Wound surface microorganism detection method based on optical fiber spectrum knowledge data dual-drive envelope

Abstract

The invention relates to the technical field of medical detection and spectral analysis, and discloses a spectrum detection method for microorganisms on a wound surface based on knowledge data dual-drive envelope fusion. The method comprises the following steps: firstly, acquiring a multi-band near-infrared spectrum signal of a wound surface through a fiber optic spectrometer, and carrying out normalization and smooth filtering pretreatment; analyzing optical characteristic parameters based on a diffuse reflection theory model, constructing a knowledge-driven characteristic sample space, extracting spectral high-order characteristics by using a one-dimensional deep convolutional neural network, and constructing a data-driven characteristic sample space; then generating a grain envelope sample through a double-flow feature fusion mechanism; and finally, constructing a lightweight dual-task classification model based on the particle envelope sample, and realizing rapid detection of fungal infection of the wound surface and identification of seven common bacteria. According to the method, the interpretability, the recognition precision and the robustness of characteristic pathology are considered, and an efficient and reliable technical scheme is provided for early clinical diagnosis of wound infection.

Description

Technical Field

[0001] This invention belongs to the field of near-infrared spectroscopy wound microbial detection technology, specifically involving a wound microbial detection method based on a dual-driven envelope of fiber optic spectral knowledge data. Background Technology

[0002] In recent years, near-infrared spectroscopy has demonstrated significant application value in the field of medical testing due to its non-invasive and rapid characteristics. Early diagnosis of bacterial and fungal infections of wounds is crucial for the development of clinical treatment plans, but existing detection methods all have shortcomings. Based on their detection principles, existing methods can be broadly categorized into direct detection methods and indirect inference methods. Direct detection methods, such as microbial culture and tissue biopsy, can provide relatively accurate diagnostic results, but they suffer from being highly invasive, complex in operation, and having long testing cycles. Indirect inference methods, such as clinical sign observation and empirical judgment, while possessing the advantages of speed and convenience, lack objective quantitative evidence, and the diagnostic results are easily influenced by subjective factors, resulting in insufficient accuracy.

[0003] Near-infrared spectroscopy, by analyzing the characteristic spectra of microbial metabolites, theoretically enables rapid and non-invasive detection of infection status. However, two technical challenges remain in the feature extraction stage: (1) Although optical parameters based on optical theoretical models (knowledge-driven) have strong pathological interpretability, they are subject to interference from tissue heterogeneity, resulting in insufficient characteristic stability and generalization. (2) Although deep learning methods based on pure data can extract high-order features, they lack interpretability due to the “black box effect”, and their results are difficult to guarantee clinical credibility.

[0004] Therefore, existing feature extraction methods struggle to balance "pathological interpretability" and "robustness," hindering the promotion and application of near-infrared spectroscopy in the early diagnosis of clinical infections. Summary of the Invention

[0005] The present invention aims to address the technical problem in the prior art of near-infrared spectroscopy for detecting microorganisms in wounds, where feature extraction is difficult to simultaneously consider the pathological interpretability of spectral features and the recognition efficiency of machine learning, resulting in a bottleneck in the accuracy of single-modal feature extraction.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting wound microorganisms based on dual-driven envelope of fiber optic spectral knowledge data, comprising the following steps: S1. Use a fiber optic spectrometer to collect multi-band near-infrared spectral signals of wound tissue, and perform normalization and noise reduction preprocessing on the raw spectral data. S2. Based on the diffuse reflection theoretical model, the set of optical characteristic parameters, including the reduced scattering coefficient, is extracted from the preprocessed spectral signal. and absorption coefficient Construct a knowledge-driven feature sample space; S3. A one-dimensional deep convolutional neural network 1D-CNN is used to extract high-order abstract features of spectral signals through multi-layer convolutional kernels to construct a data-driven feature sample space. S4. Establish a dual-stream feature fusion mechanism, realize spatial dimension fusion through feature cascading, and perform embedding mapping on the fused high-dimensional feature samples based on manifold dimensionality reduction method, and generate granular envelope samples based on the dimensionality-reduced knowledge-data samples. S5. Input the particle envelope samples as the training set into the lightweight classifier for model training. Establish a high-dimensional mapping relationship from spectral features to microbial species through iterative optimization. After processing the spectral data of the wound to be tested, input it into the model and output the classification and discrimination results of 7 common microorganisms. The theoretical model expression for diffuse reflection is as follows:

[0007] In the formula: , , , ; It is the radius of the light collection area. =0.1; parameter Depends on the refractive index of the medium , =3.2.

