covers various imaging conditions, such as plain scan, contrast-enhanced scan, thin-slice scan, and thick-slice scan, to ensure a comprehensive reflection of the patient's fat and muscle condition. For example, for a patient about to undergo abdominal surgery, preoperative chest CT images under different scanning modes are retrieved from the database, including the general morphology of the tissue clearly shown on plain scans, the relationship between blood vessels and fat and muscle tissue more clearly presented after contrast-enhanced scans, and the tissue details at different resolutions provided by thin-slice and thick-slice scans. The image data is in DICOM format, a standard format commonly used in the field of medical imaging, facilitating data storage, transmission, and processing.

[0017] Gene sequencing technology is used to test blood or tissue samples from target patients. Sample collection is the first step in gene testing, and blood and tissue samples are important sources of genetic information. Blood samples are collected via venipuncture, drawing 5-10 ml of peripheral venous blood using blood collection tubes containing anticoagulants to prevent clotting. Tissue samples are obtained during surgery or biopsy; for example, for patients undergoing abdominal surgery, a small amount of adipose or muscle tissue can be collected during the procedure. Collected samples need to be processed as soon as possible to prevent nucleic acid degradation. For blood samples, plasma, white blood cells, and other components are separated using methods such as centrifugation to extract DNA. Tissue samples require a series of processing steps, such as chopping and digestion, to release their DNA. This extracted DNA is the basic material for subsequent gene sequencing.

[0018] When sequencing the genes of a target patient, the extracted DNA is first fragmented into short segments suitable for sequencing. These fragments are typically a few hundred base pairs in length and can be achieved through methods such as ultrasonic disruption or enzyme digestion. Next, specific adapter sequences are added to both ends of the fragments. These adapters contain primer binding sites for PCR amplification and sequencing reactions. PCR amplification enriches the number of DNA fragments, ensuring sufficient signal for subsequent sequencing reactions. On the Illumina HiSeq platform, the sequencing reaction is based on the principle of sequencing-by-synthesis (SBS). During sequencing, four dNTPs (deoxyribonucleotides) with different fluorescent labels are sequentially added to the synthesizing DNA strand according to the base complementarity principle. Each time a base is added, a specific color fluorescent signal is emitted. These signals are detected by an optical system to determine the type of base added, thereby enabling the determination of the DNA sequence. This high-throughput sequencing method can sequence a large number of DNA fragments in a single experiment, greatly improving sequencing efficiency.

[0019] The FTO gene is an important gene closely related to lipid metabolism, and variations in the FTO gene are associated with an increased risk of obesity. It influences fat accumulation in the body by affecting appetite regulation and energy metabolism. When conducting genetic testing on target patients, special attention should be paid to variations at specific sites in the FTO gene. Mutations at certain sites may lead to alterations in FTO protein function, thereby affecting adipocyte differentiation and metabolism, making the body more prone to fat storage. Sequencing analysis of the FTO gene can reveal whether obesity-related genetic factors exist in the patient's genetic background, which is crucial for assessing the risk of abnormal fat content and distribution. The PPAR (peroxisome proliferator-activated receptor) family of genes, including PPARα, PPARβ / δ, and PPARγ, plays a key role in lipid metabolism, cell differentiation, and inflammatory responses. PPARα is mainly expressed in tissues such as the liver and skeletal muscle, participating in fatty acid oxidation metabolism and regulating energy balance. PPARγ is highly expressed in adipose tissue and plays an important regulatory role in adipocyte differentiation and lipid storage. Detecting the expression levels and variations of PPAR family genes can provide deeper insights into the molecular mechanisms of lipid metabolism in patients. If the PPARγ gene undergoes certain mutations, it can affect its ability to regulate adipocyte differentiation, leading to lipid metabolism disorders and consequently impacting the patient's fat-muscle status. The MYOD1 gene is a key regulatory gene for muscle growth and development. Belonging to the myogenic regulatory factor family, it promotes myoblast differentiation and muscle fiber formation. During muscle development, the expression of the MYOD1 gene dynamically changes, regulating the expression of a series of genes related to muscle growth.

[0020] like Figure 2As shown, the preset fat and muscle segmentation model is built based on a deep learning network, employing the UNet segmentation model. Its unique architecture effectively captures local and global features in images, demonstrating good adaptability to segmentation tasks involving complex morphologies and structures like fat and muscle tissue. The model adopts an Encoder-Decoder architecture. The Encoder part performs downsampling operations through multiple consecutive convolutions, progressively extracting image feature information while reducing the image size, allowing the model to focus on key image features. During this process, the number of channels in the feature map gradually increases, while the spatial resolution gradually decreases. For example, starting with a single-channel CT image, after a series of convolutions and pooling operations, the number of channels in the feature map gradually increases from 1 to 64, 128, etc., while the image size decreases accordingly. The Decoder part uses multiple consecutive convolutions for upsampling recovery, progressively enlarging the feature map extracted by the Encoder to restore it to a size similar to the original image. Simultaneously, it fuses features from corresponding layers to improve segmentation accuracy. This corresponding layer feature fusion method fully utilizes feature information from different levels, preserving image details and resulting in more accurate segmentation results.

[0021] In the Unet model, starting from the high-level feature maps of the Encoder part, the image size is gradually restored through upsampling operations. Simultaneously, features from corresponding layers are stitched and fused to finally output the segmentation result. The model is designed with single-channel CT images as input because CT images provide rich information about adipose and muscle tissue. The single-channel setting simplifies the model input and focuses on the key features of the CT images. The input size is (batch, 1, 512, 512), where batch refers to the amount of data required for one network iteration for training; this application sets it to 8. This means that in each training iteration, the model processes 8 single-channel CT image datasets of size 512×512 simultaneously. This batch setting helps improve the efficiency and stability of model training. Output parameters: The output is a 4-channel (background, SMM, SAT, VAT) segmentation result, with an output size of (batch, 4, 512, 512). Each channel corresponds to the background, skeletal muscle (SMM), subcutaneous adipose tissue (SAT), and visceral adipose tissue (VAT), respectively. The numbers in the channel represent the probability values ​​belonging to that channel, distributed from 0 to 1. Ultimately, the result for each location is classified using the highest probability across multiple channels, forming a four-value map to achieve accurate segmentation of different tissues. During training, the model continuously adjusts its parameters to gradually reduce the DiceLoss value, thus optimizing the model. Network iterative parameter optimization: Cross-entropy loss is used as network supervision, and stochastic gradient descent is employed to allow the network to learn through backpropagation from errors in predictions and manually labeled results, eventually leading to convergence and completing the entire learning process. During this process, the model continuously adjusts the weight parameters in the network based on the gradient information of the loss function, enabling the model to better fit the training data. Parameters such as the learning rate play a crucial role in the training process; setting a reasonable learning rate can control the speed and effectiveness of model convergence, avoiding overfitting or slow convergence.

