Bone injury ultrasonic detection method and equipment based on multi-scale features and machine learning

Abstract

The invention discloses a bone injury ultrasonic detection method and equipment based on multi-scale features and machine learning, and the method comprises the steps: collecting a UGW signal, carrying out the standardization and envelope processing of the UGW signal, removing the signal noise, and keeping the key injury information. Feature extraction in multiple signal processing fields such as time domain, frequency domain and time-frequency domain is combined, and for multi-scale characteristics and complexity of bone injury signals, complementary set empirical mode decomposition energy features, energy entropy features and ASDDF features are provided. In order to ensure scale unification between features and eliminate redundant features, a genetic algorithm is adopted to perform feature selection. And finally, inputting multiple features subjected to GA optimization into the designed full-connection neural network model, and training through a K-fold cross validation method to obtain an efficient bone injury automatic identification model. The detection accuracy and adaptability of the method to bone injury in a complex environment are better than those of a traditional method, and an efficient and accurate solution is provided for early screening of osteoporosis and fracture.

Description

Technical Field

[0001] The present invention relates to the field of UGW non-destructive testing technology, and in particular to a UGW bone damage detection method and device based on multi-scale features and machine learning. Background Art

[0002] Osteoporosis is a skeletal disease characterized by decreased bone mass and deterioration of bone microarchitecture, leading to increased bone fragility and a significantly elevated risk of fracture. With the aging of the global population, the prevalence of osteoporosis and the risk of fractures in people over 50 years of age have increased significantly. Osteoporosis not only significantly reduces patients' quality of life but also increases the risk of disability due to fracture complications and contributes to a greater socioeconomic burden. Therefore, strengthening the prevention, early diagnosis, and management of osteoporosis is key to improving individual health and alleviating the societal burden. Currently, bone mineral density (BMD) measurement is the standard method for assessing bone health, with dual-energy X-ray absorptiometry (DXA) being the most commonly used technique. However, DXA technology has numerous limitations, including high cost, lack of portability, and potential ionizing radiation exposure. Furthermore, DXA faces challenges in assessing cortical bone, making it difficult to fully assess bone microstructure and geometry. As an alternative, conventional non-destructive ultrasound testing has attracted attention due to its lack of ionizing radiation, portability, and relative low cost. However, traditional ultrasound testing is usually limited to local measurements, making it difficult to accurately assess the geometry and microstructure of bones, limiting its widespread clinical application. Patent 1 (patent number CN97227129.5, entitled "X-ray bone density tester") and patent 2 (patent number CN99203394.2, entitled "Portable X-ray hand bone image converter") represent important advances in bone injury detection technology, providing solutions based on X-ray imaging. Although these patents have made significant achievements in bone density measurement, they still face some challenges and limitations in practical applications. These challenges include the risk of radiation exposure, high equipment costs, insufficient sensitivity for early bone injury detection, and application adaptability issues in diverse clinical settings.

[0003] In addition, most existing feature extraction methods rely on traditional features that fail to be optimized for the unique characteristics of the test object, resulting in a limited range of feature selection. The complexity of bone health status means that a single feature (such as sound velocity or attenuation) cannot fully reflect the overall condition of the bone. In the field of machine learning, support vector machines (SVMs) are widely used in signal classification tasks, especially in UGW signal analysis. However, SVMs often face overfitting problems when processing small samples and complex signals, especially in the case of high noise or complex feature sets, their performance may drop significantly, which limits the application effect of existing technologies in bone injury detection. Summary of the Invention

[0004] The ultrasonic detection method for bone damage based on multi-scale features and machine learning proposed in the present invention can solve at least one of the above technical problems.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a multi-scale feature and machine learning-based ultrasound detection method for bone damage, comprising the following steps:

[0007] Use UGW to collect original bone sample signals, normal signals and damage signals respectively, and convert the signals into digital format;

[0008] Perform Z-score normalization and envelope processing on the collected signals;

[0009] Perform multi-scale feature extraction on the processed signal;

[0010] After feature extraction, feature selection is performed based on GA;

[0011] Establish the MFI-FCNN classifier model and perform K-fold cross-validation training optimization;

[0012] The optimized MFI-FCNN classifier model is used to judge whether the processed guided wave signal is damaged.

[0013] Furthermore, the collected signal is subjected to Z-score normalization and envelope processing, including the following steps:

[0014] The formula for the first step of preprocessing is as follows:

[0015]

[0016] Where y(m) is the original waveguide signal, z′(m) is the signal after the first step of preprocessing, L represents the length of y(m), and the normalized signal is then Hilbert transformed.