[0008] Preferably, step S1 specifically includes: S11. Preheat the spectrometer. Turn on the spectrometer, light source and supporting software SpectraSuite in sequence. Use the diffuse reflectance standard plate to collect the control spectrum. Turn off the light source to collect the dark spectrum. Turn the light source back on to collect the NIR spectrum of the diffuse reflectance standard plate and set the reflectance amplitude to 100%. After stabilizing for 5 minutes, collect the NIR spectrum signal of the wound. S12. Normalize the original spectral signal. The normalization calculation formula is as follows: ; In the formula, The spectral signal value; It is the broadband light intensity measured from the wound surface; It is the background spectrum when the light source is off; Represents the spectrum measured from the reflectance standard; S13. The normalized spectral signal is smoothed using a Savitzky-Golay filter. The calculation formula is as follows: ; In the formula, Represents the filter function. To be relative to the current point Offset Spectral signal values ​​at the location, This represents the number of data points within the filter window.

[0009] Preferably, in step S2, the optical characteristic parameters are extracted by fitting using the least squares method, and the fitting optimization formula is as follows: ; In the formula, The spectral signal value, This is the theoretical value for the spectrum.

[0010] Preferably, step S3 specifically includes: S31. Establish a one-dimensional convolutional neural network classification model using one-dimensional convolution, pooling, and fully connected layers. S32. Train the preprocessed spectral signal and use the Adam optimization algorithm to adjust the parameters of each layer. When the model is optimized to the point where the loss function no longer decreases, save the final model and stop the optimization iteration. S33. Use the final model to extract high-order abstract features of the spectral signal.

[0011] Preferably, in step S3, the convolution operation formula for the one-dimensional convolutional neural network is: ; The pooling operation formula is: ; The feature transformation formula for a fully connected layer is: ; In the formula, W is the weight parameter of the l-th layer, which is the convolution kernel in the convolutional and pooling layers, and the weight matrix in the fully connected layer; H is the output feature of the l-th layer, which is the feature map in the convolutional and pooling layers, and the feature vector in the fully connected layer; b is the bias term.

[0012] 6. The wound microbial detection method based on dual-driven envelope of fiber optic spectral knowledge data according to claim 1, characterized in that: in step S3, cross-entropy is used as the loss function, and the formula is: .

[0013] In the formula, This represents the cross-entropy loss value. This represents the total number of categories in the classification task. This is a real label; This represents the probability of the sample class predicted by the model.

[0014] Preferably, step S4 specifically includes: S41. Spatial dimension fusion is performed by cascaded optical parameter features and deep convolution features to obtain knowledge-data samples; S42. Perform manifold embedding mapping on high-dimensional knowledge data samples to preserve local neighborhood structure and remove redundant information. S43. First, the entire dataset is initialized as a single particle envelope, with its initial center being the average value of all samples. At this point, the purity of the initial particle envelope is the lowest, which does not meet the set threshold condition. S44. Split the particle envelope: Randomly select K-1 remaining heterogeneous points as new centers, and redistribute all samples in the particle envelope to these K centers to generate K new particle envelopes; S45. Calculate the purity value of each sub-envelope. If the set purity threshold is not met, continue splitting the sub-envelope until the threshold is reached.

[0015] As a preferred option, the spatial dimension fusion formula for feature concatenation in step S4 is: ; In the formula, Indicates cascading splicing. and Features of knowledge-based samples These are features of data-type samples.

[0016] Preferably, the objective function for manifold dimensionality reduction in step S4 is:

[0017] In the formula, Represents the projection matrix. Z represents the transpose of a matrix, and Z represents the data matrix. Represents sample points, Represents the Laplace matrix, Degree matrix, Represents the adjacency matrix. This represents the adjacency weight between two samples before dimensionality reduction; that is, the closer the distance, the greater the weight, and the smaller the distance, the greater the weight.

