Muscle ultrasonic image analysis system and method for small animals

Abstract

The invention discloses a muscle ultrasonic image analysis system and method for small animals, and relates to the field of medical image analysis. The method is based on a small sample transfer learning strategy and comprises four core steps of multi-modal data acquisition, data preprocessing and registration, multi-modal fusion analysis and result output and storage, and the diagnosis precision in a small sample scene is improved through cross-modal feature fusion and semi-supervised learning. According to the method, the problems of scarcity of small animal muscle ultrasonic samples and low precision of complex pathological diagnosis are solved, sample dependence is reduced, diagnosis suitability and interpretability are improved, and the method is suitable for small animal muscle evaluation in scientific research and preclinical experiments.

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis and computer vision technology, specifically to a system and method for analyzing muscle ultrasound images of small animals. Background Technology

[0002] In medical research and preclinical experiments involving small animals (such as rats, mice, and rabbits), the morphological and functional assessment of muscle tissue is a core component in studying the mechanisms of diseases such as muscle atrophy, muscle injury, and myositis, as well as the efficacy of drugs. Ultrasound imaging, due to its non-invasive, real-time, and low-cost advantages, has become the mainstream method for small animal muscle detection. However, existing technologies suffer from two major drawbacks: First, sample scarcity and high annotation costs. Small animal muscle ultrasound samples require standardized experimental feeding and professional operation for collection, making acquisition difficult. Furthermore, sample annotation requires professional veterinarians or pathologists, resulting in generally insufficient data. Traditional deep learning models, reliant on large datasets for training, struggle to adapt. Second, insufficient single-modal information. Relying solely on the structural information of ultrasound images makes it difficult to accurately distinguish complex pathological states such as muscle fibrosis and complex muscle injuries, limiting diagnostic accuracy and failing to meet the needs of in-depth analysis of pathological mechanisms in research. Current technologies often focus on optimizing models with single-modal data, neglecting to incorporate multimodal information to improve diagnostic accuracy. Multimodal fusion solutions, on the other hand, rely on training with a large number of labeled samples, making them unsuitable for scenarios where small animal samples are scarce. Therefore, there is an urgent need for a method for analyzing small animal muscle ultrasound images that balances small sample adaptability with multimodal accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a muscle ultrasound image analysis system and method for small animals to solve the problems of scarce muscle ultrasound samples and low accuracy in complex pathological diagnosis in small animals.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A method for analyzing muscle ultrasound images in small animals is proposed. Based on a few-sample transfer learning strategy, this method improves diagnostic accuracy in small-sample scenarios through cross-modal feature fusion and semi-supervised learning, achieving precise segmentation, parameter quantification, and pathological diagnosis of small animal muscle regions. Specifically, it includes the following four steps: Step 1: Multimodal data acquisition. Two or more modalities of data from the same muscle region of a small animal are acquired, and corresponding metadata is recorded synchronously. Time-series synchronization of dynamic data is achieved through timestamps. Step 2: Data preprocessing and registration. Noise reduction, enhancement, or feature extraction preprocessing is performed on each modality of data. Spatial registration of multimodal data is achieved using a feature point matching algorithm. Step 3: Multimodal fusion analysis. Three sub-steps are executed sequentially: multimodal pre-training, few-sample fine-tuning fusion, and quantification analysis and diagnosis, to complete the effective fusion of multimodal features and pathological correlation analysis. Step 4: Result output and storage. A standardized analysis report is generated, a multimodal database is constructed to store various types of data, and the core analysis model is pruned and quantified to adapt to laboratory terminal equipment. First, we define multimodal small sample fusion and small sample transfer learning, clarify the overall technical framework and four core processes, directly address the two core pain points of "sampling scarcity" and "low accuracy of single-modal diagnosis", construct the overall logical closed loop of the technical solution, lay the technical foundation of the entire invention, clarify the collaborative relationship of each link, and ensure that the solution is not only suitable for small sample scenarios, but also improves the diagnostic accuracy through multimodal fusion.