[0017] The second step of preprocessing is to calculate the Hilbert transform as follows: Construct the parsing signal: Thus, the envelope signal of the signal is obtained

[0018] Furthermore, multi-scale feature extraction is performed on the processed signal. The proposed multi-scale features include multi-scale features Multi-scale features Multi-scale features 3-FE3 Multi-scale features Multi-scale feature 5-FE5(max(x′)-min(x′)), multi-scale feature Multi-scale features Multi-scale Features 8-FE8 Multi-scale features Multi-scale features Multi-scale features 11-FE 11 (max(x′(i))) and multi-scale features Multi-scale features Multi-scale features and multi-scale features 15-FE 15 (max(f));

[0019] Among them, multi-scale features 1-FE1 to multi-scale features 12-FE 12 It is the time domain feature, which is the mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square value, crest factor, shape factor, pulse factor, gap factor, maximum value, and waveform factor. It directly reflects the change law of the signal in the time dimension; multi-scale feature 13-FE 13 to multi-scale features 15-FE 15 are frequency domain features, namely the average frequency, standard deviation frequency, and maximum frequency. The frequency domain features convert the signal from the time domain to the frequency domain through fast Fourier transform, which can reveal the impact of bone damage on the signal frequency distribution. x′ is the time domain guided wave signal, f is the frequency, and P(f) is the power spectrum corresponding to the frequency f.

[0020] Furthermore, the proposed multi-scale features also include: Multi-scale features 17-FE 17 (max|wt(m,k)|) and multi-scale features Among them, the multi-scale feature 16-FE 16 , Multi-scale Features 17-FE 17 and multi-scale features 18-FE 18 are the average time-frequency value, the maximum time-frequency value, and the standard deviation time-frequency value. The time-frequency domain features are transformed through continuous wavelet transform, which can capture the local changes of the signal in time and frequency at the same time, and can describe the transient characteristics of bone damage more finely; wt(m,k) is the wavelet transform coefficient of the signal at the mth moment and the kth scale, L is the length of the signal, k is the scale number of the wavelet transform, and |wt(m,k)| is the amplitude of the wavelet transform coefficient.

[0021] Furthermore, it also includes extracting features from the z′(m) signal using a CEEMD method;

[0022] Applying CEEMD to z′(m) results in the decomposition of the signal into a set of IMFs, each representing the behavior of the signal at different frequency components; z′(m) is written as where b i(m) is the i-th IMF, N is the number of IMFs, r m (m) is the Nth residual component, which represents the central tendency of z′(m);

[0023] The CEEMD energy characteristics are calculated by the intrinsic mode functions obtained above. For each IMF, its energy E is calculated. i To reflect the energy distribution of this frequency band, the specific formula is: The average energy characteristics of CEEMD (FE 19 ) is defined as: The energy entropy feature H is defined as follows: First, the total energy of the signal z′(m) is calculated Calculate the ratio of IMF energy to total energy where p i Represents the energy proportion of the i-th IMF, then the energy entropy characteristic is The energy entropy feature H is used to describe the distribution complexity of signal energy. The higher the entropy value, the more dispersed the energy distribution.

[0024] Furthermore, after feature extraction, feature selection based on GA specifically includes:

[0025] Use GA to select the optimal subset of features, is the feature set of class j, where m j is the number of samples in the jth class, and the dimension of each feature vector is g; 0 and 1 respectively indicate whether the corresponding feature is selected, 1 indicates selection, and 0 indicates non-selection. GA uses the fitness function to evaluate the quality of each individual, that is, the quality of the feature selection scheme;

[0026] The fitness function uses the ratio of the intra-class distance to the inter-class distance. The intra-class distance of the jth class is expressed as j=12,...,p, where b i is the i-th sample in the j-th class, is the center of the j-class sample, i.e., the average vector, ‖.‖ represents the Euclidean distance; p is the number of classes, so the total intra-class distance is defined as Similarly, the inter-class distance is The fitness function is defined as

[0027] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0028] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0029] In summary, the present invention proposes an ultrasonic bone injury detection method that combines multi-scale features and machine learning, aiming to improve detection accuracy in complex signal and small sample data scenarios. This method collects UGW signals and performs normalization and envelope processing on them, effectively removing signal noise and retaining key injury information. By combining feature extraction designed in multiple signal processing fields, such as the time domain, frequency domain, and time-frequency domain, and targeting the multi-scale characteristics and complexity of bone injury signals, complementary ensemble empirical mode decomposition (CEEMD) energy features, energy entropy features, and ASDDF features are proposed. To ensure scale uniformity between features and eliminate redundant features, the present invention uses a genetic algorithm (GA) for feature selection. Finally, the multi-features optimized by the GA are input into a fully connected neural network (MFI-FCNN) model designed and trained using the K-fold cross-validation method to obtain an efficient automatic bone injury recognition model. This method has superior accuracy and adaptability to traditional methods for bone injury detection in complex environments, providing an efficient and accurate solution for early screening of osteoporosis and fractures.

[0030] This paper utilizes a self-designed automated machine learning algorithm classification model, MFI-FCNN. FCNN efficiently classifies selected features and effectively prevents overfitting through regularization techniques. Compared to traditional support vector machines (SVMs), the model employed in this paper demonstrates higher accuracy in bone injury detection, particularly with small sample data sets. Consequently, MFI-FCNN significantly improves the accuracy of bone injury detection, demonstrating its potential for application in injury detection, particularly in environments with complex signals and small sample data.