[0022] The training sample set was selected from a large number of historical medical records accumulated in the hospital, containing complete preoperative and postoperative data. These samples included patients of different ages, genders, and disease types to ensure the model's generalization ability. For example, 1000 preoperative CT image data and corresponding gene testing information from patients with different diseases such as liver cancer and gastric cancer were selected, including 500 male patients and 500 female patients, aged between 30 and 70 years old. These sample data were organized, with the CT image data preprocessed according to uniform standards, such as adjusting the image size to 512×512 pixels and normalizing pixel values ​​to the [0,1] interval; the gene testing information was categorized and organized, and the variation of different gene loci was marked. The organized CT image data and gene testing information were matched one-to-one to form a structured training sample set for subsequent training of the preset fat-muscle segmentation model.

[0023] S102 preprocesses the preoperative CT image data of the target patient to generate musculoskeletal imaging features, subcutaneous adipose tissue imaging features, and visceral adipose tissue imaging features.

[0024] In one implementation, denoising is performed on the preoperative CT image data of the target patient to generate low-noise CT images. In actual clinical environments, CT images may be affected by various types of noise, such as electronic noise and quantum noise, which can affect image quality and the accuracy of subsequent analysis. A nonlocal mean filtering algorithm is used to denoise the preoperative CT image data of the target patient. This algorithm uses similar local regions in the image to estimate the true value of each pixel, thereby effectively removing noise. Taking the patient preparing for abdominal tumor surgery as an example, the original CT images show significant noise in the lung tissue and abdominal fat regions. After nonlocal mean filtering, the noise in the image is significantly reduced, and the originally blurred tissue boundaries become clearer. For example, the boundary between the liver edge and the surrounding fat tissue is more distinct, providing a clearer base image for subsequent processing.

[0025] Window width and level adjustments were made to low-noise CT images to generate CT images highlighting soft tissue. Based on clinical experience and the imaging characteristics of soft tissue, adjusting the window width and level can highlight the imaging features of specific tissues. For this patient's low-noise CT images, the window width was set to 350-450 HU (e.g., 400 HU), and the window level was set to 30-50 HU (e.g., 40 HU). This adjustment enhanced the contrast of soft tissue in the image, making the details of muscles, fat, and other tissues more apparent. The muscle texture of the patient's abdomen, as well as the distribution of subcutaneous and visceral fat, can be clearly seen. Fat layers that were difficult to distinguish under normal window width and level are now clearly visible after adjustment. The CT images highlighting soft tissue were then normalized to generate normalized CT images. To ensure that the CT image data from different patients have a uniform scale and range, facilitating model processing and comparative analysis, the CT images highlighting soft tissue were normalized. A linear normalization method was used to map the pixel values ​​of the image to the [0,1] interval. After normalization, all pixel values ​​in the patient's CT images were standardized to a uniform range, making the differences in grayscale values ​​between different tissues more stable and comparable within this standard range. For example, regardless of the pixel value range of other patients' original CT images, after normalization, they can be analyzed at the same scale as the patient's images, avoiding model training bias caused by differences in data scale.

[0026] Normalized CT images were segmented into muscular regions to generate target muscular images. The Unet model was used for muscular region segmentation on the normalized CT images. This model, by learning from a large amount of labeled CT image data, can accurately identify muscular regions. In the image processing of this patient, the model successfully segmented the abdominal muscular region from other tissues, generating the target muscular image. The complete outlines of the rectus abdominis, external oblique, and other abdominal muscles were clearly visible, providing precise regions for subsequent feature extraction. Feature extraction was performed on the target muscular image to generate muscular image features. Image feature extraction algorithms, such as the HOG (Histogram of Oriented Gradients) algorithm, were used to extract features from the target muscular image. The HOG algorithm describes the texture and shape features of an image by calculating the gradient orientation histogram of local regions. For this patient's muscular image, the HOG algorithm extracted features such as muscle texture orientation and edge intensity. In the rectus abdominis region, HOG features clearly reflected the direction and arrangement of muscle fibers, which can be used to assess the health and functional status of the muscles.

[0027] Normalized CT images were segmented to generate target subcutaneous adipose tissue images. Feature extraction was then performed on these images to generate subcutaneous adipose tissue image features. The Unet model accurately identified the extent of subcutaneous adipose tissue, generating target subcutaneous adipose tissue images. These images clearly showed the thickness and distribution of subcutaneous fat in the patient's abdomen. Next, the Gray-Level Co-occurrence Matrix (GLCM) algorithm was used to extract features from the target subcutaneous adipose tissue images. The GLCM algorithm calculates the spatial correlation of gray values ​​in the image, thereby obtaining texture features. Through the GLCM algorithm, texture features of the subcutaneous adipose tissue, such as contrast, correlation, energy, and entropy, were extracted. These features reflect the uniformity and structural characteristics of the subcutaneous adipose tissue, which is of great significance for assessing the patient's fat metabolism.

[0028] Normalized CT images were segmented to generate target visceral adipose tissue images. Feature extraction was then performed on these images to generate visceral adipose tissue image features. Taking this patient's image data as an example, the model separated the visceral adipose tissue within the abdominal cavity from other tissues, generating target visceral adipose tissue images. These images clearly show the distribution of visceral fat surrounding the patient's liver, intestines, and other organs. After generating the target visceral adipose tissue image, a combination of morphological feature analysis and deep learning feature extraction was used to obtain its features. Morphological analysis calculated geometric features such as area, perimeter, and roundness. For example, the area of ​​the visceral adipose tissue in the patient's CT images was measured as [X] square centimeters, the perimeter as [Y] centimeters, and the roundness value reflected the regularity of its shape. These geometric features can intuitively demonstrate the scale and morphological characteristics of the visceral adipose tissue, aiding in the assessment of the degree and distribution of fat accumulation. Pre-trained convolutional neural networks (CNNs), such as ResNet, were used to extract features from the target visceral adipose tissue images. CNNs can automatically learn complex texture and structural features in images. By inputting an image of the target visceral adipose tissue into a pre-trained ResNet model, the output of specific layers is extracted as depth features. These depth features contain rich information, such as the microscopic texture of the adipose tissue and its relationship with surrounding tissues.

