A deep learning-based multi-source feature fusion method for sarcopenia ultrasound MRC automatic grading

Abstract

The application provides a sarcopenia ultrasound MRC automatic grading method based on deep learning multi-source feature fusion. The method comprises the following steps: constructing a multi-dimensional data set, wherein the multi-dimensional data set comprises an ultrasound image, an anatomical label, a morphological parameter and an MRC grading label; processing data in the multi-dimensional data set to obtain multi-dimensional input data; constructing an MRC grading model, wherein the MRC grading model comprises a data processing module, a feature extraction module and a feature fusion and classification module; training the MRC grading model based on the multi-dimensional input data; and processing to-be-processed data by using the trained MRC grading model to obtain a sarcopenia ultrasound MRC automatic grading result. The technical scheme provided by the application can realize objective, stable and repeatable automatic grading of the muscle strength level of sarcopenia.

Description

Technical Field

[0001] This application belongs to the field of medical image processing technology, and relates to an automatic ultrasound MRC grading method for sarcopenia, and in particular to an automatic ultrasound MRC grading method for sarcopenia based on deep learning multi-source feature fusion. Background Technology

[0002] Aging, prolonged bed rest, inflammatory responses associated with critical illness, muscle loss, and ICU-acquired weakness (ICU-AW) are all associated with this condition. ICU-AW-induced muscle loss is a common cause of sarcopenia. Decreased muscle strength is considered a core indicator for the diagnosis of sarcopenia, and Manual Muscle Testing (MMT) and its quantitative system, the Medical Research Council (MRC) five-level grading system, remain the most commonly used methods for clinically assessing patient muscle strength. However, traditional MRC grading is highly dependent on the assessor's experience and is subject to interference from factors such as assessment perspective, patient compensatory movements, and differences in physician subjective judgment. Its inter-assessor consistency is limited, with reported inter-rater reliability (Kappa value) generally only between 0.4 and 0.65, failing to meet the clinical need for objective consistency and quantitative assessment. Summary of the Invention

[0003] This application provides an automatic ultrasound MRC grading method for sarcopenia based on deep learning multi-source feature fusion, which is used to achieve objective, stable, and repeatable automated grading of muscle strength levels in sarcopenia.

[0004] In a first aspect, embodiments of this application provide an automatic ultrasound MRC grading method for sarcopenia based on deep learning multi-source feature fusion. The method includes: constructing a multi-dimensional dataset, the multi-dimensional dataset comprising ultrasound images, anatomical labels, morphological parameters, and MRC grading labels; processing the data in the multi-dimensional dataset to obtain multi-dimensional input data; constructing an MRC grading model, the MRC grading model including a data processing module, a feature extraction module, and a feature fusion and classification module; training the MRC grading model based on the multi-dimensional input data; and using the trained MRC grading model to process the data to be processed to obtain automatic ultrasound MRC grading results for sarcopenia.

[0005] In some implementations, constructing a multi-dimensional dataset includes: acquiring ultrasound images of the bilateral biceps brachii, vastus intermedius, and / or rectus femoris muscles using an ultrasound acquisition device, wherein the ultrasound images include transverse and longitudinal images of the muscles; standardizing the ultrasound images and inputting the processed ultrasound images into a pre-trained muscle segmentation model to obtain a probability map of the muscle region; binarizing the probability map of the muscle region and filling holes in the processed image to obtain the largest connected region as the region of interest.

[0006] In some implementations, the method further includes: determining the collection object based on a preset data collection standard, wherein the data collection standard includes inclusion criteria and exclusion criteria.

[0007] In some implementations, the morphological parameters include muscle thickness and cross-sectional area. Constructing a multi-dimensional dataset further includes: for the longitudinally slicing image, obtaining the maximum distance along the vertical direction of the muscle fibers within the region of interest as the muscle thickness; for the transversely slicing image, obtaining the product of the total number of pixels and the distance between pixels within the region of interest as the cross-sectional area, where the distance between pixels is the pixel spacing.

[0008] In some implementations, constructing a multidimensional dataset also includes: performing data augmentation on images of regions of interest in the training set; and / or standardizing the morphological parameters.

[0009] In some implementations, the data augmentation includes random horizontal flipping, rotation, brightness adjustment, and / or contrast adjustment.

[0010] In some implementations, the multi-dimensional input data includes a dense vector corresponding to the muscle type label, a dense vector corresponding to the imaging direction label, and morphological parameters selected according to the imaging direction.