[0018] The adjacency weight calculation formula is:

[0019] In the formula, and These represent two different high-dimensional feature samples.

[0020] Preferably, the formula for calculating particle envelope purity in step S4 is:

[0021] In the formula, This represents the number of samples in the i-th class.

[0022] Compared with the prior art, the technical effects and advantages of the present invention are: (1) In the prior art, although knowledge-driven optical parameters have clear pathological significance, they are easily affected by tissue heterogeneity, and although data-driven deep learning features can capture high-order information, they suffer from the "black box effect". This invention, through knowledge-data dual-driven feature fusion, not only retains the biomedical interpretability of optical parameters based on diffuse reflection theory, but also combines high-order abstract features extracted by one-dimensional convolutional neural networks, effectively overcoming the limitation of a single modality in balancing "interpretability" and "robustness".

[0023] (2) This invention designs a dual-stream feature fusion mechanism and a particle envelope sample construction method, which cascades and fuses the two types of features and optimizes dimensionality reduction to fully explore the biophysical differences and statistical regularities in the spectral signals and reduce redundant information and noise interference. Compared with traditional single feature extraction methods, this approach can capture microbial-related features more comprehensively, significantly improve the rapid detection capability of wound fungal infections and the accuracy of identifying seven common bacteria, and shows stronger generalization on different patient samples.

[0024] (3) This invention adopts a lightweight dual-task classification model, which reduces computational overhead while ensuring recognition performance, and can complete the processing and classification of spectral data of the wound to be tested without complicated operations. Compared with the long cycle and strong invasiveness of traditional direct detection methods such as microbial culture, and the subjectivity of indirect inference methods such as clinical sign observation, this technology can realize rapid and non-invasive detection of microorganisms in wounds, providing efficient support for early clinical diagnosis.

[0025] (4) This invention optimizes multiple aspects (spectral preprocessing, dual-drive feature modeling, envelope fusion, and lightweight classification) to form a complete wound microbial spectral detection technology system. It not only outperforms existing common wound spectral analysis algorithms in performance but also provides clinically relevant detection results that are both "interpretable" and "accurate," helping medical staff to more scientifically determine the type of infection, providing a reliable basis for subsequent treatment plans, and promoting the application of near-infrared spectroscopy technology in the field of clinical infection diagnosis. Attached Figure Description

[0026] Figure 1 Examples of wound photographs and near-infrared spectral imaging for different infection types according to this invention; Figure 2 This is a diagram showing the effect of spectral preprocessing and feature parameter extraction in this invention.

[0027] Figure 3 This is a flowchart of the method of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] The following combination Figures 1 to 3 This application will be described in further detail. Studies have shown that the near-infrared spectral signals of wound tissue not only contain optical differences caused by microbial metabolic activity, but are also closely related to the structural characteristics of the tissue itself. This dual information is of great value in improving the accuracy of microbial detection. Based on this, to address the contradiction between insufficient pathological interpretability and unstable model performance in the aforementioned feature extraction process, this invention provides a wound microbial spectral detection method based on dual-driven envelope fusion of knowledge data. After spectral signal acquisition and preprocessing, and before classifier training, this method performs dual-driven feature modeling and envelope fusion on the original spectral samples, thereby generating novel sample features that combine biomedical significance and machine learning performance for subsequent classification and identification. As shown in the figure, the method includes the following steps: S1. Acquire multi-band near-infrared spectral signals of wound tissue using a fiber optic spectrometer, and preprocess the raw spectral data. S2. Based on the diffuse reflection theoretical model, the set of optical characteristic parameters, including the reduced scattering coefficient, is extracted from the preprocessed spectral signal. ) and absorption coefficient ( Construct a knowledge-driven feature sample space; S3. A one-dimensional deep convolutional neural network (1D-CNN) is used to extract high-order abstract features of spectral signals through multi-level convolutional kernels to construct a data-driven feature sample space. S4. Establish a dual-stream feature fusion mechanism, realize spatial dimension fusion through feature cascading, and perform embedding mapping on the fused high-dimensional feature samples based on manifold dimensionality reduction method. Generate granular envelope samples based on the dimensionality-reduced knowledge-data samples. S5. In this step, the previously generated particle envelope samples are used as input. Algorithms such as XGBoost are used to learn the complex mapping relationship between fusion features and microbial species, and a lightweight dual-task classification model is trained: ① Fungal classification branch: The XGBoost classifier is used to output the judgment result of whether the wound is infected with fungi; ② Bacterial classification branch: The XGBoost classifier is used to first determine whether the wound is infected with bacteria, and then further identify 7 common bacterial species.

