Exosome automatic identification and analysis method based on microscopic imaging

Abstract

The invention provides an exosome automatic identification and analysis method based on microscopic imaging, and belongs to the technical field of exosome identification and analysis, and the method comprises the steps: collecting an exosome fluorescence microscopic image, carrying out the preprocessing, and obtaining an exosome region through adaptive threshold segmentation and morphological processing. A GPU is adopted to process and extract morphological, texture and statistical feature matrixes in parallel, and meanwhile, the CPU recognizes an abnormal region. And combining the features to form a region feature set, inputting the region feature set into the preliminary screening model for scoring and screening high-credibility regions. Deep feature extraction is carried out on the high-credibility region, and multiple features are fused by using an attention fusion mechanism to construct vectors; and inputting a depth classifier to obtain a classification probability matrix and a screening result. And finally outputting the number, position, classification and credibility score of the exosomes. The problems that an existing exosome analysis method often depends on manual recognition, and automatic and high-throughput analysis is difficult to achieve are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of exosome recognition. Specifically, it relates to an automatic recognition and analysis method for exosomes based on microscopic imaging. Background Art

[0002] Exosomes are membrane vesicle structures secreted by cells, usually with diameters between 30 - 150 nm. They carry information molecules such as proteins, nucleic acids, and lipids within cells and play important roles in physiological processes such as intercellular communication, immune regulation, and signal transduction. Therefore, accurately detecting and analyzing the characteristics of exosomes is of great significance for in - depth understanding of biological processes, disease diagnosis, and treatment.

[0003] Currently, commonly used exosome detection and analysis techniques include dynamic light scattering, immunoaffinity chromatography, and electron microscopy imaging, etc. Among them, fluorescence microscopy imaging technology based on fluorescence labeling has been widely used in the field of exosome research due to its intuitive, sensitive, and high - throughput characteristics. This technology can visually obtain information such as the morphology, size, and quantity of exosomes by fluorescence - labeling exosomes and then using a fluorescence microscope for imaging observation.

[0004] However, existing exosome analysis methods often rely on manual recognition and are difficult to achieve automated and high - throughput analysis. Summary of the Invention

[0005] In view of this, the present invention provides an automatic recognition and analysis method for exosomes based on microscopic imaging, which does not require manual recognition and can perform automated and high - throughput analysis.

[0006] The present invention is implemented as follows:

[0007] The present invention provides a method for automatic recognition and analysis of exosomes based on microscopic imaging, comprising the following steps: collecting fluorescence microscopic images of exosomes, and preprocessing the fluorescence microscopic images of exosomes to obtain preprocessed images; performing adaptive threshold segmentation and morphological processing on the preprocessed images to obtain exosome regions; using a graphics processing unit parallel processing stream to extract features from the exosome regions, and calculating to obtain the morphological feature matrix, the texture feature matrix, and the statistical feature matrix, and combining them into a region feature set; inputting the region feature set into a pre-trained preliminary screening model for preliminary screening to obtain a credibility score matrix of the regions; and according to a preset credibility threshold, screening high-credibility regions, inputting them into a feature enhancement equation set for deep feature extraction to obtain a multi-scale deep feature matrix; using an attention fusion mechanism to perform adaptive weight fusion on the region feature set to construct a fusion feature vector; inputting the fusion feature vector into a pre-trained deep classifier to obtain an exosome classification probability matrix; and performing result screening according to a confidence threshold, and outputting exosome recognition results, including the number of exosomes, position coordinates, classification labels, and credibility scores. Specifically, it includes:

[0008] S10. Collect fluorescence microscopic images of exosomes, and preprocess the fluorescence microscopic images of exosomes for denoising, contrast enhancement, and illumination compensation to obtain preprocessed images;

[0009] S20. Perform adaptive threshold segmentation and morphological processing on the preprocessed images to obtain exosome regions;

[0010] S30. Use a graphics processing unit parallel processing stream to extract features from the exosome regions, including: a morphological feature matrix containing area, perimeter, roundness, and gray value, a texture feature matrix containing gray-level co-occurrence matrix features and local binary pattern features, and a statistical feature matrix containing mean, variance, skewness, and kurtosis; at the same time, start the central processing unit to extract abnormal regions, including edge-blurred regions, overlapping regions, incomplete regions, and regions with abnormal sizes; combine the morphological feature matrix, the texture feature matrix, and the statistical feature matrix to form a region feature set;