[0011] In some implementations, the feature extraction module is built based on a pre-trained deep learning model, freezing the parameters of some layers during training while training only the parameters of the remaining layers; and / or the data processing module uses a two-level fully connected network to process numerical features; and / or the feature fusion and classification module concatenates image features, label embedding features, and parameter features, and outputs hierarchical probabilities through a classification network, and obtains the prediction probabilities of each level through a softmax function.

[0012] In some implementations, the method further includes: evaluating and validating the trained MRC grading model.

[0013] In some implementations, the evaluation metrics for assessing the trained MRC grading model include accuracy, precision, recall, and / or F1 score.

[0014] As described above, this application provides an automatic ultrasound MRC grading method for sarcopenia based on deep learning multi-source feature fusion, which addresses the following problems existing in current MRC grading technologies:

[0015] (1) The grading is highly subjective: Traditional MRC often relies on doctors’ manual assessment for grading, lacks objective quantitative indicators, and has low consistency among different doctors’ assessments, which can easily lead to misjudgment.

[0016] (2) Insufficient use of information: Existing automatic grading methods mostly use only the visual features of ultrasound images, without combining anatomical labels such as muscle type and imaging direction, as well as morphological parameters such as muscle thickness and cross-sectional area, resulting in an incomplete characterization of sarcopenia pathological features by the model.

[0017] (3) Insufficient feature adaptability: Existing technologies do not consider the correlation between ultrasound imaging direction and morphological parameters (longitudinal section corresponds to thickness, transverse section corresponds to area), and adopt a fixed feature input mode, which cannot give full play to the clinical diagnostic value of different parameters.

[0018] (4) Weak generalization ability of the model: Most existing deep learning models are designed for single muscle groups or fixed imaging patterns, and have not adapted to the anatomical differences of key muscle groups in sarcopenia (biceps brachii, vastus intermedius, rectus femoris), resulting in unstable grading performance in different clinical scenarios.

[0019] To address the aforementioned issues, this application innovatively integrates ultrasound image features, muscle anatomy labels (type, direction), and morphological parameters (muscle thickness, cross-sectional area) in its embodiments. It also dynamically selects appropriate morphological parameters based on the imaging direction, fully leveraging the clinical diagnostic value of information from each dimension. This approach solves the problems of insufficient information utilization and inadequate feature adaptability in existing technologies.

[0020] This application proposes an integrated model that includes "image feature extraction - label embedding - parameter processing - feature fusion". It realizes the learnable representation of discrete labels through the embedding layer, enhances the expressive power of numerical parameters through a dedicated fully connected module, and combines the fine-tuning strategy of the pre-trained backbone network to balance hierarchical accuracy and training efficiency.

[0021] In this application embodiment, a sarcopenia ultrasound dataset covering key muscle groups, dual imaging directions, and multi-dimensional labels was established. By formulating unified standards for acquisition, segmentation, annotation, and preprocessing, high-quality data support can be provided for model training and clinical promotion, which is conducive to improving the reproducibility and generalization ability of the technology. Attached Figure Description

[0022] Figure 1 The diagram shown is a structural schematic of an electronic device in an embodiment of this application.

[0023] Figure 2 The flowchart shown is a process for an automated ultrasound MRC grading method for sarcopenia provided in an embodiment of this application.

[0024] Figure 3A The flowchart shown is for obtaining the region of interest in this embodiment of the application.

[0025] Figure 3B The diagram shown is an architecture diagram of nnUNet in the embodiments of this application.

[0026] Figure 4 The diagram shown is an architecture diagram of the MRC hierarchical model in this application embodiment.

[0027] Component designation explanation

[0028] 1 electronic devices 11 memory 12 monitor 13 processor 131 CPU 132 Neural Network Processor S21~S25 step S31~S32 step Detailed Implementation