[0005] A further embodiment includes the following steps for multimodal data acquisition: acquiring two or more modal data of the same muscle region in a small animal, and synchronously recording the corresponding metadata; achieving time-series synchronization of dynamic data through timestamps, wherein the time-series synchronization satisfies the following formula: In the formula: This is the timestamp for the acquisition of the first dynamic modal data in the nth frame. This is the timestamp for the acquisition of the nth second dynamic modal data. The timing synchronization error threshold is set to a value ranging from 0 to 10 ms; the modal data includes any combination of ultrasound images, CT images, and electromyography signals; the metadata includes information on animal species, experimental group, and detection site; and the sample size for a single modality is controlled to be 50-200 cases. The "data collection" stage is refined, specifying the combination types of multimodal data, metadata recording requirements, time-series synchronization standards, and sample size range. This addresses the issues of "insufficient information in a single modality" and "untraceable and asynchronous data," while limiting the sample size to meet core requirements. This ensures that the collected data possesses "multimodal complementarity," "time-series consistency," and "traceability." The explicit limitation on the sample size of a single modality provides a concrete basis for small-sample transfer learning, avoiding the blind pursuit of data collection.

[0006] A further embodiment includes the following data preprocessing and registration steps: performing noise reduction, enhancement, or feature extraction preprocessing on each modality of data, and using a feature point matching algorithm to achieve spatial registration of the multimodal data; wherein, image enhancement employs an adaptive histogram equalization (CLAHE) algorithm, the grayscale mapping formula of which is: In the formula: To enhance the grayscale value of the image at the (x,y) coordinates, M is the maximum grayscale level (value 255). This represents the grayscale value of the original image at the (x,y) coordinates. Let A be the histogram frequency of gray level i, A be the area of ​​the image sub-block, and L be the total number of gray levels (value 256). Focusing on "data quality optimization," this study addresses the issues of noise interference, lack of feature prominence, and spatial inconsistency in multimodal data through three main operations: noise reduction, enhancement, and registration. It also clarifies the specific formulas and parameters of the CLAHE algorithm to ensure reproducibility of the preprocessing process, thereby improving the purity and feature recognition of each modality. Spatial registration ensures accurate fusion of multimodal data, providing a high-quality data foundation for subsequent feature extraction and analysis. The quantization parameters of the CLAHE algorithm make the image enhancement effect controllable.

[0007] A further embodiment of the scheme includes three sub-steps for the multimodal fusion analysis step: multimodal pre-training, small-sample fine-tuning fusion, and quantitative analysis and diagnosis. The small-sample fine-tuning fusion sub-step employs a cross-modal contrastive learning strategy, with the loss function formula as follows: In the formula: This represents the cross-modal contrastive learning loss value, where N is the number of samples. This is the cosine similarity calculation function. Let i be the ultrasonic modal feature vector of the i-th sample. Let i be the non-ultrasonic modal feature vector of the i-th sample. This is a temperature parameter, with a value range of 0.05-0.1. This paper breaks down the "core fusion analysis" into sub-steps, focusing on clarifying the cross-modal contrastive learning strategy. It quantifies the training objective through a loss function formula, addressing the problem of "difficulty in aligning multimodal features in small sample sizes," and providing algorithmic support for the effective fusion of features from different modalities. By using contrastive learning to narrow the distance between different modal features of the same sample, it improves the correlation and consistency of feature fusion. The limited parameter range of the loss function makes the model training process controllable, avoiding feature shift in small sample scenarios.

[0008] In a further embodiment, the small-sample fine-tuning fusion sub-step also employs a semi-supervised pseudo-label strategy, with the pseudo-label generation and screening formula as follows: In the formula: The pseudo-label for the i-th unlabeled sample. The parameter set of the pre-trained model. To predict the probability that the i-th sample belongs to class k for the model, The confidence threshold is 0.7 to 0.9. Unlabeled means that no pseudo-label will be generated for this sample.

[0009] Another core strategy for supplementing small-sample fine-tuning is to generate high-quality pseudo-labels using unlabeled samples, addressing the pain point of "scarcity of labeled samples and insufficient training data." Simultaneously, a confidence threshold is used to filter pseudo-labels, ensuring the quality of the training data. This expands the amount of training data without increasing labeling costs, and the confidence threshold prevents low-quality pseudo-labels from interfering with model training, allowing for more thorough model optimization in small-sample scenarios and improving the model's generalization ability.