[0011] S40. Input the region feature set into a pre-trained preliminary screening model for preliminary screening to obtain a credibility score matrix of the regions; and according to a preset credibility threshold, screen high-credibility regions;

[0012] S50. Input the high-credibility regions into a feature enhancement equation set for deep feature extraction to obtain a multi-scale deep feature matrix;

[0013] S60. Use an attention fusion mechanism to adaptively weight-fuse the morphological feature matrix, the abnormal region features, and the multi-scale depth feature matrix to construct a fused feature vector;

[0014] S70. Input the fused feature vector into a pre-trained deep classifier to obtain an exosome classification probability matrix; and perform result screening according to a confidence threshold;

[0015] S80. Output the exosome recognition results, including the number of exosomes, position coordinates, classification labels, and credibility scores.

[0016] Based on the above solution, an automatic exosome recognition and analysis method based on microscopic imaging according to the present invention can also be improved as follows:

[0017] Among them, the feature enhancement equation set includes a scale mapping equation, a feature enhancement equation, a contrast optimization equation, and a noise suppression equation, and there are cascade and feedback relationships among the equations in the feature enhancement equation set.

[0018] Further, the scale mapping equation is used to construct feature representations at different scales. The inputs include the original image scale, target scale parameters, scale transformation coefficients, scale interpolation parameters, scale weight coefficients, and it receives the feedback features of the feature enhancement equation. The output is a feature mapping matrix at multiple scale levels.

[0019] Further, the feature enhancement equation is used to enhance the local structural features of exosomes. The inputs include the feature mapping matrix, local structure tensor, enhancement weight matrix, structural similarity threshold, local contrast parameter, local variance threshold, neighborhood size parameter, and it receives the contrast feedback information of the contrast optimization equation. The output is an enhanced feature matrix;

[0020] The contrast optimization equation is used to enhance the contrast between exosomes and the background. The inputs include the enhanced feature matrix, contrast threshold parameter, brightness equalization coefficient, gray-scale stretching factor, global contrast gain, local contrast gain, edge enhancement coefficient, and it receives the noise evaluation feedback of the noise suppression equation. The output is an optimized contrast feature matrix;

[0021] The noise suppression equation is used to suppress random noise in the feature extraction process. The inputs include the optimized contrast feature matrix, noise evaluation parameters, noise threshold, smoothing coefficient, edge preservation parameter, detail preservation coefficient, noise variance estimation, and it feeds back the noise suppression result to the scale mapping equation. The output is a final denoised feature matrix.

[0022] Further, the edge-blurred region is a region where the gray-scale value gradient is less than a preset threshold, and the value range of the preset threshold is 15 to 45.

[0023] Further, the overlapping region is a region where multiple regions overlap locally, and the value range of the overlap rate threshold is from 0.2 to 0.6.

[0024] Further, the incomplete region is an incomplete region located at the edge of the image, and the value range of the edge distance threshold is from 5 to 20 pixels.

[0025] Further, the region with abnormal size is a region where the area or perimeter significantly deviates from the normal range. The value range of the area deviation threshold is plus or minus 30%, and the value range of the perimeter deviation threshold is plus or minus 25%.

[0026] Further, the classification labels include pure exosomes, degraded exosomes, exosome aggregates, and non-exosome particles.

[0027] Further, the preliminary screening model is used to preliminarily evaluate and screen exosome regions. The input is the region feature set, and the output is the region credibility score matrix. The structure includes a feature input layer, multiple convolutional layers, a fully connected layer, and an output layer;

[0028] The steps for obtaining the training data set of the preliminary screening model include: collecting exosome fluorescence microscopic image samples, performing image annotation, extracting features, constructing a training set and a validation set;

[0029] The training steps of the preliminary screening model include: data preprocessing, model parameter initialization, iterative training, model validation, and model saving;

[0030] The deep classifier is used to accurately classify exosomes. The input is the fusion feature vector, and the output is the exosome classification probability matrix. The structure includes a feature extraction module, an attention mechanism module, and a classification module;

[0031] The steps for obtaining the training data set of the deep classifier include: collecting multi-type exosome samples, making a standard data set, feature extraction, data augmentation, and data set division;

[0032] The training steps of the deep classifier include: data standardization, model structure definition, loss function optimization, batch training, and model evaluation.