[0029] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0030] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the illustrations only show the components related to this application and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0031] In this application, the terms "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0032] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0033] Muscle ultrasound is non-invasive, real-time, portable, and can be performed at the bedside. It can be used to assess quantitative parameters such as muscle structure, muscle fiber arrangement, muscle thickness, and cross-sectional area. The inventors discovered that ultrasound can display the cross-sectional area (CSA) and muscle thickness (TH) of muscles in transverse and longitudinal sections, respectively, both of which are closely related to sarcopenia-related structural changes. However, ultrasound-based sarcopenia assessment techniques have several limitations: First, most methods still rely on manual measurement of muscle thickness or cross-sectional area, lacking an objective mapping relationship directly corresponding to MRC grading; second, different muscle groups (such as biceps brachii, vastus intermedius, and rectus femoris) have differentiated diagnostic value under different imaging orientations. For example, longitudinal imaging is more suitable for thickness measurement, while transverse imaging better reflects the overall degree of atrophy, but existing models rarely utilize this directionality and anatomical difference; third, ROI (Region of Interest) acquisition methods often rely on manual or simple threshold-based image processing, which has limited stability when faced with noise from different devices, speckle interference, and individual differences, affecting subsequent feature extraction and grading results; fourth, most deep learning model methods rely only on image texture features and do not integrate multi-source information such as muscle anatomical labels (muscle type, imaging orientation) and morphological parameters (TH, CSA), making it difficult to fully reflect the complex pathological features of sarcopenia at the structural, morphological, and functional levels.

[0034] To address some or all of the aforementioned problems, this application provides an automated ultrasound MRC grading method for sarcopenia based on deep learning-based multi-source feature fusion. This method integrates multi-dimensional information processing methods from ultrasound images, anatomical labels, and morphological parameters to achieve objective, stable, and repeatable automated grading of sarcopenia muscle strength levels (e.g., MRC 1-5), thereby compensating for the lack of consistency in subjective assessments and improving the objective quantitative capabilities of sarcopenia screening and rehabilitation follow-up. Furthermore, based on automatic ultrasound image segmentation, this application introduces muscle type, imaging direction, and matching morphological parameters to construct a clinically significant comprehensive feature expression system. This provides theoretical and technical support for the intelligent quantitative assessment of sarcopenia and is of great significance for improving grading accuracy, reducing subjective errors, and promoting standardized diagnosis of sarcopenia.

[0035] The automatic ultrasound MRC grading method for sarcopenia based on deep learning multi-source feature fusion provided in this application embodiment can, for example, be applied to... Figure 1 Electronic device 1 is shown. (e.g.) Figure 1 As shown, electronic device 1 includes a memory 11, a display 12, and at least one general-purpose processor 13.

[0036] The memory 11 may include volatile memory, such as random access memory (RAM) or cache. The memory 11 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD). The memory 11 can be used to store program instructions for a processor to call and execute at least a portion of the automatic ultrasound MRC grading method for sarcopenia provided in the embodiments of this application.

[0037] The display 12 may specifically include a display screen (display panel). In some implementations, the display panel may be configured using a liquid crystal display (LCD), an organic light-emitting diode (OLED), or other similar forms. Alternatively, the display 12 may also be a touch panel (touchscreen, touch screen), which may include a display screen and a touch-sensitive surface. When the touch-sensitive surface detects a touch operation on or near it, it transmits the information to a general-purpose processor 13 to determine the type of touch event. Subsequently, the general-purpose processor 13 provides corresponding visual output on the display device based on the type of touch event.

[0038] The general-purpose processor 13 can be any type of device capable of processing electronic instructions. In this embodiment, the electronic device 1 may include one or more general-purpose processors 13, such as one or both of a central processing unit (CPU) 131 and a neural-network processing unit (NPU) 132. Furthermore, it may include one or more of a graphics processing unit (GPU), microprocessor, microcontroller, main processor, controller, and ASIC (Application Specific Integrated Circuit). The general-purpose processor 13 is configured to execute various types of digitally stored instructions, such as software or firmware programs stored in memory 11, enabling the electronic device 1 to provide various services. For example, the processor 11 can execute programs or process data to perform at least a portion of the automatic ultrasound MRC grading method for sarcopenia provided in this embodiment.

[0039] The main function of the central processing unit 131 is to parse computer instructions and process data in computer software, realize the overall control of electronic device 1, and control and allocate all hardware resources of electronic device 1 (such as storage resources, communication resources, I / O interfaces, etc.).

[0040] Neural network processors are a general term for new types of processors based on neural network algorithms and acceleration. They are specifically designed for artificial intelligence and are used to accelerate neural network operations, solving the problem of low efficiency of traditional chips in neural network operations.

[0041] It should be noted that the name of the neural network processor does not constitute a limitation of this application. For example, in other application scenarios, the neural network processor can also be modified or replaced by other processors with similar functions, such as a tensor processing unit (TPU), a deep learning processing unit (DPU), and so on.