[0031] The proposed ultrasonic bone injury detection method, which leverages multi-scale features and machine learning, not only more accurately assesses fracture severity but also provides orthopedic surgeons with a more reliable diagnostic basis. Compared to traditional methods, this method demonstrates higher accuracy in complex signal and small sample data scenarios, demonstrating significant clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is the algorithm flow chart of the present invention;

[0033] Figure 2 It is a schematic diagram of the UGW signal acquisition device of the present invention;

[0034] Figure 3 This is a comparison chart before and after the first step of data preprocessing in an embodiment of the present invention;

[0035] Figure 4 is a three-dimensional graph of original features and dedimensionalized features of an embodiment of the present invention;

[0036] Figure 5 is a flowchart of the steps of GA in an embodiment of the present invention;

[0037] Figure 6 is a three-dimensional feature graph selected by GA in an embodiment of the present invention;

[0038] Figure 7 Schematic diagram of the structure of the MFI-FCNN model according to an embodiment of the present invention;

[0039] Figure 8 is a schematic diagram of an experimental device according to an embodiment of the present invention;

[0040] Figure 9 1 is a schematic diagram of the accuracy loss curve of the automatic classification model training in Example 1 of the present invention;

[0041] Figure 10 Schematic diagram of the accuracy loss curve of the automatic classification model training in Example 2 of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0043] The embodiment of the present invention proposes an innovative UGW-based bone injury detection method, specifically a UGW bone injury detection method based on multi-scale features and machine learning. This method aims to avoid radiation exposure, reduce costs, improve portability, and enhance the ability to detect early bone injuries. This method overcomes the shortcomings of traditional detection technologies. By extracting features from multiple angles such as the time domain, frequency domain, and time-frequency domain, a multidimensional feature space is constructed to comprehensively integrate the microstructure of the bone. GA is used to automatically screen the optimal features and reduce redundant information, thereby improving model efficiency and detection accuracy. To prevent overfitting, the present invention proposes an MFI-FCNN model, which effectively improves its performance through regularization. This method has the potential to improve the accuracy of bone injury detection and be applied to early screening of osteoporosis and fractures.

[0044] Specifically, a multi-scale feature and machine learning-based ultrasound detection method for bone damage in an embodiment of the present invention includes the following steps:

[0045] (1) The original sample signal was collected using UGW equipment. In vitro bovine bones (length 22 cm) were used as samples to obtain normal bone signals and damaged bone signals. The damaged bone signal was simulated by coupling iron blocks. Different iron block sizes were used to simulate the severity of the injury. The distance between the iron block and the transducer was 2 cm, 4 cm, and 7 cm. The excitation signal had a center frequency of 65 kHz. The sampling interval for data acquisition was approximately 100 ms. Each position corresponding to the iron block was sampled 50 times to remove random noise, thereby improving the signal-to-noise ratio, and converting the signal into a digital format.

[0046] (2) Before extracting features, the signal needs to be preprocessed to improve data quality. The goal is to eliminate interference and ensure that the extracted features accurately reflect the actual state of the sample, thereby improving the classification performance of the MFI-FCNN model. The formula for the first step of preprocessing is as follows:

[0047]

[0048] Where y(m) is the original waveguide signal, z′(m) is the signal after the first step of preprocessing, L represents the length of y(m), and the normalized signal is then Hilbert transformed. The calculation formula for the second step of preprocessing is as follows:

[0049] Construct the parsing signal: Thus, the envelope signal of the signal is obtained

[0050] (3) The preprocessed signal optimizes the data to a certain extent, but it still contains irrelevant information; through feature extraction, more concentrated and stable damage information can be extracted from the preprocessed signal, which helps to improve the recognition performance of the model. The proposed multi-scale features include multi-scale features Multi-scale Features 2-FE2 Multi-scale features Multi-scale features Multi-scale feature 5-FE5(max(x′)-min(x′)), multi-scale feature Multi-scale features 7-FE7 Multi-scale features Multi-scale features Multi-scale features Multi-scale features 11-FE 11 (max(x′(i))) and multi-scale features Multi-scale features Multi-scale features and multi-scale features 15-FE 15(max(f)). Among them, multi-scale features 1-FE1 to multi-scale features 12-FE 12 It is a time domain feature, which is (mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square value, crest factor, shape factor, pulse factor, gap factor, maximum value, waveform factor). The time domain feature directly reflects the change law of the signal in the time dimension and can describe the impact of bone damage on signal amplitude, energy and other characteristics. Multi-scale feature 13-FE 13 to multi-scale features 15-FE 15 Frequency domain features include mean frequency, standard deviation frequency, and maximum frequency. Frequency domain features convert signals from the time domain to the frequency domain using a fast Fourier transform (FFT), revealing the impact of bone injury on the signal's frequency distribution. x' represents the time-domain guided wave signal, f represents the frequency, and P(f) represents the power spectrum corresponding to frequency f. The 12 extracted time domain features comprehensively quantify the characteristics of bone injury signals. The mean and RMS value provide an energy baseline, enabling rapid screening for overall signal anomalies. The standard deviation and peak-to-peak value measure the intensity of signal fluctuations and are closely related to the extent of injury. The maximum value can be used to locate the critical point of signal attenuation and is particularly sensitive to bone density loss. The skewness and kurtosis reflect morphological changes in the signal distribution, enabling identification of asymmetry in the fracture region and transient impact characteristics. Furthermore, the crest factor, impulse factor, and gap factor collaborate to detect localized signal abrupt changes. The gap factor is sensitive to microcracks, the crest factor helps locate fractures, and the impulse factor captures signal anomalies caused by irregular bone surface structures. The shape factor and waveform factor comprehensively assess signal morphology, effectively distinguishing chronic injuries. In terms of frequency domain features (mean frequency, standard deviation frequency, maximum frequency), these indicators can accurately capture the changes in frequency distribution caused by bone injuries. Among them, the mean frequency reflects the main frequency band of energy, and its offset can be used to distinguish different types of injuries. The standard deviation frequency quantifies the degree of discreteness of frequency domain energy. High values usually indicate complex injuries, such as bone fractures accompanied by soft tissue injuries. The maximum frequency is used to identify the high-frequency cutoff point, and its attenuation can serve as an early warning signal for osteoporosis. The combination of these time domain and frequency domain features greatly improves the sensitivity and specificity of injury classification, providing a more reliable basis for the automatic detection of bone injuries.