[0033] Compared with the prior art, the present invention can overcome the limitations that existing exosome analysis methods often rely on manual recognition and are difficult to achieve automated and high-throughput analysis, and can achieve rapid detection and accurate classification of exosomes. The method mainly includes the following technical innovations and improvements:

[0034] 1. The image preprocessing module adopts algorithms such as adaptive denoising, contrast enhancement, and illumination compensation, effectively improving the image quality and laying a good foundation for subsequent analysis.

[0035] 2. The feature extraction module integrates morphological features, texture features, statistical features, and abnormal region features, comprehensively describing the geometric size, gray-scale distribution, and internal structure of exosomes. Compared with traditional simple morphological indicators, it has been greatly improved.

[0036] 3. Deep learning technology is used to construct a preliminary screening model and an accurate classifier, which can adaptively learn the complex pattern features of exosomes and greatly improve the accuracy of classification and recognition.

[0037] 4. Adaptive mechanisms such as scale mapping, feature enhancement, contrast optimization, and noise suppression are introduced in feature extraction and fusion, realizing the dynamic optimization of features and further enhancing the robustness of the algorithm.

[0038] 5. The entire analysis process is highly automated, realizing the rapid detection and classification of a large number of exosome samples and meeting the high-throughput requirements of biomedical research.

[0039] In summary, the automatic recognition and analysis method for exosomes proposed by the present invention, through technological innovations such as image preprocessing, multi-dimensional feature extraction, adaptive feature fusion, and deep classification, significantly improves the intelligent level and analysis performance of exosome analysis, providing strong analysis support for related research in the biomedical field. Brief Description of the Drawings

[0040] Figure 1 is the flow chart of the method of the present invention;

[0041] Figure 2 is the image after preprocessing the exosome region in Example 2;

[0042] Figure 3 is the distribution characteristics of exosomes from different sample sources in Example 2. Detailed Embodiments

[0043] To make the objectives, 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 with reference to the accompanying drawings in the embodiments of the present invention.

[0044] As Figure 1 shown, it is the flow chart of an automatic recognition and analysis method for exosomes based on microscopic imaging provided by the present invention. This method includes the following steps:

[0045] S10. Collect fluorescence microscopic images of exosomes, and preprocess the fluorescence microscopic images of exosomes for denoising, contrast enhancement, and illumination compensation to obtain a preprocessed image;

[0046] S20. Perform adaptive threshold segmentation and morphological processing on the preprocessed image to obtain an exosome region;

[0047] S30. Use a graphics processing unit parallel processing stream to extract features from the exosome region, including: a morphological feature matrix containing area, perimeter, roundness, and gray value, a texture feature matrix containing gray-level co-occurrence matrix features and local binary pattern features, and a statistical feature matrix containing mean, variance, skewness, and kurtosis; at the same time, start the central processing unit to extract abnormal regions, including edge-blurred regions, overlapping regions, incomplete regions, and regions with abnormal sizes; combine the morphological feature matrix, the texture feature matrix, and the statistical feature matrix to form a region feature set;

[0048] S40. Input the region feature set into a pre-trained preliminary screening model for preliminary screening to obtain a credibility score matrix for the region; and screen high-credibility regions according to a preset credibility threshold;

[0049] S50. Input the high-credibility regions into a feature enhancement equation set for deep feature extraction to obtain a multi-scale deep feature matrix;

[0050] S60. Use an attention fusion mechanism to adaptively weight-fuse the morphological feature matrix, the abnormal region features, and the multi-scale deep feature matrix to construct a fusion feature vector;

[0051] S70. Input the fusion feature vector into a pre-trained deep classifier to obtain an exosome classification probability matrix; and perform result screening according to a confidence threshold;

[0052] S80. Output the exosome recognition results, including the number of exosomes, position coordinates, classification labels, and credibility scores.