[0010] In a further embodiment, the quantitative analysis and diagnostic sub-steps achieve multimodal feature fusion through a Cross-Attention mechanism, and the fusion weight calculation formula is as follows: In the formula: The feature fusion weights for ultrasonic and non-ultrasonic modes are used. This is the attention weight matrix. This represents the feature vectors of modalities other than m-modality; muscle core parameters are extracted based on fused features, including at least the texture entropy value, calculated using the following formula: In the formula: H is the texture entropy value, Let L be the probability of gray level i appearing, and L be the total number of gray levels (value 256). This study clarifies the specific mechanism (Cross-Attention) and core quantification parameter (texture entropy) of multimodal feature fusion. It utilizes weighting formulas to emphasize key features and quantification parameters to objectively assess pathological conditions, addressing the issues of "blind multimodal feature fusion" and "lack of quantitative basis for diagnosis." The Cross-Attention mechanism adaptively allocates fusion weights to each modality, highlighting the contribution of key information; quantification parameters such as texture entropy provide objective data support for pathological diagnosis, enhancing the scientific rigor and credibility of diagnostic results and avoiding subjective judgment errors.

[0011] In a further embodiment, the result output and storage steps include: generating a standardized analysis report containing image annotation results, a table of quantitative parameters, and pathological diagnostic conclusions; constructing a multimodal database to store raw data, preprocessed data, analysis results, and metadata in association, supporting data traceability and model iterative optimization; pruning and quantifying the core analysis model to adapt it to laboratory terminal equipment for real-time on-site analysis.

[0012] For the "results implementation and iteration" stage, the standardization of output content, the relevance of data storage, and the lightweight requirements of the model are clearly defined, solving the problems of "non-standard results, difficulty in data traceability, and difficulty in model deployment," and forming a complete closed loop of "collection-analysis-output-iteration." Standardized reports are convenient for researchers to use directly, multimodal databases support subsequent model iteration and optimization, and lightweight deployment allows the system to adapt to laboratory scenarios for real-time analysis, improving the practicality and scalability of inventions.

[0013] A muscle ultrasound image analysis system for small animals is disclosed. The system adopts a multimodal small sample data fusion architecture, including four core units: a data acquisition module, a preprocessing and registration module, a core analysis module, and a result output and storage module. Each unit works collaboratively and, based on a small sample transfer learning strategy, achieves accurate segmentation, parameter quantification, and pathological diagnosis of muscle regions in small animals. The system also improves diagnostic accuracy in small sample scenarios through cross-modal feature fusion and semi-supervised learning.

[0014] The present invention has the following beneficial effects: This invention effectively reduces the dependence on labeled samples through small-sample transfer learning and semi-supervised learning strategies, and can complete model training without a large number of samples, which greatly reduces the cost of sample collection and labeling, and can well adapt to the needs of the sample accumulation stage in the early stage of scientific research. Meanwhile, by integrating complementary information from multiple modal data and combining attention mechanisms to strengthen the association of effective features, the diagnostic accuracy for complex pathological states has been significantly improved, which is better able to meet the needs of in-depth analysis of pathological mechanisms compared to single-modal analysis schemes. The core analysis model, after being optimized for lightweight design, can be adapted to commonly used laboratory terminal equipment, supporting real-time on-site analysis. It is also compatible with the muscle detection needs of different breeds of small animals, making it suitable for a wide range of scenarios. In addition, this method can output the contribution interpretation of each modality feature, clarify the diagnostic basis, provide strong support for the analysis of pathological mechanisms in scientific research, and improve the interpretability and credibility of diagnostic results. Attached Figure Description

[0015] Figure 1 This is a logic block diagram of the present invention.

[0016] Figure 2 This is a logic diagram of data acquisition in this invention. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention. These embodiments strictly follow the technical features defined in the claims, and by supplementing key information such as equipment parameters, operating details, quality control standards, and result verification, the abstract technical solution of the claims is made repeatable, operable, and scientific.