[0042] Figure 2 The flowchart shown is a process for an automated ultrasound MRC grading method for sarcopenia provided in an embodiment of this application. Figure 2 As shown, the method includes the following steps S21 to S25.

[0043] S21. Construct a multi-dimensional dataset. This multi-dimensional dataset includes ultrasound images, anatomical labels, morphological parameters, and MRC grading labels.

[0044] S22 processes the data in the multidimensional dataset to obtain multidimensional input data.

[0045] S23, Construct the MRC hierarchical model. This MRC hierarchical model includes a data processing module, a feature extraction module, and a feature fusion and classification module.

[0046] S24, The MRC grading model is trained based on multi-dimensional input data.

[0047] S25. The trained MRC grading model is used to process the data to obtain the automatic ultrasound MRC grading results for sarcopenia.

[0048] In some implementations, constructing a multidimensional dataset involves: collecting ultrasound images of patients with sarcopenia, extracting the ROI regions of key muscle groups (such as biceps brachii, vastus intermedius, and rectus femoris) from the ultrasound images, labeling the imaging direction (transverse / longitudinal section) and muscle type labels, and measuring the corresponding morphological parameters (TH, CSA). Based on the above data, a complete dataset containing "image-label-parameter-grading" is established.

[0049] For example, ultrasound images of patients with sarcopenia can be collected according to specific data collection criteria. These criteria may include inclusion and exclusion criteria.

[0050] Inclusion criteria include, for example: ① Age 60-90 years; ② Meeting the definition of "possible sarcopenia", i.e., having decreased muscle strength (e.g., grip strength: male <27 kg, female <16 kg) or prolonged 5 sit-to-stand tests (≥15 s); ③ Being able to cooperate in completing ultrasound examinations and adjusting standing / sitting positions; ④ Having complete clinical data, including basic information, muscle strength assessment, gait speed or functional assessment data; ⑤ Agreeing to participate in the study and signing an informed consent form.

[0051] Exclusion criteria include, for example: ① a clear diagnosis of a disease affecting the structure of the muscle itself, such as muscular dystrophy, sequelae of poliomyelitis, peripheral neuropathy, severe inflammatory myopathy, etc.; ② a history of severe muscle injury, fracture, muscle rupture, or related surgery within the past 6 months; ③ significant edema, subcutaneous fluid accumulation, or large amounts of fat deposition in the limbs, which may affect ultrasound penetration and image quality; ④ a previous diagnosis of end-stage severe cardiopulmonary disease, advanced malignant tumor, or inability to cooperate with the examination; ⑤ use of drugs that significantly affect muscle metabolism (such as long-term high-dose corticosteroids); ⑥ suspected severe motor neuron disease or difficulty in achieving a stable body position; ⑦ ultrasound image quality that does not meet the analysis requirements (such as scanning artifacts, image defocus, blurred muscle boundaries).

[0052] For example, high-resolution ultrasound diagnostic instruments or other ultrasound image acquisition devices can be used to acquire ultrasound images. The instrument parameters can be set to, for example, constant values ​​such as: grayscale gain of 70 dB, depth of focus of approximately 3.0 cm, thermal index of 0.0, and mechanical index of 1.3. During acquisition, the ultrasound probe should be in light contact with the skin and perpendicular to the bone cortex to ensure stable probe position during scanning, avoiding pressure that could cause muscle deformation and thus reducing errors.

[0053] For example, the acquisition range includes: bilateral biceps brachii (midpoint of the muscle belly when the elbow is flexed at 90°), vastus intermedius (10cm above the patella), and rectus femoris (midpoint of the line connecting the anterior superior iliac spine and the patella). For each muscle, several frames of transverse (perpendicular to the long axis of the muscle fiber) and longitudinal (parallel to the long axis of the muscle fiber) images are acquired, and the frames with the best quality are selected for inclusion in the dataset.

[0054] Please see Figure 3A In some implementations, extracting the ROI region of key muscle groups based on ultrasound images includes the following steps S31 and S32.

[0055] S31, standardize the ultrasound image and input the processed ultrasound image into the pre-trained muscle segmentation model to obtain the probability map of the muscle region.

[0056] For example, in step S31, the ultrasound image can be standardized to normalize the pixel values ​​to [0, 1], and salt-and-pepper noise can be removed by methods such as Gaussian filtering. The processed image can then be input into a pre-trained muscle segmentation model to obtain a probability map of the muscle region.