[0030] This method aims to address the accuracy bottleneck of single-modality feature extraction in existing methods by constructing a knowledge-driven and data-driven dual-stream feature fusion module to encapsulate spectral signals and generate granular envelope samples that possess both pathological interpretability and deep abstract representation. During envelope construction, the model adaptively combines optical parameters obtained from diffuse reflection theory with high-order features extracted by deep convolutional networks to achieve complementary modeling of the knowledge and data spaces, more effectively capturing biophysical differences and statistical regularities in spectral signals. Simultaneously, through a feature cascade dual-stream fusion mechanism, the interrelationships between different modal features are further explored to obtain the final granular envelope samples, thereby significantly improving the accuracy and robustness of the classifier in the rapid identification of various microorganisms.

[0031] (1) Step S1: Acquisition and preprocessing of wound spectral signals The near-infrared spectral signal of wound tissue contains complex tissue structure and microbial metabolic characteristics. The raw signal is often accompanied by noise and baseline drift. Therefore, standardization calibration is required during acquisition, and preprocessing is necessary after acquisition. Step S1 specifically includes the following sub-steps: S11. Spectrometer Preparation and Signal Acquisition: Warm up the fiber optic spectrometer to ensure it is in a stable acquisition state. Before acquisition, perform the following operations in sequence: turn on the spectrometer, light source, and its accompanying software SpectraSuite; acquire the control spectrum using a diffuse reflectance standard plate at the sample location; turn off the light source and acquire the dark spectrum; turn the light source back on and acquire the diffuse reflectance standard plate spectrum, setting its reflectance amplitude to 100%. After approximately 5 minutes of stable operation, the spectrometer will enter normal working condition. At this point, place the wound sample and acquire its multi-band near-infrared (NIR) spectral signals.

[0032] S12. Normalization Processing: The acquired raw spectral signals are normalized to eliminate amplitude differences under different sampling conditions, ensuring the stability and comparability of subsequent analyses. The formula for normalizing diffuse reflectance spectra is as follows:

[0033] The spectral signal value; It is the broadband light intensity measured from the wound surface; It is the background spectrum when the light source is off; This represents the spectrum measured from the reflection standard.

[0034] S13. Smoothing Filtering: The normalized spectral signal is smoothed using a Savitzky-Golay filter. This filter calculates local fitting parameters through polynomial convolution operations, effectively removing noise while preserving the spectral peak shape characteristics, thereby improving the quality of the spectral signal. The calculation formula is as follows:

[0035] In the formula, Represents the filter function. To be relative to the current point Offset Spectral signal values ​​at the location, This represents the number of data points within the filter window.

[0036] (2) Step S2: Construct a knowledge-driven feature sample space The preprocessed spectral signal remains in its raw form and lacks clear pathological interpretation. To further extract biophysically significant features, this step analyzes the spectral signal based on a diffuse reflectance theory model to obtain a set of optical property parameters reflecting tissue structure and metabolic state. Specifically, this includes the reduced scattering coefficient (…). ) and absorption coefficient ( ), It is directly proportional to the level of collagen in the wound, and it changes with the presence of a large number of inflammatory cells in the wound, providing a basis for whether the wound is infected. This parameter is directly related to the water content of the wound and changes when the wound becomes infected. Based on this, a knowledge-driven feature sample space is constructed, providing pathologically interpretable baseline features for subsequent data-driven feature learning and fusion modeling. Step S2 specifically includes the following steps: S21. A complex diffuse reflectance model is established between near-infrared spectral intensity and optical characteristic parameters, and its expression is:

[0037] In the formula: , , , ; It is the radius of the light collection area. =0.1; parameter Depends on the refractive index of the medium , =3.2.