[0029] S103, preprocess the training sample set, adjust the window width and window level and normalize it to generate target training image data.

[0030] In one implementation, the training sample set is denoised to generate low-noise CT images. During medical image acquisition, CT equipment is affected by various factors, resulting in noise in the CT images of the training sample set. This noise interferes with subsequent analysis and model training. Therefore, a Gaussian filtering algorithm is used to denoise each CT image in the training sample set. The impact of noise is reduced by weighted averaging of each pixel and its neighboring pixels. For the CT images of these 1000 patients, in the images of liver cancer patients, noise blurs the boundaries between the fatty and muscle tissues surrounding the liver. After Gaussian filtering, the noise in the image is reduced, and the tissue boundaries become clearer. For example, the previously blurred boundary between the liver and the surrounding fat becomes distinct, providing a foundation for subsequent image processing.

[0031] Window width and level adjustments were made to low-noise CT images to highlight soft tissue image features and generate normalized training image data. Due to differences in scanning equipment and parameters, the display of soft tissue in CT images varies among different patients. To highlight soft tissue image features and facilitate subsequent analysis of fat and muscle tissue, window width and level adjustments were made to the low-noise CT images based on clinical experience and the imaging characteristics of soft tissue. The window width was set to 400 HU, and the window level to 40 HU. After this adjustment, the contrast of soft tissues such as muscle and fat was enhanced in the images of these 1000 patients. Muscle textures and fat layers that were previously difficult to distinguish under normal window width and level are now clearly displayed. For example, in the CT images of gastric cancer patients, the distribution of muscle and fat tissue around the stomach can be observed more clearly, facilitating subsequent analysis and annotation. Subsequently, to ensure that the CT image data from different patients have a uniform scale and range, facilitating model training and comparative analysis, a linear normalization method was used to process the adjusted CT images, mapping the pixel values ​​of the images to the [0,1] interval. In this way, all patients' CT image data are processed at the same scale, avoiding model training bias caused by differences in data scale.

[0032] Normalized training image data was screened and labeled to clearly define the skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue regions, generating labeled training image data. Images with poor quality, obvious artifacts, or incompleteness were removed from the normalized training image data. From 1000 data points, 800 high-quality images were selected for subsequent labeling, clearly defining the skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue regions on these images. For example, during the labeling process, experts carefully delineated the outlines of abdominal skeletal muscles such as the rectus abdominis and external oblique muscles, as well as the extent of subcutaneous and visceral fat, in each image, generating labeled training image data. The skeletal muscle regions in the labeled training image data were segmented and feature extracted to generate skeletal muscle training features. From the 1000 labeled training image data, an image of a liver cancer patient was selected. The skeletal muscle region in this image was segmented using the Unet segmentation model based on deep learning. The Unet model, trained on a large amount of image data with skeletal muscle annotations, accurately identified the skeletal muscle range of the patient's abdomen, clearly outlining muscles such as the rectus abdominis and external oblique. After segmentation, the HOG algorithm was used to extract features from the segmented skeletal muscle images. The HOG algorithm describes the texture and shape features of an image by calculating the gradient orientation histogram of local image regions. On the patient's skeletal muscle images, the HOG algorithm extracted features such as muscle texture direction and edge intensity. For example, it was found that the patient's abdominal muscles had atrophied due to liver cancer, and the direction of the muscle texture changed compared to the normal state, with different edge intensities. These features were accurately captured by the HOG algorithm, forming the training features for the patient's skeletal muscles.

[0033] Subcutaneous adipose tissue regions in the labeled training image data were segmented and feature extracted to generate training features for subcutaneous adipose tissue. From 1000 labeled training images, an image of a liver cancer patient was selected. This training image was segmented using another trained segmentation model to accurately identify the musculoskeletal region of the patient's abdomen, such as clearly outlining the rectus abdominis and external oblique muscles. After segmentation, the HOG algorithm was used to extract features from the segmented musculoskeletal image, describing the texture and shape features of the image by calculating the gradient orientation histogram of local image regions. On the patient's musculoskeletal image, the HOG algorithm extracted features such as muscle texture direction and edge intensity. For example, it was found that the patient's abdominal muscles had atrophied due to liver cancer, and the direction of the muscle texture changed compared to the normal state, with different edge intensities. These features were accurately captured by the HOG algorithm, forming the training features for the patient's musculoskeletal system.

[0034] Similar to the processing of musculoskeletal regions, subcutaneous adipose tissue was first accurately segmented from the labeled training image data. This helps in the precise analysis of subcutaneous adipose tissue characteristics, which are important for understanding the patient's fat metabolism, degree of obesity, and association with disease. These characteristics are also key data for training the model to identify subcutaneous adipose tissue. Using the labeled training image data of the aforementioned liver cancer patient, a method combining threshold segmentation and morphological operations was employed to segment the subcutaneous adipose tissue region. First, an appropriate threshold was set based on the grayscale value range of subcutaneous adipose tissue on CT images to initially segment the subcutaneous adipose tissue. Then, morphological dilation and erosion operations were used to remove some noise and small missegmented regions, resulting in a more accurate image of the subcutaneous adipose tissue. Next, the Gray-Level Co-occurrence Matrix (GLCM) algorithm was used to extract features from the segmented subcutaneous adipose tissue image. The GLCM algorithm obtains texture features by calculating the spatial correlation of grayscale values ​​in the image. From the patient's subcutaneous adipose tissue image, the GLCM algorithm extracted features such as contrast, correlation, energy, and entropy. For example, the calculated contrast feature reflects the degree of difference between different gray-level regions within subcutaneous adipose tissue. A high contrast indicates that the adipose tissue is unevenly distributed. The correlation feature reflects the similarity of gray-level values ​​between pixels and can be used to determine whether the texture of adipose tissue is regular. The energy feature reflects the uniformity of the gray-level distribution of the image. The higher the energy value, the more regular the texture. Entropy measures the complexity of the gray-level distribution in the image. A higher entropy value means that the texture of adipose tissue is more complex.