[0057] For example, the pre-trained muscle segmentation model is, for instance, the nnUNet segmentation model, whose architecture diagram is shown below. Figure 3B As shown.

[0058] S32, binarize the probability map of the muscle region, fill holes in the processed image, and obtain the largest connected region as the region of interest.

[0059] For example, a 0.5 threshold can be used to binarize the probability map, and holes can be filled using morphological closing operations, retaining the largest connected component as the final ROI to ensure the integrity of the ROI boundary. The binarization formula is as follows:

[0060] ;

[0061] in, Represents pixels The probability of belonging to the muscle region.

[0062] In some implementations, constructing a multi-dimensional dataset may also include labeling muscle types and imaging orientations. For example, the labeled muscle types might be... Where 0, 1, and 2 correspond to the biceps brachii, vastus intermedius, and rectus femoris muscles, respectively. The labeled imaging direction is, for example,... , where 0 and 1 correspond to the transverse and longitudinal directions, respectively.

[0063] In some implementations, morphological parameters include muscle thickness and cross-sectional area. Constructing a multidimensional dataset can also include calculating parameters within the ROI region using the following formula:

[0064] ;

[0065] ;

[0066] in, and These represent muscle thickness and cross-sectional area, respectively. Indicates the region of interest. The pixel pitch (mm / pixel) can be obtained from the parameters of the ultrasound image acquisition device. This represents the total number of pixels within the region of interest. According to the formula above, the muscle thickness in this embodiment is... Equal to the maximum distance along the vertical direction of the muscle fiber in the longitudinal image, cross-sectional area The total number of pixels in the region of interest in the cross-section image and The product of.

[0067] In some implementations, the MRC classification can be determined by relevant medical personnel based on manual muscle testing (MMT). The target value can be used as a training target for the model, but this application is not limited to it.

[0068] In some implementations, constructing a multidimensional dataset may also include: performing data augmentation on images of regions of interest in the training set, represented as... ,in The transformation set can include any combination of one or more of the following transformations: random horizontal flipping, rotation, brightness adjustment, and contrast adjustment. It should be understood that the above data augmentation methods are merely examples of several feasible methods in the embodiments of this application, but this application is not limited thereto.

[0069] In some implementations, constructing a multidimensional dataset also includes standardizing the morphological parameters.

[0070] For example, the morphological parameters can be Z-score normalized: ,in, For the original parameters (TH or CSA). and The mean and standard deviation of the training set parameters.

[0071] In some implementations, the multi-dimensional input data includes dense vectors corresponding to muscle type labels, dense vectors corresponding to imaging direction labels, and corresponding morphological parameters selected according to the imaging direction.

[0072] For example, integer encoding (0-2) can be used to embed muscle type labels through an embedding layer. The mapping to the corresponding dense vector is represented as follows: ,in The learnable embedding matrix has a size of, for example, 3×8. This is the dense vector obtained after mapping.

[0073] For example, integer encoding (0-1) can be used to embed the imaging orientation label through the embedding layer. The mapping to the corresponding dense vector is represented as follows: ,in The learnable embedding matrix has a size of, for example, 2×8.

[0074] For example, corresponding morphological parameters can be selected based on the imaging direction to add multi-dimensional input data, as shown below:

[0075] .

[0076] After standardizing the selected morphological parameters, they can be used as numerical features input into the MRC grading model.

[0077] In some implementations, morphological parameters can be dynamically selected based on the imaging direction (longitudinal images match muscle thickness, and transverse images match cross-sectional area), with the ROI image, muscle type label, imaging direction label, and adapted morphological parameters used as input parameters for the model.

[0078] Figure 4 This is a diagram showing the architecture of the MRC hierarchical model in some implementations. For example... Figure 4 As shown, the MRC hierarchical model includes a data processing module, a feature extraction module, and a feature fusion and classification module, which is an end-to-end "feature extraction-processing-fusion-classification" architecture.

[0079] The feature extraction module is built based on a pre-trained deep learning model, freezing the parameters of some layers during training while training only the parameters of the remaining layers. For example, a pre-trained ResNet18 can be used as the backbone for image feature extraction, and a label embedding module and a morphological parameter processing module can be designed. Multi-dimensional information is integrated through a feature fusion network to output MRC classification results.