[0051] (4) Also includes: multi-scale features Multi-scale features 17-FE 17 (max|wt(m,k)|) and multi-scale features Among them, the multi-scale feature 16-FE 16 , Multi-scale Features 17-FE 17 and multi-scale features 18-FE 18The time-frequency domain features are the average time-frequency value, the maximum time-frequency value, and the standard deviation time-frequency value. Using continuous wavelet transform (CWT), the time-frequency domain features simultaneously capture local variations in the signal in time and frequency, enabling a more detailed description of the transient characteristics of bone damage. wt(m,k) represents the wavelet transform coefficients of the signal at time m and scale k, L represents the signal length, k represents the number of wavelet transform scales (i.e., the number of wavelet bases), and |wt(m,k)| represents the amplitude of the wavelet transform coefficients. In bone damage detection applications, the average time-frequency value reflects the overall energy distribution of the bone damage area and helps distinguish healthy from damaged bone tissue. In UGW detection, the energy distribution of healthy bone tissue is relatively uniform, while damaged areas may exhibit energy attenuation or abnormal energy enhancement. Therefore, the average time-frequency value can serve as an effective indicator for damage detection. The maximum time-frequency value is used to identify local high-energy response points caused by bone damage. For example, in the early stages of crack or microdamage formation, the energy of the ultrasonic signal will locally concentrate at the damage site, resulting in peak enhancement. This phenomenon can be captured using the maximum time-frequency value, enabling early damage detection. Furthermore, the standard deviation of the time-frequency value can describe the volatility of time-frequency characteristics. In bone injury detection, it can be used to analyze the degree of signal variability within the damaged area. Severe injuries often lead to increased non-stationary characteristics of the signal, manifesting as increased volatility of the time-frequency characteristics. Therefore, the standard deviation of the time-frequency value can help identify different types and severity of injuries.

[0052] (5) In addition to the features extracted from the time domain, frequency domain, and time-frequency domain, some custom features are also used. The CEEMD method extracts features from the z′(m) signal. CEEMD is an adaptive signal decomposition method. When applied to z′(m), the signal is decomposed into a set of IMFs, each of which represents the behavior of the signal at different frequency components. z′(m) can be written as where b i (m) is the i-th IMF, N is the number of IMFs, r m (m) is the Nth residual component, which represents the central tendency of z′(m); the CEEMD energy characteristic is calculated by the intrinsic mode function obtained above. For each IMF, its energy E is calculated. i To reflect the energy distribution of this frequency band, the specific formula is: The average energy characteristics of CEEMD (FE 19 ) is defined as: The energy entropy feature H is defined as follows: First, the total energy of the signal z′(m) is calculated Calculate the ratio of IMF energy to total energy where p i Represents the energy proportion of the i-th IMF, then the energy entropy characteristic is The energy entropy feature H describes the complexity of the signal energy distribution. Higher entropy values indicate a more dispersed energy distribution. The CEEMD average energy feature reflects the energy distribution of each IMF component after decomposition. Changes in energy concentration can reveal the impact of damage on the signal. The energy entropy feature measures the complexity and uncertainty of the signal energy distribution. The energy distribution of healthy bone tissue is relatively stable, while damage causes an increase in energy entropy. Therefore, this feature can be used to assess the severity of damage. Higher entropy values may indicate more severe bone damage. These features can describe the changes in bone damage signals from different perspectives. When combined with deep learning models, they can improve the accuracy and robustness of damage classification, providing higher sensitivity and reliability for UGW detection.