[0035] The visceral adipose tissue region in the labeled training image data is segmented and its features are extracted to generate visceral adipose tissue training features. Taking a liver cancer patient from the selected 1000 data points as an example, the visceral adipose tissue region in the labeled training image data is segmented. After segmentation, features are obtained using a combination of morphological analysis and deep learning feature extraction. Through morphological analysis, the geometric features of the patient's visceral adipose tissue, such as area, perimeter, and roundness, are calculated. For example, the area of ​​the visceral adipose tissue in this patient's CT image is measured to be 80 square centimeters, the perimeter to be 60 centimeters, and the roundness value to be 0.7 (the closer the roundness value is to 1, the more regular the shape). These geometric features visually demonstrate the scale and morphological characteristics of the visceral adipose tissue. The images are input into a ResNet model, and the output of specific layers is extracted as depth features. These depth features contain rich information such as the microscopic texture of the adipose tissue and its relationship with surrounding tissues. For example, by analyzing these depth features, abnormal boundary features were found between the patient's visceral adipose tissue and the surrounding liver tissue, suggesting the impact of the disease on the relationship between the liver and adipose tissue. These features together constitute the training features of the patient's visceral adipose tissue.

[0036] Training features from skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue are integrated to generate target training image data. First, the feature vectors of skeletal muscle are ranked, followed by those of subcutaneous adipose tissue, and finally those of visceral adipose tissue, forming a comprehensive feature vector. This integration process is performed on each patient in a dataset of 1000 cases, ultimately yielding target training image data containing the comprehensive feature vectors of all patients. This data will serve as training samples, inputting into a pre-defined adipose-muscle segmentation model based on a deep learning network. This allows the model to learn the patterns and characteristics of various adipose-muscle tissue features under different disease states, enabling more accurate segmentation and analysis of adipose-muscle tissue when processing new patient CT images, thus providing strong support for preoperative assessment.

[0037] S104, Based on the target training image data, the preset fat and muscle segmentation model based on the deep learning network is trained to generate the target fat and muscle segmentation model.

[0038] In one implementation, the number of features for skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue in the target training image data is counted, and a sampling ratio is generated based on the balanced distribution of these features across different tissues. After obtaining the target training image data, it is necessary to analyze the number of features for each of the skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue. These features include information on texture, shape, grayscale, and other aspects. The sampling ratio is generated based on the balanced distribution of these features across different tissues to avoid model training bias caused by an excessive or insufficient number of features in certain tissues, ensuring that the model can learn the features of various tissues. Statistical analysis revealed that skeletal muscle tissue has 50 different texture features and 30 different shape features; subcutaneous adipose tissue has 40 texture features and 25 different shape features; and visceral adipose tissue has 45 texture features and 28 different shape features. Analysis showed that the texture features of skeletal muscle tissue are relatively concentrated in the dataset, while the shape features of subcutaneous adipose tissue are more dispersed. To balance the proportion of different tissue features in subsequent training, the sampling ratio for skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue is calculated to be 3:2:2 based on the distribution of these features. This ratio means that in subsequent sampling, samples will be selected from the features of different tissues according to this ratio to ensure that the model can fully learn the characteristics of various tissues.

[0039] Based on the sampling ratio, stratified sampling processing is performed on the target training image data to generate a preset number of sampled feature combinations from skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue features. According to the previously generated sampling ratio, stratified sampling is performed on the target training image data. Stratified sampling is a method of sampling from different levels or categories according to a certain proportion to ensure that representative samples can be obtained from the features of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue. Using a sampling ratio of 3:2:2, assuming 100 sampled feature combinations are to be generated, 30 feature sets are randomly selected from the numerous features of skeletal muscle tissue (according to the ratio 3 / (3+2+2)×100=30), 20 feature sets are selected from subcutaneous adipose tissue features (2 / (3+2+2)×100=20), and 20 feature sets are selected from visceral adipose tissue features. For example, in one sampling, feature combinations representing muscle texture direction and edge sharpness were selected from skeletal muscle tissue; feature combinations reflecting the uniformity of fat distribution and texture coarseness were selected from subcutaneous adipose tissue; and feature combinations reflecting its relationship with surrounding organs and shape regularity were selected from visceral adipose tissue. These feature combinations from different tissues were combined to form a sampling feature set containing information from multiple tissues. A total of 100 such sets were generated to prepare for subsequent analysis.

[0040] Features were selected from all available features to differentiate between skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue. These key features are core elements that effectively reflect the differences between these tissues. A similarity measurement analysis was performed on each sampled feature combination with these key features to quantitatively determine the degree of similarity between each sampled feature combination and different tissues. Based on the analysis results, the data was divided into high-quality and low-quality sample groups. Data samples in the high-quality sample group showed high similarity to the key features, more accurately representing the characteristics of various tissues, and thus having greater value for model training; the low-quality sample group showed the opposite. Each sample group contained a pre-defined number of data samples, ensuring that at least one data sample carried labeling information. This labeling information was used during training to clarify the category of the sample, helping the model learn and identify different tissues. From the previously processed data of 1000 patients, after in-depth analysis and screening of various features, 10 key features were identified, including specific texture patterns (such as the fibrous texture of skeletal muscle and the uniform texture of adipose tissue) and shape indices (such as the compactness of visceral adipose tissue and the distribution range index of subcutaneous adipose tissue). For the 100 generated sampled feature combinations, a cosine similarity algorithm was used to calculate their similarity to the 10 key features. Sampled feature combinations with a similarity higher than 0.7 were classified as high-quality sample groups, and those with a similarity lower than 0.5 were classified as low-quality sample groups. For example, a sampled feature combination containing typical fiber texture features and specific shape features of skeletal muscle has a cosine similarity of 0.8 with the key features, so this combination is classified as a high-quality sample group; while another sampled feature combination, due to its less typical features, has a cosine similarity of only 0.4 with the key features, and is classified as a low-quality sample group. Ultimately, the high-quality sample group contains 40 sampled feature combinations, and the low-quality sample group contains 60 sampled feature combinations, with at least one sample in each group having a clear label, such as "liver cancer patient - skeletal muscle" or "gastric cancer patient - subcutaneous fat," so that the tissue category of the sample can be clearly distinguished during model training.