[0080] For example, the feature extraction module can be built based on a pre-trained ResNet18, retaining the first 17 convolutional and pooling layers (removing the original fully connected layers), and outputting 512-dimensional features. Specifically, a 224×224×3 ROI image can be normalized and input into the feature extraction module. The feature extraction module can extract features through the convolutional layers of ResNet18, and the final output is: .

[0081] For example, in the embodiments of this application, a parameter fine-tuning strategy can be adopted to freeze the parameters of the first 14 layers (retaining the ImageNet pre-trained weights) and train only the parameters of the last 3 convolutional layers to balance feature extraction capability and training efficiency.

[0082] The data processing module can use a two-level fully connected network to process numerical features, as shown in the following formula:

[0083] ;

[0084] in, For learnable weight matrix, This is the activation function.

[0085] The feature fusion and classification module concatenates image features, label embedding features, and parametric features, outputs hierarchical probabilities through a classification network, and obtains the prediction probabilities of each level through a softmax function.

[0086] For example, the feature fusion and classification module can concatenate image features, label embedding features, and parameter features into a 544-dimensional vector, represented as: .

[0087] For example, the feature fusion and classification module can output hierarchical probabilities through a fully connected layer with dropout, as shown below:

[0088] ;

[0089] in, The dropout rate is 0.5. The dropout rate is 0.3%. This corresponds to a 5-level classification.

[0090] For example, the feature fusion and classification module can obtain the prediction probabilities at each level through the softmax function, as shown below:

[0091] .

[0092] In some implementations, a hierarchical partitioning strategy can be used to construct training, validation, and test sets. The model parameters can be optimized using the cross-entropy loss function, the Adam optimizer, and regularization techniques to ensure hierarchical accuracy and generalization ability.

[0093] For example, when training the MRC classification model, the cross-entropy loss function can be used to adapt to five classification tasks. The formula for the loss function is: ,in One-hot encoding for the real label. =1 indicates that the actual level is k, otherwise it is 0.

[0094] For example, the Adam optimizer can be used for parameter updates, with the following parameter update rules:

[0095] ;

[0096] ;

[0097] ;

[0098] in, For model parameters, , For first- and second-order momentum estimation, For example, 0.9, For example, 0.999. For example, 1e-8, initial learning rate For example, 1e-4. It should be noted that the values ​​of the above parameters are merely examples of embodiments of this application, and this application does not limit the values ​​of these parameters.

[0099] For example, when training an MRC hierarchical model, the number of training epochs and batches can be set according to actual needs, and gradient clipping (maximum norm 10) can be used during training to prevent gradient explosion.

[0100] For example, in the embodiments of this application, the accuracy can be calculated on the validation set after each round of training, and the model weights corresponding to the highest accuracy can be saved.

[0101] For example, in this embodiment of the application, if the validation set loss does not decrease for three consecutive rounds, the learning rate can be multiplied by 0.1 to achieve learning rate decay.

[0102] For example, the automatic ultrasound MRC grading method for sarcopenia provided in the embodiments of this application may further include: evaluating and validating the trained MRC grading model to assess the model's performance on an independent test set and verify its clinical applicability.

[0103] For example, evaluation metrics may include accuracy, precision, recall, and / or F1 score.

[0104] For example, the formula for calculating accuracy is:

[0105] ;

[0106] For example, the formula for calculating accuracy is:

[0107] ;

[0108] The formula for calculating recall rate is, for example:

[0109] ;

[0110] The formula for calculating the F1 score is as follows:

[0111] ;

[0112] in, The total number of samples, For prediction classification, For indicator functions, , and These represent the number of true positive samples, false positive samples, and false negative samples at level k, respectively.

[0113] The scope of protection of the automatic ultrasound MRC grading method for sarcopenia provided in this application is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0114] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the automatic ultrasound MRC grading method for sarcopenia provided in this application. Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing a processor. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof. The storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0115] This application also provides an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program; the processor is communicatively connected to the memory, and when the computer program is invoked, it executes the automatic ultrasound MRC grading method for sarcopenia provided in this application.

[0116] This application embodiment may also provide a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application embodiment are generated. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0117] When the computer program product is executed by a computer, the computer performs the method described in the foregoing method embodiments. The computer program product can be a software installation package; when the foregoing method is required, the computer program product can be downloaded and executed on the computer.