[0018] I. Experimental Preparation Laboratory animals and model construction Animal selection: Sixty SPF-grade SD rats (male, 6-8 weeks old, weighing 200-220g) were selected from a certain experimental animal center and were acclimatized for 1 week before the experiment.

[0019] Group design: The animals were randomly divided into a model group (n=30) and a control group (n=30). They were housed in single cages. The rearing environment was controlled with a temperature of 22±2℃, humidity of 50±5%, 12h light / 12h dark cycle, and free access to food and water.

[0020] Construction of the myofibrosis model: The model group received local injections of bleomycin (5 mg / mL, 0.1 mL / animal) into the right gastrocnemius muscle once a week for 4 consecutive weeks; the control group received an equal volume of physiological saline, with the injection site and frequency consistent with the model group. Multimodal data were collected one week after the modeling was completed to ensure stable model formation.

[0021] Experimental equipment and parameter selection Experimental ethics and quality control prerequisites This experiment strictly followed the 3R principle (replace, reduce, optimize) in animal experiments.

[0022] Equipment calibration: One week before the experiment, the ultrasound probe and CT equipment were calibrated for accuracy. The ultrasound image resolution error was ≤5% and the CT slice thickness deviation was ≤1μm. The electromyography sensor was zero-point calibrated to ensure that the baseline noise met the requirements.

[0023] Operator qualifications: All data acquisition personnel have undergone professional training and are familiar with equipment operation procedures and animal experiment standards to avoid human error.

[0024] II. Multimodal Data Acquisition This step strictly follows the core principles of "same region, multimodal synchronization, and small sample controllability," and the specific operation is as follows: 1. Animal anesthesia and fixation Isoflurane inhalation anesthesia was used (induction concentration 5%, maintenance concentration 2%). During anesthesia, the rats' blood oxygen saturation (≥95%) and heart rate (300-400 beats / min) were monitored using a pulse oximeter to ensure anesthesia safety.

[0025] Anesthetized rats were fixed prone on a constant-temperature operating table (37°C), and hair was removed from the right gastrocnemius muscle area (hair removal cream was applied for 1 minute and then rinsed with saline) to avoid hair interfering with ultrasound imaging and electromyography signal acquisition.

[0026] 2. Modal data acquisition operation Ultrasound data acquisition: After applying coupling gel to the ultrasound probe, gently place it on the surface of the right gastrocnemius muscle, adjust the probe angle to ensure the long axis of the muscle is fully displayed, and acquire 30 frames of static images (resolution 1280×720 pixels); trigger passive contraction of the rat's hind limb (by gently pulling the ankle joint), and acquire a 5-second dynamic contraction sequence (frame rate 15fps) to obtain structural change data of the muscle under movement.

[0027] CT data acquisition: The fixed rat was transferred to the Micro-CT scanning chamber, the right gastrocnemius muscle region was located as the scanning center, the scanning range was set to 5 mm above and below (covering the entire gastrocnemius muscle), and the anatomical data was acquired using the spiral scanning mode. The raw data format was DICOM.

[0028] Electromyography (EMG) data acquisition: Two EMG electrodes were attached along the direction of the right gastrocnemius muscle fibers (5 mm apart), and a reference electrode was attached to the skin of the ipsilateral ankle joint. EMG signals were acquired during the 5-second passive contraction phase of the muscle (sampling frequency 2000 Hz), and the baseline signal value (≤1 μV) was recorded simultaneously.

[0029] 3. Metadata and Time Series Synchronization Metadata recording: Metadata is synchronously entered through customized data acquisition software (developed based on Python). Fields include: animal number, breed (SD rat), sex, weight, experimental group, modeling duration, detection site (right gastrocnemius muscle), collection date, and operator ID, ensuring that the data is unique and traceable.

[0030] Timing synchronization implementation: Activate the clock calibration function of each device, and calibrate the clocks of ultrasound, CT, and electromyography equipment based on the time of the portable terminal system, with an error ≤1ms. Dynamic data synchronization employs a hardware-triggered method, synchronously initiating electromyographic signal acquisition via the "contraction trigger signal" of the ultrasound device, following a timing synchronization formula. Perform matching, set The software automatically removes data frames that exceed the error threshold. Static data (CT images, ultrasound static frames) are correlated through acquisition timestamps to ensure that the time difference of multimodal static data of the same animal is ≤30s.