[0038] S22. The model is extracted using the least squares method to reasonably obtain the required optical characteristic parameters of the wound. The original measured diffuse reflectance spectral data is then fitted. and the calculated theoretical spectral data By minimizing the sum of squared errors within the wavelength range To optimize the fit, the optical characteristic parameters of each band are obtained. and Its formula is

[0039] (3) Step S3: Constructing a data-driven feature sample space While the spectral signal extracted using knowledge-driven feature extraction possesses clear biophysical interpretability, it may still miss higher-order statistical features and complex pattern information. To fully exploit the potential higher-order features in the spectral signal, this step employs a one-dimensional deep convolutional neural network (1D-CNN) to automatically learn local and global spectral patterns through multiple convolutional kernels. The resulting data-driven feature sample space can supplement the shortcomings of knowledge-driven features, providing rich abstract representations for subsequent dual-stream feature fusion and particle envelope generation. Step S3 includes the following specific steps: S31. A one-dimensional convolutional neural network classification model is established using one-dimensional convolution, pooling, and fully connected layers. The formula for convolution can be expressed as follows: The pooling formula can be expressed as: The feature transformation of a fully connected layer can be expressed as .

[0040] S32. Train the model on the preprocessed spectral signal and adjust the parameters of each layer using the Adam optimization algorithm. When the model optimization reaches a point where the loss function no longer decreases, save the final model and stop the optimization iteration. The cross-entropy loss function is... In the formula, This represents the cross-entropy loss value. This represents the total number of categories in the classification task. This is a real label; This represents the probability of the sample class predicted by the model.

[0041] S33. Use the final model to extract high-order abstract features of the spectral signal. Assume the trained 1D-CNN model is denoted as... The extracted data-driven features can be represented as .

[0042] (4) Step S4: Establish a dual-stream feature fusion mechanism. Spatial dimension fusion is achieved through feature concatenation, and then manifold embedding mapping is performed on the fused high-dimensional feature samples to generate granular envelope samples in the dimensionality-reduced embedding space; S41. Concatenate the knowledge-driven feature sample space and the data-driven feature sample space to form a unified high-dimensional fusion feature representation, thereby balancing the pathological interpretability and deep abstract representation capability of spectral signals. Assume the knowledge-based sample features are... and The characteristics of data-type samples are Their spatial dimensions are then fused as follows:

[0043] This indicates cascading splicing.

[0044] S42. Perform manifold dimensionality reduction on the fused features. While preserving the local geometric structure and global distribution characteristics, remove redundant and noisy features to obtain a compact and discriminative low-dimensional feature representation. Its objective function can be written as:

[0045] In the formula, Represents the projection matrix. Representation matrix The transpose of Z, where Z represents the data matrix. Represents sample points, Represents the Laplace matrix, Degree matrix, Represents the adjacency matrix. This represents the adjacency weight between two samples before dimensionality reduction; that is, the closer the samples are, the greater the weight, and the smaller the distance, the greater the weight. Its formula is expressed as follows:

[0046] S43. The entire dataset after dimensionality reduction is initialized into a granular envelope, whose initial center is the average of all samples. At this point, the purity of the initial granular envelope is the lowest, failing to meet the set threshold condition. Assume there are k classes of samples in the granular envelope, and... Let represent the number of samples in class i. The formula for calculating purity can be expressed as:

[0047] S44. Split the particle envelope: Randomly select K-1 remaining heterogeneous points as new centers, and redistribute all samples in the particle envelope to these K centers to generate K new particle envelopes.

[0048] S44. Calculate the purity value of each sub-envelope. If the set purity threshold is not met, continue splitting the sub-envelope until the threshold is reached.

[0049] S5. Train a lightweight classifier (such as XGBoost) on the particle envelope sample space to achieve rapid identification and classification of 7 common wound microorganisms.

[0050] The following is an analysis of the experiments and results: To verify the effectiveness of the knowledge-data dual-driven envelope fusion-based wound bacterial spectral detection and application method proposed in this embodiment, this section of the experiment collects spectral data of patients' wounds through clinical trials, and uses microbial culture results as the gold standard to study the spectral characteristics of wound infection.

[0051] We first divided the training and test sets according to the patients, as shown in Table 1, to ensure that the test set samples and training set samples come from different patients during the model training process. Table 1. Patient segmentation for different task training and test sets.