[0053] The following describes the specific implementation manners of the above steps in detail:

[0054] The specific implementation manner of step S10 is: First, collect fluorescence microscopic images of exosomes. To improve the accuracy and stability of subsequent processing, it is necessary to preprocess the collected original images. The preprocessing includes three sub-steps: denoising, contrast enhancement, and illumination compensation:

[0055] 1. Denoising processing: Since there may be random noise interference during the acquisition process, a low-pass filter is needed to denoise the image. A Gaussian low-pass filter or a median filter can be selected. By adjusting the filter kernel size and standard deviation parameters, noise can be effectively suppressed while retaining the target edge information.

[0056] 2. Contrast enhancement: There is a certain intensity difference between the exosome fluorescence signal and the background. To improve the recognizability of the target area, the image needs to be processed for contrast enhancement. The histogram equalization algorithm can be used. By adjusting the gray-scale distribution of the image and stretching the gray-scale range of the target area, the contrast can be enhanced.

[0057] 3. Illumination compensation: Since there may be an uneven illumination distribution during the fluorescence imaging process, resulting in local differences in image brightness. Therefore, illumination compensation is required. The method of estimating and subtracting the illumination model based on Fourier transform can be used to effectively eliminate the influence of non-uniform illumination in the image.

[0058] After the above preprocessing steps, the noise of the original image is suppressed, and the contrast and brightness distribution are optimized, creating good conditions for subsequent exosome region segmentation and feature extraction.

[0059] The specific implementation of step S20 is: performing adaptive threshold segmentation and morphological processing on the preprocessed image to obtain the region of exosomes. It specifically includes the following sub-steps:

[0060] 1. Adaptive threshold segmentation: The OTSU method is used to automatically calculate the threshold T, and the image is binarized using this threshold. The OTSU algorithm can adaptively determine the optimal segmentation threshold by maximizing the between-class variance principle, thus effectively separating the exosome region.

[0061] 2. Morphological processing: Morphological processing such as opening operation and closing operation is performed on the binarized image. The opening operation can remove noise points and small fragments, and the closing operation can fill small holes in the region. Through these morphological operations, the boundary and shape of the exosome region can be further optimized.

[0062] 3. Region labeling: The connected component analysis method is used to label the regions of the segmentation result. Through 8-neighborhood connectivity judgment, adjacent exosome regions can be segmented into independent target blocks, laying a foundation for subsequent feature extraction.

[0063] After these processes, the exosome regions in the original image have been accurately extracted, providing reliable input data for subsequent in-depth analysis and recognition.

[0064] The specific implementation of step S30 is as follows: Use a graphics processing unit (GPU) with parallel processing to extract features from the extracted exosome regions, including three categories: morphological features, texture features, and statistical features. At the same time, start the central processing unit (CPU) to extract the features of abnormal regions. These feature matrices will be combined into a region feature set to provide support for subsequent classification and recognition. The detailed process is as follows:

[0065] 1. Extraction of morphological features: It includes four indicators: area, perimeter, roundness, and gray value. The area and perimeter are obtained by pixel statistical counting, and the roundness is calculated according to the geometric relationship between the area and perimeter. The gray value feature is represented by the average value of the pixels within the region. These morphological features comprehensively describe the geometric scale and gray distribution characteristics of exosomes.

[0066] 2. Extraction of texture features: Based on the gray-level co-occurrence matrix, four texture features, namely contrast, homogeneity, entropy, and correlation, are extracted. The gray-level co-occurrence matrix can effectively reflect the correlation of the pixel gray distribution in the image, and these statistical features can describe the texture information of exosomes.

[0067] 3. Extraction of statistical features: It includes four indicators: gray mean, standard deviation, skewness, and kurtosis. These higher-order statistical moments describe the detailed features of the gray distribution of the exosome region and help to distinguish different types of exosomes.

[0068] 4. Extraction of features of abnormal regions: For possible edge-blurred regions, overlapping regions, incomplete regions, and regions with abnormal sizes, corresponding indicators are extracted for quantitative description. These features can effectively identify and quantitatively evaluate various abnormal situations.

[0069] Through GPU parallel acceleration, the feature extraction of a large number of exosome regions can be efficiently completed. These feature matrices comprehensively reflect the morphological, texture, and statistical characteristics of exosomes, laying a solid foundation for subsequent classification and recognition.