[0031] 4. Sample size control and data quality control Single-modal sample size: 40 valid single-modal samples were collected from both the model group and the control group (40 each from ultrasound, CT, and electromyography), for a total of 80 cases, which is in line with the small sample size of 50-200 cases; Data quality control standards: ultrasound images should have no obvious artifacts (artifact area ≤5%), CT images should have clear muscle boundaries (grayscale contrast ≥30), and electromyography signals should have no saturation distortion (peak value ≤5mV). Samples that do not meet the standards should be re-acquired, with an acquisition failure rate ≤3%.

[0032] III. Data Preprocessing and Registration This step improves data quality through a standardized preprocessing workflow and ensures registration accuracy by using "feature point matching + error verification". The specific operations are as follows: 1. Details of preprocessing for each modal data Ultrasound image preprocessing: Noise reduction: A phase filtering algorithm (cutoff frequency 0.5MHz) was used to remove breathing and limb micro-motion artifacts. The signal-to-noise ratio (SNR) of the filtered image was ≥25dB. Enhancement processing: The CLAHE algorithm is used, with the following parameters: image sub-block area A = 64×64 pixels, number of sub-blocks 8×8, contrast threshold limited to 2.0, and grayscale mapping formula applied. Calculations were performed where M=255, L=256, and the gray level difference between the enhanced muscle and surrounding tissue was ≥20 gray levels.

[0033] CT image preprocessing: Noise reduction: Gaussian filtering (standard deviation σ=1.0) is used to smooth image noise while preserving anatomical boundaries; Gray-level normalization: CT values ​​(-1000-1000HU) are linearly mapped to the gray-level range of [0,255], using the following formula: ,in , This highlights the distinction between muscle tissue (grayscale value 80-150) and bones (grayscale value 200-255).

[0034] Electromyographic signal preprocessing: Noise reduction: The db4 wavelet basis function is used to decompose the signal into 4 layers to remove power frequency interference (50Hz) and motion artifacts, and the reconstructed signal SNR is ≥30dB. Feature extraction: Extract time-domain features (peak value, root mean square value, integral electromyography value) and frequency-domain features (average power frequency, peak frequency) to form a 12-dimensional electromyography feature vector.

[0035] Multimodal spatial registration process and quality control Registration algorithm implementation: The SIFT+RANSAC combined algorithm is adopted and implemented based on OpenCV 4.8.0. The specific steps are as follows: Feature point extraction: SIFT feature points are extracted from ultrasound enhanced images and CT normalized images respectively, with ≥500 feature points in each image; Initial matching: Feature point matching is performed using the K-nearest neighbor algorithm (K=2), with a matching threshold set to 0.7, retaining high-quality initial matching pairs; False match removal: The RANSAC algorithm is used for iterative optimization (1000 iterations, interior point threshold of 2 pixels) to remove false matches, and the final matching accuracy is ≥95%. Spatial mapping: Based on the correct matching points, the affine transformation matrix (6 degrees of freedom) is solved to map the anatomical structure of the CT image to the ultrasound image coordinate system, thereby achieving spatial alignment.

[0036] Registration effect evaluation: Mutual information (MI), mean squared error (MSE), and normalized cross-correlation coefficient (NCC) are used as evaluation indicators. The requirements are MI≥1.2, MSE≤10, and NCC≥0.85. Samples that do not meet the indicators are re-registered.

[0037] IV. Multimodal Fusion Analysis This step strictly follows the "pre-training-fine-tuning-quantization-diagnosis" process, supplementing the model structure, training parameters, and pseudo-label iteration logic. The specific operations are as follows: 1. Multimodal pre-training (basic model construction) Pre-training dataset: The publicly available human muscle multimodal dataset (MuscleNetv2.0) is used, which includes 10,000 ultrasound images, 8,000 CT images, and 5,000 sets of electromyography signals. The data have been labeled with pathological types (normal, muscle fibrosis, muscle injury).