[0052] The spectral data extraction process of this method includes four core steps: wound area delineation, raw spectrum acquisition, data preprocessing, and feature parameter analysis. The aim is to obtain high-quality, classifiable feature information from the raw signal.

[0053] Methods that rely solely on wound appearance to diagnose infection type are significantly subjective and uncertain, such as... Figure 1 As shown, the appearance characteristics of different infected wounds are difficult to distinguish directly. By analyzing the near-infrared spectral imaging results of the wound and extracting the spectral curves of each pixel (…),… Figure 2 This provides an objective quantitative basis for identification.

[0054] The following are Figure 2 The five drawings in the document will be described in detail: OriginalSpectra: This graph presents the raw near-infrared spectral signal of untreated wound tissue. The horizontal axis represents wavelength (nm), and the vertical axis represents the spectral signal value. The graph shows the raw peak shapes and fluctuations of the spectrum at different wavelengths, containing spectral information related to wound tissue structure and microbial metabolites. However, it also contains interference such as noise and baseline drift, serving as the basic data for subsequent spectral preprocessing.

[0055] SGSmoothedSpectra (Savitzky-Golay smoothed spectrum): This figure shows the result of smoothing the original spectrum using a Savitzky-Golay filter. The horizontal axis still represents wavelength (nm), and the vertical axis represents the spectral signal value. After smoothing, the noise in the spectrum is effectively removed, the curve is smoother and more continuous, and the key peak characteristics of the original spectrum are preserved. This provides high-quality preprocessed data for subsequent extraction of valuable spectral information and construction of a feature sample space.

[0056] Top 15 Most Important Features: This graph, with feature number on the horizontal axis and feature importance score or corresponding wavelength as the relevant dimension, presents the 15 most critical features for microbial classification selected from spectral data. These features are core spectral information that the model learns to determine, effectively distinguishing different infection types (such as fungi and bacteria, different types of bacteria), providing high-value feature input for subsequent classification model training.

[0057] Cumulative Feature Importance (CPI) plot: This plot, with the number of features on the horizontal axis and the cumulative feature importance score on the vertical axis, shows the cumulative trend of feature importance as the number of features increases. This plot demonstrates that a small number of key features (such as the top 15 important features) can achieve high cumulative importance, indicating that the selected core features can efficiently support microbial classification tasks and providing a basis for simplifying models and improving detection efficiency.

[0058] This invention first preprocesses the raw spectra to improve the signal-to-noise ratio and enhance inter-class differentiation. It then employs a dual feature extraction strategy combining knowledge-driven and data-driven approaches to screen high-value features most sensitive to infection type from massive spectral data. Only a small number of key features are needed to achieve high cumulative importance, laying a solid foundation for accurate and automated identification.

[0059] Finally, experiments were conducted on four classifiers (RandomForest, KNeighbors, DecisionTree, and AdaBoost), and the specific classification results are shown in Table 2. Table 2 Classification results of different classifiers

[0060] As shown in Table 2, different classifiers exhibit significant differences in their performance in identifying wound infection types and seven common wound bacteria. Overall, AdaBoost outperforms other classifiers in accuracy across all infection types and bacterial species, particularly achieving 83.3% and 83.6% accuracy on Kpn and Pae, respectively, demonstrating strong generalization ability and robustness. DecisionTree follows closely, exceeding 73% accuracy on most bacterial species, indicating that while single decision trees are prone to overfitting, they can still effectively capture class discrimination patterns under the current feature representation. In contrast, RandomForest's results are relatively lower, possibly due to redundant splitting in its high-dimensional fused feature space, leading to unclear classification boundaries. KNeighbors performs well on some bacterial species (such as Kpn), but its overall performance is highly volatile, indicating its sensitivity to local density distribution in the feature space and susceptibility to noise. In summary, AdaBoost shows the most significant improvement, validating its ability to fully exploit discriminative information between features under a weak classifier ensemble strategy, providing a methodological basis for establishing high-precision and fast identification models.