[0070] The specific implementation of step S40 is as follows: Input the region feature set extracted in step S30 into a pre-trained preliminary screening model to obtain a credibility score matrix for each region. Then, according to the set credibility threshold, high-credibility exosome regions are screened out. It specifically includes the following sub-steps:

[0071] 1. Training of the preliminary screening model: Collect a large number of exosome fluorescence microscopy image samples, perform manual annotation, extract region features, and construct a training set and a validation set. Then, use these data to train a preliminary screening model that includes a feature input layer, multiple convolutional layers, a fully connected layer, and an output layer. During training, steps such as data preprocessing, parameter initialization, iterative optimization, model validation, and saving are required.

[0072] 2. Regional credibility assessment: Input the regional feature set extracted in step S30 into the trained preliminary screening model to output the credibility score matrix for each region. Here, the credibility reflects the probability that the region is an exosome.

[0073] 3. Threshold screening: Set a suitable credibility threshold, eliminate the regions with scores lower than this threshold, and only retain the exosome regions with high credibility. This step can effectively remove some false positive regions introduced by noise or segmentation errors and retain the true exosome targets.

[0074] Through the training and application of the preliminary screening model, a large number of exosome regions can be preliminarily evaluated quickly, laying a foundation for subsequent in-depth analysis and precise classification, and improving the robustness and reliability of the entire recognition process.

[0075] The specific implementation of step S50 is as follows: Input the exosome regions with high credibility screened in step S40 into the feature enhancement equation set for in-depth feature extraction, and finally obtain the multi-scale depth feature matrix. The feature enhancement equation set includes a scale mapping equation, a feature enhancement equation, a contrast optimization equation, and a noise suppression equation. These four equations achieve the adaptive optimization of features through cascading and feedback. Specifically as follows:

[0076] 1. Scale mapping equation: This equation is used to construct feature representations at different scales. It adopts the multi-scale decomposition strategy of the Gaussian pyramid, highlights the contribution of the main scale by weighted fusion of features at different scales, and introduces a feedback term to achieve dynamic adjustment of feature enhancement. The scale parameter is determined by logarithmic equidistant sampling, the Gaussian kernel parameter is proportional to the scale parameter, and the weight coefficient is optimized through cross-validation.

[0077] 2. Feature enhancement equation: This equation aims to enhance the local structural features of exosomes. It combines nonlinear mapping, differential operators, and a feedback mechanism. The first term is used for feature enhancement, the second term is used for edge detection, and the third term describes the change rate of features. The contrast is adaptively adjusted through the feedback term. The smoothing parameter is estimated by local variance, the weight coefficient is optimized by the gradient descent method, and the feedback coefficient is adjusted by an iterative method.

[0078] 3. Contrast optimization equation: This equation aims to improve the contrast between exosomes and the background. It adopts a strategy combining local and global normalization, retains edge information through the gradient term, and the feedback term is used to suppress the influence of noise on the contrast. Local statistics are calculated through a sliding window, and global statistics are obtained through histogram analysis. The weight coefficient is fitted by the least squares method.

[0079] 4. Noise suppression equation: This equation is used to suppress random noise during the feature extraction process. It designs three functional terms: Gaussian suppression, edge preservation, and detail preservation, which correspond to different types of noise features respectively. Through a feedback mechanism, it realizes collaborative optimization with scale mapping. The noise statistical parameters are estimated through the background region, while the edge and detail preservation parameters are analyzed through structural similarity. The weight coefficients are optimized through grid search.

[0080] After the cascade processing of these feature enhancement equations, the multi-scale structural features of exosomes can be fully mined, the detail information can be enhanced, the contrast can be improved, and the noise interference can be suppressed. Finally, a high-quality depth feature matrix is obtained.

[0081] The specific implementation of step S60 is as follows: An attention fusion mechanism is adopted to adaptively weight and fuse the morphological features, abnormal region features extracted in step S30, and the multi-scale depth features obtained in step S50 to construct a comprehensive fusion feature vector. The specific process is as follows:

[0082] 1. Feature matrix normalization: First, normalize various types of feature matrices (morphological, texture, statistical, abnormal region, depth) to eliminate the differences in dimension and magnitude, and prepare for subsequent weighted fusion.