[0041] The pre-defined fat and muscle segmentation model based on a deep learning network is iteratively trained using high-quality and low-quality sample groups, generating the trained model and performance metrics. During training, the model continuously adjusts its parameters based on the sample data, optimizing its ability to identify and segment different tissue features. In each iteration, the model attempts to reduce the difference between the predicted results and the ground truth labels, gradually improving the segmentation accuracy for various tissues. As training progresses, the trained model and corresponding performance metrics are generated. These metrics are used to evaluate the model's performance during training, such as accuracy, recall, and Dice coefficient, providing a direct understanding of the model's segmentation effectiveness for different tissues. The pre-defined high-quality and low-quality sample groups are input into the pre-defined fat and muscle segmentation model based on the Unet architecture. During training, the cross-entropy loss function is used as the optimization objective, and the stochastic gradient descent algorithm is employed to adjust the model's weight parameters. In each training round, the model processes the sample data, calculates the loss value between the predicted results and the ground truth labels, and then updates the weights based on the gradient information of the loss value. After 50 rounds of training, the model achieved an accuracy of 85%, a recall of 80%, and a Dice coefficient of 0.82 on the training set. These are the performance metrics of the trained model. These metrics indicate that the model can accurately identify most tissue regions when segmenting fat and muscle tissue in the training data, and the segmentation results have a high degree of consistency with the real situation.

[0042] If the model recognizes the labeled information in the training results as a feature distinguishing skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue, and the performance index reaches the preset threshold, then the trained model is used as the target adipose-muscle segmentation model. The model's ability to accurately identify the labeled information is checked, i.e., whether it can accurately classify samples with labeled information into the corresponding skeletal muscle, subcutaneous adipose tissue, or visceral adipose tissue categories. Simultaneously, the model's performance index is compared with the preset threshold. If the model can correctly identify the labeled information, and the performance index reaches or exceeds the preset threshold, it indicates that the model performs well in learning different tissue features, possessing high accuracy and reliability. This trained model can then be used as the target adipose-muscle segmentation model for subsequent segmentation and analysis of the target patient's adipose-muscle tissue. If the conditions are not met, the training parameters need to be adjusted or retraining is required. In this training, the preset performance index thresholds for the model were 80% accuracy, 75% recall, and a Dice coefficient of 0.8. After training, the model achieved an accuracy of 83%, a recall of 78%, and a Dice coefficient of 0.81 on the test set for samples with labeled information, all reaching the preset thresholds. Meanwhile, the model can accurately identify the tissue categories corresponding to different labeling information. For example, for a sample labeled "liver cancer patient - musculoskeletal," the model can accurately segment it into musculoskeletal tissue, and the segmentation results are quite consistent with the real situation in terms of shape and location. This indicates that the model performs well in learning different tissue features and has the ability to accurately segment fat and muscle tissue. Therefore, this trained model is identified as the target fat and muscle segmentation model. In subsequent applications, this model can segment musculoskeletal, subcutaneous, and visceral fat tissues in the preoperative CT images of target patients, providing doctors with accurate image analysis results and assisting in preoperative assessment. If the model's performance indicators do not reach the preset thresholds, such as an accuracy of only 75%, a recall of 70%, and a Dice coefficient of 0.75, the training process needs to be adjusted. Possible adjustments include increasing the amount of training data, reselecting key features, adjusting the model's hyperparameters (such as learning rate, number of iterations, etc.), or improving the model architecture, and then retraining until the model meets the preset performance requirements.

[0043] S105, based on the target fat and muscle segmentation model, processes the image features of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue to generate target image segmentation information.

[0044] In one implementation, feature normalization is performed on the skeletal muscle image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features based on a target fat-muscle segmentation model to generate normalized tissue image features. For the patient preparing for abdominal surgery, the pixel values ​​of the skeletal muscle image features range from 0 to 255, the pixel values ​​of the subcutaneous adipose tissue image features range from -100 to 100, and the pixel values ​​of the visceral adipose tissue image features range from 50 to 200. Using a min-max normalization method, the pixel values ​​of these features are mapped to the [0,1] interval. For the skeletal muscle image features, assuming a pixel value of 128, after normalization calculation (128-0) / (255-0)≈0.5, it is transformed to the [0,1] interval. Similar processing is performed on the subcutaneous adipose tissue and visceral adipose tissue image features, so that all tissue image features are at the same scale, which facilitates subsequent model processing.

[0045] Based on the target fat-muscle segmentation model, feature extraction and recognition are performed on the normalized tissue image features to generate preliminary segmentation results. After normalization, the target fat-muscle segmentation model is used to perform in-depth analysis on these uniform-scale image features. The model extracts representative features and identifies feature patterns related to skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue, thereby generating preliminary segmentation results. These features initially outline the possible distribution of different tissues in the image. Taking the target fat-muscle segmentation model based on the Unet architecture as an example, the convolutional layer of this model performs convolution operations on the normalized tissue image features to extract features such as texture, edges, and shape. When processing the patient's image, the model identifies some texture features and finds that a certain region has a specific texture pattern similar to the texture features of skeletal muscle tissue in the training data, thus initially judging that the region may belong to skeletal muscle tissue, generating corresponding preliminary segmentation results, and marking the probability information that the region may be skeletal muscle on the feature map.

[0046] Based on the target fat and muscle segmentation model, the preliminary segmentation results are processed for region division, clearly defining the respective regions of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue images, generating region division features, and merging and refining regions with high probabilities in the images. For example, regions with a probability greater than 0.5 are identified as the corresponding tissue regions. In the patient's images, the abdominal region is divided into skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue regions, clearly defining the boundaries of different tissues and generating region division features. These features clearly define the specific location and extent of each tissue in the image. Area calculation is performed on the region division features to obtain the regional areas of the skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue images, generating regional area features. After region division, the skeletal muscle region contains 5000 pixels. Given that the conversion relationship between pixels and actual area in this CT image is 1 pixel = 0.1 square centimeters, the area of ​​the skeletal muscle image is 5000 × 0.1 = 500 square centimeters. Similarly, the area of ​​the subcutaneous adipose tissue image was calculated to be 300 square centimeters, and the area of ​​the visceral adipose tissue image was calculated to be 200 square centimeters. These area data constitute the regional area features.