[0053] (6) Following the content of (5) above, this is also a customized feature based on demand. When different types of bone structure damage occur, the resonant frequency distribution and energy distribution of the signal will change significantly; therefore, the average spatial distance-based damage feature (ASDDF) is proposed to address this difference; ASDDF provides an effective method to describe the change of signal characteristics by quantifying the distance between the normal state and the damaged state IMFs, which is defined as: in, represents the i-th IMF under normal conditions, where represents the i-th IMF under the damage state, where The ASDDF feature combined with the fractal properties of the signal can extract the nonlinear changes caused by damage. It is suitable for detecting tiny cracks or early damage. These damages may not have caused significant energy changes, but have already shown abnormal characteristics in the signal structure. It is especially important for the early diagnosis of latent injuries such as osteoporosis.

[0054] (7) After completing the extraction of time domain, frequency domain, and time-frequency domain features, the average energy features, energy entropy features, and ASDDF features obtained by CEEMD decomposition were further extracted to comprehensively characterize the characteristic distribution of bone injury signals. The extracted features were further dedimensionalized. This processing step is crucial to improving the generalization ability and convergence speed of the model. It also helps to reduce errors and anomalies in the numerical calculation process. The method is the formula for the first step preprocessing in step (2).

[0055] (8) After feature extraction, GA is used for feature selection. Feature selection improves the performance of the classifier algorithm by reducing redundant and irrelevant features, improving generalization ability and avoiding overfitting; GA is used to select the optimal subset of features. is the feature set of class j, where m jis the number of samples in the jth class, and the dimension of each feature vector is g; 0 and 1 respectively indicate whether the corresponding feature is selected (1 means selection, 0 means no selection). GA uses the fitness function to evaluate the quality of each individual, that is, the quality of the feature selection scheme. The fitness function can be expressed as the ratio of the intra-class distance to the inter-class distance. The intra-class distance of the jth class can be expressed as where b i is the i-th sample in the j-th class, is the center of the j-class sample (i.e., the mean vector), ‖.‖ represents the Euclidean distance; p is the number of categories (2 in this study), so the total intra-class distance can be defined as Similarly, the inter-class distance is So the fitness function is defined as GA is used to select the most discriminative feature combinations from multiple feature categories to optimize the performance of bone injury detection models. However, not all features contribute to injury classification, and redundant or irrelevant features can reduce the model's generalization ability and increase computational overhead. GA simulates the natural evolutionary process to automatically select the subset of features that best distinguish injury types, improving model accuracy and computational efficiency. In practice, this feature selection method not only improves the reliability of bone injury detection but also reduces data processing and storage requirements.

[0056] In general, in order to achieve injury classification, the automatic classification machine learning algorithm model designed in the present invention is MFI-FCNN. The technical innovation of the present invention lies in the combination of UGW signal acquisition, multi-scale feature integration and machine learning classification algorithm to build a new bone injury detection system. Specific improvements include: (1) using UGW to replace traditional X-rays to achieve radiation-free detection and capture microscopic bone structure changes through high-frequency signals; (2) proposing a multi-scale feature integration method that integrates time domain, frequency domain, time-frequency domain, CEEMD energy entropy, ASDDF and other features to significantly improve feature expression capabilities; (3) designing a proprietary MFI-FCNN machine learning model for injury classification, which can still maintain an accuracy of 95.37% under small sample conditions (training data reduced by 60%), while the performance of the traditional SVM method is degraded. The model constructs a network architecture comprising a feature input layer, a two-layer fully connected hidden layer, and a classification output layer. Both the first and second fully connected layers are designed with 50 neurons, and a dropout layer with a 30% dropout rate is set after the second fully connected layer. The Adam optimization algorithm is used for network training, with an initial learning rate of 1e-2 and an L2 regularization parameter of 1e-5 to achieve the optimization process. During training, a small batch size of 16 samples and a 50-round iteration strategy are used. The model uses the ReLU activation function between hidden layers, and the Softmax function in the output layer to achieve binary classification probability output. This invention significantly improves the model's generalization ability and training efficiency while ensuring classification accuracy, making it particularly suitable for classification applications involving small and medium-sized datasets.

[0057] As can be seen above, the embodiments of the present invention utilize UGW technology to acquire signals. This technology is favored for its radiation-free nature, cost-effectiveness, portability, and high sensitivity for detecting early bone damage. Compared to traditional X-ray imaging, UGW technology provides a safer and more economical means of assessing bone density and structural integrity, significantly reducing the risk of radiation exposure while improving accessibility and accuracy.

[0058] Specifically, after using UGW technology to collect the signal to be measured, the UGW signal is first normalized to a mean of 0 and a variance of 1, eliminating outliers and noise. Next, the signal envelope, E(m), is calculated. This envelope effectively extracts the changing characteristics of the UGW signal, providing key information for subsequent feature extraction and classification analysis.

[0059] The present invention takes into account that bone conditions are affected by uncertain factors such as noise interference, and a single feature is difficult to achieve stable and accurate damage prediction. Therefore, this paper extracts time domain, frequency domain, and time-frequency domain features from the signal to fully capture the multi-scale information and complex changes of bone damage signals. In addition, the present invention also proposes and applies the average energy feature, energy entropy feature, and ASDDF feature of CEEMD. These features, combined with the microstructure and mechanical properties of bones, can effectively distinguish normal bone tissue from damaged bone tissue, significantly improving the accuracy and reliability of bone damage detection.