[0061] Table 3. Comparison of effectiveness under different sample data (based on AdaBoost classifier)

[0062] Table 3 presents the effectiveness comparison results under different sample data, aiming to verify the advantages of the proposed method. PSO-Cls represents a knowledge-driven sample classification method, which obtains optical characteristic parameters by analyzing spectral signals and adaptively fine-tunes the classifier parameters using a particle swarm optimization algorithm to improve classification performance. BdasNet represents a data-driven sample classification method, which uses a one-dimensional convolutional neural network (1D-CNN) to extract multi-level features from spectral signals and completes classification in an end-to-end manner. The results in the table show that the three sample types exhibit a progressively improving performance trend when using the AdaBoost classifier. While the BdasNet algorithm, relying solely on data-driven samples extracted by a deep convolutional network, already shows some advantage in classification accuracy for various bacteria, its overall performance is still limited by the model's insufficient interpretability for pathology. In contrast, the PSO-Cls algorithm, based on knowledge-driven samples obtained from diffuse reflection theory, achieves higher accuracy across all bacterial species, indicating that optical characteristic parameters with clear biophysical significance can provide more robust feature support for classification. Building upon both methods, the envelope sample, obtained through feature cascading and manifold dimensionality reduction combined with envelope learning, demonstrated the best performance. It achieved the highest classification accuracy across all seven bacterial species, with Sau, Kpn, and Pae reaching 82.8%, 83.3%, and 83.6%, respectively, significantly outperforming the single-modality approach. This result indicates that the envelope sample can effectively overcome the performance bottleneck of single-modality feature extraction by fully exploring local structural relationships and enhancing the discriminative power between features, while simultaneously considering both pathological interpretability and data-driven representation capabilities. This provides a more reliable solution for the rapid and accurate identification of bacteria in wounds.

[0063] The wound microbial detection method involved in this invention achieves accurate capture and efficient identification of microbial spectral features through a dual-driven envelope fusion mechanism of knowledge data. Although the embodiments of this invention use seven common bacteria (Sau, Eco, Efa, Pvu, Kpn, Aba, Pae) and fungi as research objects for experimental verification, the application scope of this method is not limited to the above-mentioned specific microbial species. Since the metabolic activities of different bacteria and fungi can lead to characteristic near-infrared spectral responses in wound tissue, the biophysical laws contained in their spectral signals are universal. This invention, through a dual-driven mode of analyzing optical characteristic parameters using diffuse reflection theory and extracting high-order abstract features using a one-dimensional convolutional neural network, can effectively adapt to the spectral feature extraction needs of various microorganisms. At the same time, the dual-stream feature fusion mechanism and the particle envelope sample construction method can adaptively mine the common and specific laws of different microbial spectral signals, and after model training, it can be extended to the detection and identification of other bacteria and fungi. Therefore, the technical solution of the present invention can be optimized by expanding the spectral sample library of corresponding microorganisms according to actual detection needs, thereby achieving rapid and non-invasive detection of more types of bacteria and fungi, and providing flexible and reliable technical support for the early diagnosis of microbial infections of various wounds in clinical practice.

[0064] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting microorganisms on a wound surface based on double driving envelope of optical fiber spectrum knowledge data, characterized in that, The detection method comprises the following steps: S1, using a fiber spectrometer to collect multi-band near-infrared spectrum signals of wound tissue, and performing normalization and denoising preprocessing operations on the original spectrum data; S2, based on the diffuse reflection theory model, the optical characteristic parameter set is parsed from the pretreated spectral signal, including the reduced scattering coefficient and the absorption coefficient , a knowledge-driven feature sample space is constructed; S3, using a one-dimensional deep convolutional neural network 1D-CNN to extract high-order abstract features of the spectrum signals through multiple convolution kernels, and constructing a data-driven feature sample space; S4, establishing a double-flow feature fusion mechanism, realizing spatial dimension fusion through feature cascading, and embedding and mapping the fused high-dimensional feature samples based on a manifold dimension reduction method, and generating a particle envelope sample according to the knowledge-data sample after dimension reduction; S5, inputting the particle envelope sample as a training set into a lightweight classifier for model training, establishing a high-dimensional mapping relationship from spectrum features to microbial species through iterative optimization, inputting the to-be-tested wound spectrum data into the model after processing, and outputting classification and discrimination results of 7 common microorganisms; Wherein, the expression of the diffuse reflection theory model is: ； In the formulae: , , , ; is the radius of the light collection zone, = 0.1 ; parameter depends on the refractive index of the medium , = 3.

2.