[0047] Density calculations were performed on the region segmentation features to obtain the average density of the skeletal muscle image, subcutaneous adipose tissue image, and visceral adipose tissue image, generating the average density feature. After region segmentation, for the skeletal muscle region, assuming it contains 5000 pixels with a total pixel value of 300,000, the average density of the skeletal muscle image is 300,000 ÷ 5000 = 60 (HU, Henlein units, the unit of CT value). For the subcutaneous adipose tissue region, if it contains 3000 pixels with a total pixel value of -60,000, its average density is -60,000 ÷ 3000 = -20 (HU). The visceral adipose tissue region contains 2000 pixels with a total pixel value of 100,000, and its average density is 100,000 ÷ 2000 = 50 (HU). These calculated average density values ​​constitute the average density feature. The average density of different tissues varies significantly. For example, subcutaneous adipose tissue has a low average density, which is consistent with the characteristics of adipose tissue. In contrast, skeletal muscle and visceral adipose tissue have relatively high average densities and different numerical ranges. Through these average density characteristics, doctors can further understand the density of different tissues in the patient's body, which can help in diagnosis.

[0048] Integrate preliminary segmentation features, region area features, and average density features to generate target image segmentation information. Combine preliminary segmentation features (e.g., feature maps indicating different tissue possibilities), region area features (e.g., skeletal muscle area 500 sq cm, subcutaneous adipose tissue area 300 sq cm, visceral adipose tissue area 200 sq cm), and average density features (e.g., skeletal muscle average density 60 HU, subcutaneous adipose tissue average density -20 HU, visceral adipose tissue average density 50 HU). This information can be stored in a structured data format, such as a table containing fields for tissue type, segmentation result, region area, and average density. In this table, each row corresponds to a specific tissue, recording detailed information about that tissue.

[0049] S106 processes the target image segmentation information and the target patient's gene detection information to generate preoperative warning information for the target patient.

[0050] In one implementation, morphological analysis is performed on the segmentation results of the skeletal muscle image, subcutaneous adipose tissue image, and visceral adipose tissue image in the target image segmentation information to generate tissue morphological feature information. For the skeletal muscle image segmentation results, morphological analysis is performed using specialized image processing software (such as ITK-SNAP). The software's measurement tools revealed changes in the shape of the abdominal skeletal muscles (such as the rectus abdominis). Normally, the rectus abdominis is a relatively regular, elongated shape, but in this patient, due to the tumor, the rectus abdominis showed local distortion and atrophy, resulting in an irregular shape. Regarding size, software calculations showed that compared to healthy individuals of the same age and body type, the patient's rectus abdominis area was significantly smaller, indicating possible muscle atrophy due to disease. Furthermore, the smoothness of the edges was examined; normal muscle tissue has relatively smooth edges, while the edges of this patient's rectus abdominis became rough and discontinuous, possibly due to tumor infiltration or a stress response. Similarly, analysis of subcutaneous adipose tissue images revealed uneven distribution, with some areas showing excessive fat accumulation, appearing as irregular clumps of varying shapes and sizes, and with blurred edges. This may be related to metabolic disorders or disease-induced fat redistribution. For visceral adipose tissue images, the distribution around organs such as the liver and intestines changed, becoming disordered from its previously relatively regular pattern. In some areas, the fat layer thickness increased, with protrusions and depressions appearing, and the edges becoming less smooth. These changes all indicate that disease affects fat metabolism and storage.

[0051] Statistical analysis was performed on the area and average density data of the target image segmentation information to generate image quantitative feature information, including area proportion and density distribution difference characteristics. In the target image segmentation information, the area of ​​the skeletal muscle image is known to be 400 square centimeters, the area of ​​the subcutaneous adipose tissue image is 350 square centimeters, and the area of ​​the visceral adipose tissue image is 250 square centimeters. By calculating the proportion of each tissue area to the total abdominal area (assuming the total abdominal area is 1000 square centimeters), the proportion of skeletal muscle area was found to be 40% (400 ÷ 1000 × 100%), the proportion of subcutaneous adipose tissue area was 35%, and the proportion of visceral adipose tissue area was 25%. Comparison with reference data on the proportion of abdominal adipose and muscle tissue area in normal individuals of the same age, sex, and body type showed that the proportion of skeletal muscle area in normal individuals is typically between 45% and 55%, the proportion of subcutaneous adipose tissue area is between 20% and 30%, and the proportion of visceral adipose tissue area is between 15% and 20%. The patient's skeletal muscle area was below the normal range, while the subcutaneous and visceral fat areas were relatively high, suggesting muscle wastage and abnormal fat metabolism. These changes in area proportions preliminarily indicate an imbalance in the patient's body composition, affecting their postoperative recovery and overall health.

[0052] Regarding mean density, the mean density of skeletal muscle images was 50 HU (Henness units), the mean density of subcutaneous adipose tissue images was -70 HU, and the mean density of visceral adipose tissue images was -50 HU. Normally, the mean density of skeletal muscle is between 40-60 HU, the mean density of subcutaneous adipose tissue is between -90 and -70 HU, and the mean density of visceral adipose tissue is between -60 and 40 HU. Comparison with the mean density range of normal tissues revealed that the patient's subcutaneous adipose tissue mean density was slightly below the lower limit of the normal range, indicating some changes in the composition or structure of subcutaneous fat. For example, the lipid composition within adipocytes may have changed, or there may be some degree of inflammation affecting fat density. The mean density of visceral adipose tissue was relatively higher than the normal range, possibly related to excessive fat accumulation, adipocyte hypertrophy, and local inflammatory responses. Further analysis of density distribution differences and calculation of the standard deviation of tissue density in different regions were conducted to assess the homogeneity of tissue density. Suppose that the calculated density standard deviation in the patient's visceral adipose tissue image is large, indicating uneven density distribution within the visceral adipose tissue and the possible presence of fat regions with different metabolic states. This is significant for assessing disease progression and prognosis. Statistical analysis of these regional area and average density data generates quantitative image feature information including area proportion and density distribution differences. If the patient's visceral fat area is found to be excessively high and its density distribution abnormal, it may indicate a higher risk of cardiovascular disease, requiring greater attention to cardiovascular issues during preoperative preparation and postoperative care.