[0060] After feature extraction, the features are de-dimensionalized. The purpose of de-dimensionalization is to eliminate the impact of dimensional differences between different features and ensure that all features are analyzed on the same scale, thereby improving the comparability between features and the effectiveness of the model.

[0061] After dedimensionalizing the features, feature selection is required. This method uses a GA for feature selection. By defining a fitness function to evaluate the performance of each individual (feature selection scheme), the fitness value determines which individuals will be selected, crossovered, and mutated, thereby generating a new generation of the population. This process effectively eliminates irrelevant and redundant features, further improving classification performance.

[0062] The present invention uses a K-fold cross-validation method to train and validate the MFI-FCNN model. This method divides the dataset into K disjoint subsets, selecting one of these subsets as the validation set each time, and using the remaining K-1 subsets as the training set. This process is repeated K times, with a different subset selected as the validation set each time, ensuring that every data point is used for both training and validation. Implementing K-fold cross-validation not only improves the accuracy and reliability of the MFI-FCNN model in bone lesion detection tasks, but also further verifies the model's robustness and generalization capabilities by systematically evaluating its performance on different data subsets.

[0063] The following examples illustrate:

[0064] Example 1:

[0065] like Figure 1 As shown in FIG, the flowchart of the multi-scale feature and machine learning bone injury ultrasonic detection method proposed in the present invention. Figure 2 As shown in FIG, the device used in the present invention is a UGW instrument, which is used to collect bone waveguide signals. The collected signals are pre-processed to prepare for the subsequent feature extraction step. The comparison diagram of the waveguide signal before and after pre-processing is shown in FIG. Figure 3As shown in the figure, after the preprocessed signal, it is necessary to perform feature extraction on the signal, extracting features from multiple aspects such as time domain, frequency domain, time-frequency domain, and energy spectrum. After the feature extraction is completed, the features are dedimensionalized. The purpose of dedimensionalization is to eliminate the influence of different features due to dimensional differences, ensure that all features are subsequently analyzed on the same scale, thereby improving the comparability between features and the effectiveness of the model. The three-dimensional diagram of the original features and dedimensionalized features is shown in the figure. Figure 4 As shown in Figure 1, (a) is the original feature 3D graph of the extracted preprocessed signal, and (b) is the feature 3D graph after de-dimensionalization. After feature extraction, GA is used to select features to filter out the most discriminative feature subset. This process effectively eliminates irrelevant and redundant features. The GA workflow diagram and the typical feature 3D graph after selection are shown in Figure 1. Figure 5 As shown. The model performance is evaluated by K-fold cross validation. This method divides the data set into K non-overlapping subsets, selects one of the subsets as the test set each time, and uses the remaining K-1 subsets as the training set. After completing feature selection and cross validation, the machine learning algorithm MFI-FCNN classifier model of the present invention is used for training. Compared with the traditional method SVM, the model used in the present invention has a higher accuracy in bone damage detection, especially showing a strong generalization ability on small sample data sets. The structural diagram of the MFI-FCNN model is shown in Figure 6 As shown in the figure, the process is to preprocess the collected signal, then extract the features of the signal, de-dimensionalize the extracted features, use GA to select features, and finally send the selected features into FCNN.

[0066] This experiment tests the generalization ability of the algorithm under different damage locations. The total experimental data is shown in the table.

[0067]

[0068] The experiment is divided into two stages: The first stage is to evaluate the classification performance of the model through the test set. The data used is the signal of the small iron block at the position of 2cm, 4cm and 7cm as the training data. The schematic diagram of the experimental setup is shown below. Figure 7 shown. Figure 7 In the UGW detection experiment conducted on simulated bone materials, iron blocks of different sizes were used to simulate bone damage. Figure (a) shows a rectangular defect with a size of 10mm×15mm×20mm, Figure (b) shows a defect with a size of 10mm×15mm×15mm, and Figure (c) shows a smaller defect with a size of 5mm×10mm×10mm. In order to fully demonstrate the performance of the model in the first stage, the accuracy loss curve of the MFI-FCNN model of the present invention is shown below. Figure 8As shown in the figure, the results show that the model converges quickly in the initial stage, and the training accuracy stabilizes within a short period of time, demonstrating good learning performance and adaptability. The trained model is then used to predict new data. The second stage is to verify the model's generalization ability, using an independent validation dataset for testing. This dataset contains 475 samples, covering normal signals as well as signals of small, medium, and large iron blocks at different locations (2 cm, 4 cm, and 7 cm), independent of the training and test datasets. Through this stage of verification, the model can accurately determine the presence of bone damage, further providing an accurate basis for orthopedic diagnosis. This systematic approach significantly improves the accuracy of bone injury detection. Compared with the traditional support vector machine (SVM) method, the accuracy results are shown in the table.