2. The method of claim 1, wherein the method is characterized by: Step S1 specifically comprises: S11, preheating the spectrometer, sequentially turning on the spectrometer, the light source and the matching software SpectraSuite, collecting a control spectrum with a diffuse reflection standard plate, collecting a dark spectrum by closing the light source, collecting the NIR spectrum of the diffuse reflection standard plate by re-opening the light source and setting the reflection amplitude to 100%, and collecting the NIR spectrum signal of the wound after 5 minutes of stabilization; S12, performing normalization transformation on the original spectrum signal, and the normalization calculation formula is: ； wherein represents the spectral signal value; is the broadband light intensity measured from the wound; is the background light intensity with the light source turned off; represents the light intensity measured from the reflective standard; S13, performing smoothing processing on the normalized spectrum signal by using a Savitzky-Golay filter, and the calculation formula is: ； wherein represents a filter function, is the spectral signal value at a position that is offset by from the current point , and is the number of data points within the filter window.

3. The method of claim 1, wherein the method is characterized by: In step S2, the optical characteristic parameters are extracted by least square fitting, and the fitting optimization formula is: ； wherein is the spectral signal value, is the spectral theoretical value.

4. The method of claim 1, wherein the method is characterized by: Step S3 specifically comprises: S31, establishing a one-dimensional convolutional neural network classification model through one-dimensional convolution, pooling and full connection layer; S32, training the preprocessed spectrum signal, and adjusting the parameters of each layer by using an Adam optimization algorithm, when the model optimization is stopped and the loss function no longer decreases, saving the final model and stopping optimization iteration; S33, using the final model to extract high-order abstract features of the spectrum signal.

5. The method of claim 4, wherein the method is characterized by: In step S3, the convolution operation formula of the one-dimensional convolutional neural network is: ； The pooling operation formula is: ； The feature transformation formula of the full connection layer is: ； In the formula, W is the weight parameter of the lth layer, which is a convolution kernel in the convolution layer and the pooling layer, and a weight matrix in the full connection layer; H is the output feature of the lth layer, which is a feature map in the convolution layer and the pooling layer, and a feature vector in the full connection layer; b is a bias term.

6. The method of claim 1, wherein the method is characterized by: In step S3, cross-entropy is used as the loss function, and the formula is: ； wherein is the cross-entropy loss value; is the total number of classes in the classification task; is the true label; is the probability of the sample class predicted by the model.

7. The method of claim 1, wherein the method is characterized by: Step S4 specifically comprises: S41, performing spatial dimension fusion by cascading the optical parameter features and the deep convolution features to obtain a knowledge-data sample; S42, performing manifold embedding mapping on the high-dimensional knowledge-data sample to retain the local neighborhood structure and eliminate redundant information; S43, the entire data set is first initialized as a particle envelope, the initial center of which is the average value of all samples, at this time, the purity of the initial particle envelope is the lowest and does not meet the set threshold condition; S44, split the particle envelope: randomly select K-1 remaining isomorphic points as new centers, and reassign all samples in the particle envelope to the K centers to generate K new particle envelopes; S45, calculate the purity value of each sub-envelope, if the purity threshold is not met, continue to split the sub-envelope until the threshold is reached.

8. The method of claim 7, wherein the method is characterized by: The spatial dimension fusion formula of the feature cascade in step S4 is: ； In the formula, represents concatenation splicing, and is a knowledge type sample feature, is a data type sample feature.

9. The method of claim 7, wherein the method is characterized by: The objective function of manifold dimension reduction in step S4 is: ； In the formula, denotes a projection matrix, denotes a matrix transpose of the matrix, Z denotes a data matrix, denotes a sample point, denotes a Laplacian matrix, denotes a degree matrix, denotes an adjacency matrix, denotes an adjacency weight between two samples before dimension reduction, that is, the closer the distance, the greater the weight, and the smaller the distance, the greater the weight; The abutment weight calculation formula is: ; In the formula, and respectively represent two different high-dimensional feature samples.

10. The method of claim 7, wherein the method is based on a fiber-optic spectrum knowledge data double-driven envelope of a wound microbe detection method. The particle envelope purity calculation formula in step S4 is: ； In the formula, represents the number of the i-th type of samples.

Images