[0118] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0119] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0120] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0121] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0122] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0123] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0124] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0125] In summary, this application provides an automatic ultrasound MRC grading method for sarcopenia based on deep learning multi-source feature fusion. By fusing multi-dimensional information to reduce the limitations of single features, the model can more comprehensively capture the pathological features of sarcopenia, significantly improving grading accuracy compared to existing methods, with a recall rate of ≥90% for key grades (grades 1 and 5).

[0126] Furthermore, in this embodiment, the grading is driven by objective images, labels, and parameter data, replacing the traditional subjective manual assessment. This helps to improve the consistency and reliability of the grading results, and the Kappa value of inter-doctor assessment consistency is increased from 0.65 to over 0.85.

[0127] This application also provides a dynamic adaptation mechanism for imaging direction and morphological parameters, which conforms to the actual application scenarios of clinical ultrasound diagnosis, and is conducive to improving the practicality and scalability of the technology. It can be directly integrated into existing ultrasound diagnostic equipment.

[0128] The MRC grading model provided in this application can automatically complete multi-dimensional information processing and grading, with a single image grading time of ≤0.5 seconds, which can significantly shorten the diagnosis time, reduce the workload of doctors, and assist in early screening and large-scale population screening for sarcopenia.

[0129] The customized design and data augmentation strategy for key muscle groups in this application embodiment enable the model to work stably on ultrasound images acquired from different hospitals and different devices, adapting to diverse clinical environments, with a generalization error ≤8%.

[0130] Therefore, this application effectively overcomes the various shortcomings of the prior art and has high industrial application value.

[0131] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. An automatic ultrasound MRC grading method for sarcopenia based on deep learning multi-source feature fusion, characterized in that, The method includes: A multi-dimensional dataset is constructed, which includes ultrasound images, anatomical labels, morphological parameters, and MRC grading labels; The data in the multi-dimensional dataset is processed to obtain multi-dimensional input data; Construct an MRC hierarchical model, which includes a data processing module, a feature extraction module, and a feature fusion and classification module; The MRC grading model is trained based on the multi-dimensional input data; The trained MRC grading model is used to process the data to obtain automatic ultrasound MRC grading results for sarcopenia.

2. The method according to claim 1, characterized in that, Constructing a multidimensional dataset includes: Ultrasound images of the bilateral biceps brachii, vastus intermedius, and / or rectus femoris muscles are acquired using an ultrasound acquisition device, wherein the ultrasound images include transverse and longitudinal images of the muscles. The ultrasound image is standardized and then input into a pre-trained muscle segmentation model to obtain a probability map of the muscle region. The probability map of the muscle region is binarized, and the processed image is filled with holes to obtain the largest connected region as the region of interest.

3. The method according to claim 2, characterized in that, The method further includes: determining the collection object based on a preset data collection standard, wherein the data collection standard includes an inclusion standard and an exclusion standard.

4. The method according to claim 2, characterized in that, The morphological parameters include muscle thickness and cross-sectional area. Constructing the multi-dimensional dataset also includes: For the longitudinal section image, the maximum distance along the vertical direction of the muscle fiber is obtained within the region of interest as the muscle thickness; For the cross-sectional image, the total number of pixels in the region of interest is obtained. The product of is the cross-sectional area, where This refers to the pixel spacing.

5. The method according to claim 4, characterized in that, Building multidimensional datasets also includes: Data augmentation of images in regions of interest within the training set; and / or The morphological parameters are standardized.

6. The method according to claim 5, characterized in that, The data enhancements include random horizontal flipping, rotation, brightness adjustment, and / or contrast adjustment.

7. The method according to claim 1, characterized in that, The multi-dimensional input data includes dense vectors corresponding to muscle type labels, dense vectors corresponding to imaging direction labels, and morphological parameters selected according to the imaging direction.

8. The method according to claim 1, characterized in that, The feature extraction module is built based on a pre-trained deep learning model, which freezes the parameters of some layers during training and trains only the parameters of the remaining layers; and / or The data processing module employs a two-level fully connected network to process numerical features; and / or The feature fusion and classification module concatenates image features, label embedding features, and parameter features, outputs hierarchical probabilities through a classification network, and obtains the prediction probabilities of each level through a softmax function.

9. The method according to claim 1, characterized in that, The method further includes: evaluating and validating the trained MRC grading model.

10. The method according to claim 9, characterized in that, Evaluation metrics for assessing the trained MRC grading model include accuracy, precision, recall, and / or F1 score.

Images