[0069]

[0070] Experimental results show that the classification performance of the MFI-FCNN proposed in this paper is better than that of the SVM at the positions of 2 cm, 4 cm, and 7 cm of the small iron block. Overall, the machine learning algorithm MFI-FCNN proposed in this paper exhibits stronger feature extraction capabilities and adaptability, and is particularly stable in the detection of complex damage signals, laying a good foundation for subsequent research on generalization capabilities under complex conditions.

[0071] Example 2:

[0072] like Figures 1 to 7 As shown, this embodiment follows the same algorithm flow, signal acquisition device, data preprocessing steps, feature extraction and dedimensionalization method, feature selection process, MFI-FCNN model structure and experimental device as Example 1. Specifically: Figure 1 The algorithm flow chart of the present invention is presented, and the entire process from data collection to model prediction is described in detail. Figure 2 This is a schematic diagram of the UGW signal acquisition device of the present invention, illustrating the specific settings for signal acquisition. Figure 3 The comparison before and after the first step of data preprocessing is shown, highlighting the impact of preprocessing on the signal. Figure 4 Figures a and b show the three-dimensional graphs of the original features of the extracted preprocessed signal and the dedimensionalized features, respectively, illustrating the effect of dedimensionalization. Figure 5 Flowchart of the GA feature selection steps, Figure 6 This is a three-dimensional graph of features after GA selection, showing the process and results of feature selection. Figure 7 This is a structural diagram of the MFI-FCNN model of the present invention, which describes the architecture of the model in detail. Figure 8The figure is a schematic diagram of the experimental setup, showing the configuration of the experiment. Compared with Example 1, this Example 2 makes adjustments to the dataset to verify the generalization ability of the model under different conditions. As Example 2, the present invention aims to evaluate the performance of the model under harsh conditions. To this end, six sets of test data were designed to simulate different clinical scenarios and sample conditions. First, single-position samples were prepared, including 20 samples, which were sampled at the 2cm and 4cm positions to test the performance of the model under a single injury location. Secondly, a dual-position combination sample was constructed, also containing 20 samples, which covered the combination of 2cm+4cm and 4cm+7cm positions, and the adaptability of the model was evaluated by increasing the complexity of the injury location. In addition, in order to verify the robustness of the model under insufficient data conditions, a dual-position small sample test was further designed, containing 10 samples, based on the combination of 2cm+4cm and 4cm+7cm, but with a reduced number of samples. In the first stage, these test data groups were used for training together with normal signals, and in the second stage, an independent dataset was used to verify the performance of the model. In order to fully demonstrate the performance of the model in the first stage, the accuracy loss curve of the MFI-FCNN model of the machine learning algorithm of the present invention is shown below. Figure 10 As shown, Figure 10 This is the accuracy loss curve of Example 2 of the present invention. The results show that the MFI-FCNN model of the present invention's machine learning algorithm converges quickly in the initial stage and achieves 100% training accuracy in a short period of time. The loss value gradually approaches zero, verifying the stability and efficiency of the model training. Compared with the traditional method SVM, the accuracy results are shown in the following table

[0073]

[0074] Analysis of experimental results shows that the proposed machine learning algorithm, the MFI-FCNN model, outperforms the SVM in robustness and generalization under complex conditions. In particular, MFI-FCNN maintains high classification accuracy under small sample sizes and complex position combinations. In contrast, the SVM performs poorly under small sample sizes, with a significant drop in accuracy. This phase of the experiment further validates the advantages of machine learning models in complex damage signal detection, providing theoretical support for model optimization and practical nondestructive testing applications.

[0075] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0076] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0077] In another embodiment provided in the present application, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any of the multi-scale feature and machine learning bone injury ultrasonic detection methods in the above embodiments.

[0078] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above method.

[0079] The embodiment of the present application further provides an electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus.

[0080] Memory for storing computer programs;

[0081] The processor is used to implement the above-mentioned multi-scale feature and machine learning bone injury ultrasonic detection method when executing the program stored in the memory.

[0082] The communication bus mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0083] The communication interface is used for communication between the above electronic device and other devices.

[0084] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. Alternatively, the memory may be at least one storage device located remote from the processor.

[0085] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0086] It should also be noted that electronic devices also include terminal devices, which can also be called terminals, user equipment (UE), mobile stations (MS), mobile terminals (MT), etc. Terminal devices can be mobile phones, smart TVs, wearable devices, tablet computers, computers with wireless transceiver functions, virtual reality (VR) terminal devices, augmented reality (AR) terminal devices, wireless terminals in industrial control, wireless terminals in self-driving, wireless terminals in remote medical surgery, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, wireless terminals in smart homes, etc. The embodiments of the present application do not limit the specific technology and specific device form adopted by the terminal devices.

[0087] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. 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 computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrations. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A bone injury ultrasonic detection method based on multi-scale features and machine learning, characterized by: The following steps are included: Use UGW to collect original bone sample signals, normal signals and damage signals respectively, and convert the signals into digital format; Perform Z-score normalization and envelope processing on the collected signals; Perform multi-scale feature extraction on the processed signal; After feature extraction, feature selection is performed based on GA; Establish the MFI-FCNN classifier model and perform K-fold cross-validation training optimization; The optimized MFI-FCNN classifier model is used to judge whether the processed guided wave signal is damaged.