[0038] Model structure design: Ultrasound / CT image feature extraction model: Lightweight U-Net architecture, input size 256×256 pixels, encoder contains 4 convolutional blocks (each convolutional block contains 2 layers of 3×3 convolution, BatchNorm, ReLU activation), decoder uses transposed convolution upsampling, output 64-dimensional feature vector; Electromyography (EMG) signal feature extraction model: LSTM architecture, with input being a 12-dimensional feature vector sequence (length 500), containing 2 hidden layers (128 units per layer), a Dropout rate of 0.3, and outputting a 64-dimensional feature vector.

[0039] Pre-training parameters: Optimizer is Adam (learning rate 1e-4, β1=0.9, β2=0.999), loss function is cross-entropy loss, number of iterations is 100, batch size is 32, and the feature extraction accuracy of the model on the validation set after pre-training is ≥90%.

[0040] 2. Small sample fine-tuning fusion (model adaptation and data augmentation) Data partitioning: 60 labeled samples (30 in the model group and 30 in the control group) were selected as the fine-tuning training set, and 40 unlabeled samples (20 in the model group and 20 in the control group) were selected as the semi-supervised learning set. The labeled samples were independently labeled by two senior pathologists, and the consistency test Kappa value was ≥0.85.

[0041] Cross-modal contrastive learning fine-tuning: Loss function parameters: number of samples N=60, temperature parameter τ=0.08, cosine similarity calculation uses L2 normalized dot product; Training strategy: The method of "freezing the bottom 2 layers of the encoder + fine-tuning the top layer" is adopted to avoid forgetting the pre-trained features. The optimizer is AdamW (learning rate 5e-5, weight decay 1e-5), with 50 iterations, batch size of 16, and validation every 10 rounds. The model with the lowest loss on the validation set is retained.

[0042] Semi-supervised pseudo-label generation and iteration: Initial pseudo-label generation: The pre-trained and fine-tuned model was used to predict 40 unlabeled samples with a confidence threshold of α=0.8. Samples with a confidence threshold of ≥0.8 (approximately 32 samples) were selected as pseudo-label samples. Iterative optimization: Mix the pseudo-labeled samples with the labeled samples (total sample size 92), fine-tune again (learning rate 3e-5, 30 iterations), predict the remaining 8 low-confidence samples again, update the pseudo-labels (add about 4 effective samples), and finally form 96 training samples. Pseudo-label quality control: Randomly sample the generated pseudo-label samples (sampling ratio 20%) to ensure that the accuracy of pseudo-labels is ≥90%; otherwise, adjust the confidence threshold and regenerate.

[0043] 3. Quantitative Analysis and Diagnosis (Feature Fusion and Result Output) Cross-Attention Feature Fusion: Attention weight calculation: The attention weight matrix WA is initialized as a random normal distribution (mean 0, variance 0.01) and optimized to convergence through training. During the fusion process, the weight ratio of ultrasonic modal features to non-ultrasonic modal features ranges from 0.4 to 0.6. Feature Dimension Fusion: Ultrasound (64-dimensional), CT (64-dimensional), and electromyography (64-dimensional) features are fused into a 128-dimensional comprehensive feature vector through Cross-Attention.

[0044] Core parameter quantification: Texture entropy calculation: according to the formula Calculation: L=256. pi was obtained by statistically analyzing the gray-level histogram of ultrasound images. The texture entropy value of the model group was ≥5.0, and that of the control group was ≤3.5. Other quantitative parameters: muscle thickness (vertical distance from the midpoint of the long axis of the muscle in the ultrasound image, accurate to 0.01 mm), echo intensity (mean gray value of the muscle region in the ultrasound image), and electromyographic peak frequency (frequency corresponding to the peak value of the electromyographic signal power spectrum).

[0045] Pathological diagnostic model: Based on the fusion of feature vectors and quantization parameters (a total of 131 dimensions of input), a lightweight fully connected classifier (2 hidden layers, with 64 and 32 units respectively) is constructed to output the diagnostic probability of normal / muscle fibrosis, and the diagnostic threshold is set to 0.5.