[0053] In another implementation, the genetic testing information of the target patient is screened for genes related to fat and muscle development and metabolism to generate a key gene set. The patient's genetic testing was performed using the Illumina HiSeq series high-throughput sequencing platform, yielding a large amount of genetic data. From this data, based on known medical research findings, genes such as FTO, the PPAR family (PPARα, PPARβ / δ, PPARγ), and MYOD1 were screened. FTO gene variants are associated with an increased risk of obesity, affecting fat accumulation by influencing appetite regulation and energy metabolism; PPAR family genes play crucial roles in fat metabolism, cell differentiation, and inflammatory responses; for example, PPARγ is essential for adipocyte differentiation and lipid storage; and the MYOD1 gene is a key regulatory gene for muscle growth and development. These genes constitute the key gene set information, providing core data for analyzing the genetic factors related to fat and muscle in this patient. Functional analysis of the key gene set information is then performed to generate gene function impact characteristic information. Analysis of the FTO gene revealed a mutation at a specific site in this patient's FTO gene. This mutation altered the function of the FTO protein, leading to abnormal adipocyte differentiation and metabolism, and increasing the body's tendency to store fat. Regarding the PPARγ gene, its expression level was detected to be below the normal range, which may affect normal adipocyte differentiation and lipid storage regulation, resulting in disordered fat metabolism. Abnormal expression of the MYOD1 gene was also observed, potentially affecting myoblast differentiation and muscle fiber formation, leading to poor muscle development. These analyses combined to generate characteristic information on gene function impacts, including the functional abnormalities of key genes and their potential effects on the physiological processes of fat and muscle, providing a genetic basis for subsequent assessments.

[0054] Correlation analysis was performed on tissue morphology features, imaging quantitative features, and gene function impact features to generate a comprehensive impact feature vector. Tissue morphology features showed that the patient's abdominal muscles were irregular in shape and reduced in volume, subcutaneous adipose tissue was unevenly distributed with thickening in some areas, and visceral adipose tissue exhibited disordered morphology around the liver and intestines. Imaging quantitative features indicated that the proportion of skeletal muscle area was lower than normal, while the proportions of subcutaneous and visceral fat area were higher, and the average density of adipose tissue was also abnormal. Combined with gene function impact features, mutations in the FTO gene, abnormal PPARγ gene expression, and altered MYOD1 gene expression collectively affected the normal development and metabolism of adipose muscle. For example, mutations in the FTO gene make the body more prone to fat storage, which is associated with increased subcutaneous and visceral fat area; insufficient PPARγ gene expression affects fat metabolism, further leading to abnormal fat distribution and density changes; and abnormal MYOD1 gene expression is associated with muscle atrophy and shape changes. By performing correlation analysis on this information, a comprehensive influence feature vector containing multiple characteristics is constructed, such as [degree of abnormal muscle shape, proportion of muscle area, difference in adipose tissue density, degree of influence of FTO gene mutation, degree of abnormality of PPARγ gene expression, degree of abnormality of MYOD1 gene expression, etc.]. This vector comprehensively reflects the combined influence of multiple factors on the patient's fat and muscle condition.

[0055] Risk assessment is performed on the comprehensive impact feature vector to generate risk level information related to the patient's preoperative health status. Features such as the degree of muscle atrophy, abnormal distribution and density changes of adipose tissue are given high weights in the model, and these features are significant in the patient's comprehensive impact feature vector. Abnormal features related to the FTO gene, PPARγ gene, and MYOD1 gene also have a significant impact on the results. After calculation, the model outputs a risk score. Based on pre-defined risk level classification criteria, the risk score is mapped to different risk levels, such as low risk, medium risk, and high risk. Assuming a risk score of 0-30 is low risk, 31-60 is medium risk, and 61-100 is high risk, if the patient's risk score is 70, the patient is classified as high risk. This means that the patient may have a high preoperative health risk, such as difficulty in postoperative recovery and an increased probability of complications. Preoperative warning information for the target patient is generated based on the risk level information. For patients identified as high-risk, the generated preoperative warning information might include: "This patient is classified as high-risk due to gene mutations (specific site mutation in the FTO gene, abnormal PPARγ gene expression, and altered MYOD1 gene expression), combined with abnormal morphological and quantitative characteristics of adipose-muscle tissue (muscle atrophy, uneven fat distribution and abnormal area ratio, and altered fat density). Further examinations are recommended before surgery, with close monitoring of cardiovascular function, as the patient's lipid metabolism disorder and muscle dysfunction may increase the risk of cardiovascular disease. The surgical plan should consider the impact of insufficient muscle strength on surgical tolerance, and anesthesia and surgical procedures should be adjusted accordingly. Postoperative nutritional support and rehabilitation training should be strengthened to promote muscle recovery and improve lipid metabolism, reducing the probability of postoperative complications." Such preoperative warning information provides doctors with comprehensive and targeted references, helping to improve the accuracy and effectiveness of medical decisions.

[0056] In one implementation, such as Figure 3 As shown, this application also provides a Slicer-based deep learning fat and muscle segmentation device, comprising: The acquisition module 301 is used to acquire the preoperative CT image data of the target patient, the gene detection information of the target patient, the preset fat and muscle segmentation model based on deep learning network, and the training sample set. The processing module 302 is used to preprocess the preoperative CT image data of the target patient to generate musculoskeletal image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features; preprocess the training sample set, adjust the window width and window level, and normalize it to generate target training image data; train a preset fat and muscle segmentation model based on a deep learning network based on the target training image data to generate a target fat and muscle segmentation model; process the musculoskeletal image features, subcutaneous adipose tissue image features, and visceral adipose tissue image features based on the target fat and muscle segmentation model to generate target image segmentation information; and process the target image segmentation information and the target patient's gene detection information to generate preoperative warning information for the target patient.