2. The bone injury ultrasonic detection method based on multi-scale features and machine learning according to claim 1 is characterized by: The collected signal is Z-score normalized and envelope processed, including the following steps: The formula for the first step of preprocessing is as follows: Where y(m) is the original waveguide signal, z′(m) is the signal after the first step of preprocessing, L represents the length of y(m), and the normalized signal is then Hilbert transformed. The second step of preprocessing is to calculate the Hilbert transform as follows: Construct the parsing signal: Thus, the envelope signal of the signal is obtained 3. The bone injury ultrasonic detection method based on multi-scale features and machine learning according to claim 2 is characterized by: Multi-scale feature extraction is performed on the processed signal. The proposed multi-scale features include multi-scale features 1-FE1 Multi-scale Features 2-FE2 Multi-scale features 3-FE3 Multi-scale features 4-FE4 Multi-scale feature 5-FE5(max(x′)-min(x′)), multi-scale feature 6-FE6 Multi-scale features 7-FE7 Multi-scale Features 8-FE8 Multi-scale Features 9-FE9 Multi-scale features 10-FE 10 Multi-scale features 11-FE 11 (max(x′(i))) and multi-scale features 12-FE 12 Multi-scale features 13-FE 13 Multi-scale features 14-FE 14 and multi-scale features 15-FE 15 (max(f)); Among them, multi-scale features 1-FE1 to multi-scale features 12-FE 12 It is the time domain feature, which is the mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square value, crest factor, shape factor, pulse factor, gap factor, maximum value, and waveform factor. It directly reflects the change law of the signal in the time dimension; multi-scale feature 13-FE 13 to multi-scale features 15-FE 15 are frequency domain features, namely the average frequency, standard deviation frequency, and maximum frequency. The frequency domain features convert the signal from the time domain to the frequency domain through fast Fourier transform, which can reveal the impact of bone damage on the signal frequency distribution. x′ is the time domain guided wave signal, f is the frequency, and P(f) is the power spectrum corresponding to the frequency f.

4. The method for ultrasonic bone lesion detection based on multi-scale features and machine learning according to claim 2, characterized in that: The proposed multi-scale features also include: Multi-scale feature 16-FE 16 Multi-scale features 17-FE 17 (max|wt(m,k)|) and multi-scale features 18-FE 18 Among them, the multi-scale feature 16-FE 16 , Multi-scale Features 17-FE 17 and multi-scale features 18-FE 18 They are the average time-frequency value, the maximum time-frequency value, and the standard deviation time-frequency value. The time-frequency domain features are transformed through continuous wavelet transform, which can capture the local changes of the signal in time and frequency at the same time, and can describe the transient characteristics of bone damage more finely; wt(m, k) is the wavelet transform coefficient of the signal at the mth moment and the kth scale, L is the length of the signal, k is the scale number of the wavelet transform, and |wt(m, k)| is the amplitude of the wavelet transform coefficient.

5. The bone injury ultrasonic detection method based on multi-scale features and machine learning according to claim 3 is characterized by: It also includes the use of CEEMD method to extract features from the z′(m) signal; The application of CEEMD to z′(m) results in the decomposition of the signal into a set of IMFs, each of which represents the behavior of the signal at different frequency components; z′(m) is written as where b i (m) is the i-th IMF, N is the number of IMFs, r m (m) is the Nth residual component, which represents the central tendency of z′(m); The CEEMD energy characteristics are calculated by the intrinsic mode functions obtained above. For each IMF, its energy E is calculated. i To reflect the energy distribution of this frequency band, the specific formula is: The average energy characteristics of CEEMD (FE 19 ) is defined as: The energy entropy feature H is defined as follows: First, the total energy of the signal z′(m) is calculated Calculate the ratio of IMF energy to total energy where p i Represents the energy proportion of the i-th IMF, then the energy entropy characteristic is The energy entropy feature H is used to describe the distribution complexity of signal energy. The higher the entropy value, the more dispersed the energy distribution.

6. The bone injury ultrasonic detection method based on multi-scale features and machine learning according to claim 1 is characterized by: After feature extraction, feature selection based on GA specifically includes: Use GA to select the optimal subset of features, is the feature set of class j, where m j is the number of samples in the jth class, and the dimension of each feature vector is g; 0 and 1 respectively indicate whether the corresponding feature is selected, 1 indicates selection, and 0 indicates non-selection. GA uses the fitness function to evaluate the quality of each individual, that is, the quality of the feature selection scheme; The fitness function uses the ratio of intra-class distance to inter-class distance. The intra-class distance of the jth class is expressed as where b i is the i-th sample in the j-th class, is the center of the j-class sample, i.e., the mean vector, ||.|| represents the Euclidean distance; p is the number of classes, so the total intra-class distance is defined as Similarly, the inter-class distance is The fitness function is defined as 7. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 6.

Images