[0046] V. Result Output and Storage This step involves supplementing the report format, database design, and lightweight performance metrics. The specific steps are as follows: Standardized analysis report generation Report structure: 1. Basic sample information: animal number, group, testing date, equipment parameters, etc.; 2. Image annotation results: Ultrasound-CT registration overlay image (annotating muscle regions and fibrotic lesion locations, with red solid boxes indicating lesions, line width 2 pixels), electromyography signal waveform (annotating peak values ​​and frequency characteristics); 3. Quantitative parameter table: 4. Pathological diagnosis conclusion: Clarify the diagnosis result (e.g., "right gastrocnemius muscle fibrosis"), the contribution of each modality feature (e.g., ultrasound modality contribution 45%, CT modality 30%, electromyography modality 25%), and the diagnostic confidence (e.g., 92%).

[0047] Multimodal database construction Database type: A relational database is built using MySQL 8.0, combined with MongoDB to store unstructured data (images, signals); Data table design: Sample Information Table: Stores metadata (animal ID as primary key); Original data table: Links the original ultrasound / CT / EMG data file path and sample number; Preprocessing data table: stores preprocessed data and registration parameters; Analysis results table: stores quantitative parameters, diagnostic conclusions, and feature contribution. Data traceability: The system enables end-to-end correlation between raw data, preprocessed data, and analysis results through animal identification numbers, and supports queries based on criteria such as grouping, testing date, and diagnostic results.

[0048] Model Lightweighting and Deployment Lightweight optimization: Model pruning: Structured pruning is used to remove connections in fully connected layers with an absolute weight value ≤ 0.01, with a pruning ratio of 30%. Quantization optimization: The model parameters were quantized from 32-bit floating-point to 8-bit integer, and the model size was compressed from 28MB to 7MB after quantization. Deployment performance: The optimized model is deployed on portable computing terminals, with a single sample processing time of ≤3s (preprocessing 1s + registration 0.8s + fusion analysis 1.2s) and a diagnostic accuracy of ≥92%, meeting the needs of on-site real-time analysis in laboratories.

[0049] VI. Verification of Experimental Results Diagnostic accuracy verification Using HE staining results of pathological sections as the gold standard, all 30 cases in the model group were diagnosed with myofibrosis, while all 30 cases in the control group were normal. The diagnostic performance of the method of this invention is as follows: Accuracy: 95.0% (57 / 60); Sensitivity: 93.3% (28 / 30); Specificity: 96.7% (29 / 30); Area under the ROC curve (AUC): 0.97.

[0050] Generalization verification Twenty C57BL / 6 mice (myofibrosis model) and 20 New Zealand rabbits (normal muscle) were selected and tested using the method of this invention. The diagnostic accuracy rates were 90% and 95%, respectively, demonstrating the model's suitability for different breeds of small animals.

[0051] Comparative experiments to verify Compared with single-mode ultrasound analysis methods and multi-mode methods without transfer learning, the performance advantages of the method in this invention are as follows: VII. Experimental Precautions During data collection, the rat's body temperature needs to be maintained at 37±0.5℃ to avoid abnormal muscle contraction caused by low temperature, which would affect ultrasound and electromyography data; The pressure of the ultrasound probe in contact with the skin must be uniform (pressure ≤ 5g / cm²) to avoid excessive pressure that could cause muscle deformation. The registered images must be visually checked by the operator to ensure there are no obvious misalignments (e.g., the deviation between muscle boundaries and CT anatomical structures is ≤1 pixel). An early stopping strategy is adopted during model training (training is stopped if the validation set loss does not decrease for 10 consecutive rounds) to avoid overfitting.

[0052] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for analyzing muscle ultrasound images in small animals, characterized in that, This includes step 1: multimodal data acquisition, which involves collecting two or more modal data of the same muscle region of a small animal and synchronously recording the corresponding metadata, and using timestamps to achieve time-series synchronization of dynamic data; Step 2: Data preprocessing and registration. Denoising, enhancement, or feature extraction are performed on each modality of data. The spatial registration of multimodal data is achieved using a feature point matching algorithm. Step 3: Multimodal fusion analysis. This involves performing three sub-steps in sequence: multimodal pre-training, small-sample fine-tuning fusion, and quantitative analysis and diagnosis, to complete the effective fusion of multimodal features and pathological correlation analysis. Step 4: Results output and storage. Generate standardized analysis reports, build a multimodal database to store various types of data, and prune and quantitatively optimize the core analysis model to adapt it to laboratory terminal equipment.