[0057] The computer-readable storage medium provided in the above embodiments of this application and the Slicer-based deep learning fat and muscle segmentation method provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0058] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the Slicer-based deep learning fat and muscle segmentation method, electronic device, electronic device, and readable storage medium are relatively simple in description because they are substantially similar to the Slicer-based deep learning fat and muscle segmentation method embodiments described above. Relevant parts can be referred to the descriptions of the Slicer-based deep learning fat and muscle segmentation method embodiments described above. Although this application discloses the above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A deep learning-based fat and muscle segmentation method based on Slicer, characterized in that, include: Acquire preoperative CT image data of the target patient, genetic testing information of the target patient, a pre-set fat and muscle segmentation model based on a deep learning network, and a training sample set; Preprocessing of preoperative CT image data of target patients generates musculoskeletal imaging features, subcutaneous adipose tissue imaging features, and visceral adipose tissue imaging features. The training sample set is preprocessed, the window width and window level are adjusted and normalized to generate target training image data; The target fat and muscle segmentation model is generated by training a pre-set fat and muscle segmentation model based on a deep learning network using target training image data. Based on the target fat and muscle segmentation model, the image features of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue are processed to generate target image segmentation information. The target image segmentation information includes the segmentation results, regional surface addition, and average density of each of the skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue images. The system processes the target image segmentation information and the target patient's gene testing information to generate preoperative warning information for the target patient. This includes screening the target patient's gene testing information for genes related to fat and muscle development and metabolism to generate key gene set information; performing functional analysis on the key gene set information to generate gene function impact feature information; performing morphological analysis on the segmentation results of skeletal muscle images, subcutaneous adipose tissue images, and visceral adipose tissue images in the target image segmentation information to generate tissue morphology feature information, including tissue shape, size, and edge smoothness features; performing statistical analysis on the regional area and average density data in the target image segmentation information to generate image quantitative feature information, including area proportion and density distribution difference features; performing correlation analysis on the tissue morphology feature information, image quantitative feature information, and gene function impact feature information to generate a comprehensive impact feature vector; performing risk assessment on the comprehensive impact feature vector to generate risk level information related to the preoperative patient's health status; and generating preoperative warning information for the target patient based on the risk level information.

2. The method as described in claim 1, characterized in that, Preprocessing of preoperative CT images of the target patient generates musculoskeletal imaging features, subcutaneous adipose tissue imaging features, and visceral adipose tissue imaging features, including: Denoising the preoperative CT image data of the target patient is performed to generate low-noise CT images; Adjusting the window width and window level of low-noise CT images to generate CT images of soft tissue protrusion; Normalize CT images of soft tissue protrusions to generate normalized CT images; Normalized CT images are segmented into muscular regions to generate target muscular images. Feature extraction is performed on the target skeletal muscle image to generate skeletal muscle image features; Normalized CT images are segmented into subcutaneous adipose tissue regions to generate target subcutaneous adipose tissue images; feature extraction is performed on the target subcutaneous adipose tissue images to generate subcutaneous adipose tissue image features. The visceral adipose tissue region is segmented from the normalized CT images to generate the target visceral adipose tissue image; features are extracted from the target visceral adipose tissue image to generate visceral adipose tissue image features.

3. The method as described in claim 1, characterized in that, The training sample set is preprocessed, the window width and window level are adjusted and normalized to generate target training image data, including: The training sample set is denoised to generate low-noise CT images; the window width and window level of the low-noise CT images are adjusted to highlight soft tissue image features and generate normalized training image data. Normalized training image data is screened and labeled to identify the regions of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue, and labeled training image data is generated. The skeletal muscle region in the labeled training image data is segmented and its features are extracted to generate skeletal muscle training features. The subcutaneous adipose tissue region in the labeled training image data is segmented and its features are extracted to generate subcutaneous adipose tissue training features. The visceral adipose tissue region in the labeled training image data is segmented and its features are extracted to generate visceral adipose tissue training features. Integrate musculoskeletal training features, subcutaneous adipose tissue training features, and visceral adipose tissue training features to generate target training image data.

4. The method as described in claim 3, characterized in that, Based on the target training image data, a pre-defined fat and muscle segmentation model based on a deep learning network is trained to generate the target fat and muscle segmentation model, including: The number of features of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue in the target training image data is counted, and the sampling ratio is generated based on the distribution balance of the above features in different tissues. Based on the sampling ratio, hierarchical sampling processing is performed on the target training image data to generate a preset number of sampling feature combinations from the features of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue. Key features are selected and combined with each sampling feature for similarity measurement analysis. The data is divided into high-quality sample groups and low-quality sample groups. Each sample group contains a preset number of data samples, and at least one data sample has identification information. The pre-defined fat and muscle segmentation model based on a deep learning network is iteratively trained using high-quality and low-quality sample groups to generate the trained model and performance metrics. If the identification information in the training results is recognized by the model as a feature that distinguishes skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue, and the performance index reaches the preset threshold, then the trained model will be used as the target fat and muscle segmentation model.

5. The method as described in claim 1, characterized in that, Based on the target fat and muscle segmentation model, the image features of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue are processed to generate target image segmentation information, including: Based on the target fat and muscle segmentation model, the image features of skeletal muscle, subcutaneous adipose tissue, and visceral adipose tissue are normalized to generate normalized tissue image features. Based on the target fat and muscle segmentation model, feature extraction and recognition are performed on the normalized tissue image features to generate preliminary segmentation result features. Based on the target fat and muscle segmentation model, the preliminary segmentation results are processed to divide the region, clarify the respective regions of skeletal muscle images, subcutaneous fat tissue images and visceral fat tissue images, and generate region division features. The area of ​​the region segmentation features is calculated to obtain the region area of ​​the musculoskeletal image, subcutaneous adipose tissue image and visceral adipose tissue image, and the region area features are generated. Density calculation processing is performed on the region segmentation features to obtain the average density of skeletal muscle images, subcutaneous adipose tissue images and visceral adipose tissue images, respectively, and generate average density features. By integrating the preliminary segmentation results features, regional area features, and average density features, target image segmentation information is generated.

6. A deep learning-based fat and muscle segmentation device based on Slicer, characterized in that, For implementing the method of claim 1, the apparatus includes: The acquisition module is used to acquire the target patient's preoperative CT image data, the target patient's gene testing information, the preset fat and muscle segmentation model based on deep learning network, and the training sample set; The processing module is used to preprocess the preoperative CT image data of the target patient to generate musculoskeletal, subcutaneous, and visceral adipose tissue image features; preprocess the training sample set, adjust the window width and level, and normalize it to generate target training image data; train a preset fat-muscle segmentation model based on a deep learning network based on the target training image data to generate the target fat-muscle segmentation model; process the musculoskeletal, subcutaneous, and visceral adipose tissue image features based on the target fat-muscle segmentation model to generate target image segmentation information; and process the target image segmentation information and the target patient's gene detection information to generate preoperative warning information for the target patient.

7. An electronic device, characterized in that, include: First processor; And a memory for storing executable instructions of a first processor; wherein the first processor is configured to execute the Slicer-based deep learning fat and muscle segmentation method according to any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the Slicer-based deep learning fat and muscle segmentation method according to any one of claims 1 to 5.

Images