2. The method for analyzing muscle ultrasound images in small animals according to claim 1, characterized in that, The multimodal data acquisition steps include: acquiring two or more modal data of the same muscle region of a small animal, and synchronously recording the corresponding metadata; achieving time-series synchronization of dynamic data through timestamps, with the time-series synchronization satisfying the following formula: In the formula: This is the timestamp for the acquisition of the first dynamic modal data in the nth frame. This is the timestamp for the acquisition of the nth second dynamic modal data. This is the timing synchronization error threshold, with a value range of 0-10ms. The modal data includes any combination of ultrasound images, CT images, and electromyography signals. The metadata includes information on animal species, experimental group, and detection site. The sample size for a single modality is controlled to be 50-200 cases.

3. The method for analyzing muscle ultrasound images in small animals according to claim 1, characterized in that, The data preprocessing and registration steps include: performing noise reduction, enhancement, or feature extraction preprocessing on each modality of data, and using a feature point matching algorithm to achieve spatial registration of multimodal data; The image enhancement employs an adaptive histogram equalization algorithm, whose grayscale mapping formula is as follows: In the formula: M is the maximum gray level (value 255) to enhance the gray value of the image at the (x,y) coordinates. This represents the grayscale value of the original image at the (x,y) coordinates. Let A be the histogram frequency of gray level i, A be the area of ​​the image sub-block, and L be the total number of gray levels.

4. The method for analyzing muscle ultrasound images in small animals according to claim 1, characterized in that, The multimodal fusion analysis step includes three sub-steps: multimodal pre-training, small-sample fine-tuning fusion, and quantitative analysis and diagnosis. The few-sample fine-tuning fusion sub-step employs a cross-modal contrastive learning strategy, and its loss function formula is as follows: In the formula: This represents the cross-modal contrastive learning loss value, where N is the number of samples. This is the cosine similarity calculation function. Let i be the ultrasonic modal feature vector of the i-th sample. Let i be the non-ultrasonic modal feature vector of the i-th sample. This is a temperature parameter, with a value range of 0.05-0.

1.

5. The method for analyzing muscle ultrasound images in small animals according to claim 4, characterized in that, The small-sample fine-tuning fusion sub-step also employs a semi-supervised pseudo-label strategy. The formula for generating and filtering pseudo-labels is as follows: In the formula: The pseudo-label for the i-th unlabeled sample. The parameter set of the pre-trained model. To predict the probability that the i-th sample belongs to class k for the model, The confidence threshold is 0.7 to 0.

9. Unlabeled means that no pseudo-label will be generated for this sample.

6. The method for analyzing muscle ultrasound images in small animals according to claim 4, characterized in that, The quantitative analysis and diagnostic sub-steps achieve multimodal feature fusion through the Cross-Attention mechanism, and the fusion weight calculation formula is as follows: In the formula: The feature fusion weights for ultrasonic and non-ultrasonic modes are used. This is the attention weight matrix. Represents the eigenvectors of modes other than the m-mode; Muscle core parameters are extracted based on fusion features, including at least the texture entropy value. The formula for calculating the texture entropy value is as follows: In the formula: H is the texture entropy value, Let L be the probability of gray level i appearing, and L be the total number of gray levels.

7. The method for analyzing muscle ultrasound images in small animals according to claim 1, characterized in that, The result output and storage steps include: generating a standardized analysis report containing image annotation results, a quantitative parameter table, and pathological diagnosis conclusions; A multimodal database is constructed to store raw data, preprocessed data, analysis results, and metadata, supporting data traceability and model iteration optimization; the core analysis model is pruned and quantitatively optimized to adapt to laboratory terminal equipment for real-time on-site analysis.

8. A system for analyzing the ultrasound images of muscles in small animals to implement the analysis method according to any one of claims 1-7, characterized in that, The system adopts a multimodal small sample data fusion architecture, including a data acquisition module, a preprocessing and registration module, a core analysis module, and a result output and storage module.

Images