[0083] 2. Attention mechanism calculation: A learnable attention mechanism is adopted to assign adaptive weight coefficients to each type of feature matrix. Specifically, by calculating the dot product of each type of feature vector and a group of learnable parameter vectors, and then performing Softmax normalization to obtain the weight coefficients. In this way, the contribution of different features in the fusion feature can be dynamically adjusted according to their importance.

[0084] 3. Feature fusion: Multiply each type of normalized feature matrix by the corresponding attention weight, and then splice them together to form the final fusion feature vector. This fusion feature comprehensively describes the multi-faceted features of exosomes and lays a foundation for subsequent accurate classification.

[0085] The attention fusion mechanism can adaptively adjust the weight of each feature in the final feature vector according to its importance, improving the robustness and discriminability of feature expression, and creating good input conditions for subsequent deep classifier training.

[0086] The specific implementation of step S70 is as follows: Input the fusion feature vector obtained in step S60 into a pre-trained deep classifier model to output the classification probability matrix of exosomes, and perform result screening according to the set confidence threshold. It specifically includes the following sub-steps:

[0087] 1. Deep classifier model training: First, a large number of exosome samples of different types need to be collected to create a standard training dataset. Then, extract the features of these samples and perform preprocessing such as data augmentation. Next, define the model structure of the deep classifier, including a feature extraction module, an attention mechanism module, and a classification module. Use the training dataset to perform end-to-end supervised learning training on the model, optimize the loss function until the model performance converges.

[0088] 2. Exosome classification prediction: Input the fusion feature vector constructed in step S60 into the trained deep classifier model, and a probability matrix of each exosome region belonging to different categories can be output. Here, the categories can be exosome subtypes, such as exosomes, microvesicles, metabolosomes, etc.

[0089] 3. Result screening: Screen the classification results according to a preset confidence threshold. Only exosome regions with classification probabilities exceeding this threshold are finally determined as valid targets. This step can further improve the accuracy and reliability of recognition.

[0090] Through the training and application of the deep learning classifier, accurate identification of different types of exosomes can be achieved. The input of the fusion feature vector and the setting of a reasonable confidence threshold are the keys to ensuring classification performance.

[0091] The output results of step S80 include: the number of exosomes, position coordinates, classification labels, and confidence scores. These output information comprehensively describe the detection and recognition of exosomes in the image, providing valuable analysis basis for subsequent biomedical applications.

[0092] The following details the calculation processes or equations involved in the present invention.

[0093] 1. The calculation of the morphological feature matrix is specifically represented as follows:

[0094]

[0095] In the formula, F m is the morphological feature matrix; A i is the area of the i-th region, obtained by counting pixel points; P i is the perimeter of the i-th region, obtained by counting boundary pixel points; C i is the circularity of the i-th region, and the calculation formula is G i is the average gray value of the i-th region, and the calculation formula is where R i is the i-th region, I(x, y) is the gray value of the pixel point; n is the total number of regions.

[0096] 2. The calculation of the texture feature matrix is specifically expressed as follows:

[0097]

[0098] In the formula, F t is the texture feature matrix; H ij is the j-th gray-level co-occurrence matrix feature of the i-th region, and the calculation is as follows:

[0099] is the contrast;

[0100] is the homogeneity;

[0101] is the entropy;

[0102] is the correlation;

[0103] where p(k, l) is the element of the gray-level co-occurrence matrix; L is the number of gray levels; μ x , μ y is the mean of the marginal distribution of the gray-level co-occurrence matrix.

[0104] 3. The calculation of the statistical feature matrix is specifically expressed as follows:

[0105]

[0106] In the formula, F s is the statistical feature matrix; μ i is the gray mean of the i-th region:

[0107] σ i is the gray standard deviation of the i-th region:

[0108] s i is the gray skewness of the i-th region:

[0109] k i is the gray kurtosis of the i-th region:

[0110] 4. The calculation of the abnormal region feature is specifically expressed as follows:

[0111]

[0112] In the formula, F a is the abnormal region feature matrix; E i is the edge blur degree of the i-th region: where B i is the region boundary, is the image gradient;

[0113] O i is the overlap degree of the i-th region: where A overlap is the overlapping area with adjacent regions;

[0114] I i is the integrity of the i-th region: where d i is the minimum distance from the region to the image edge, and D is the length of the image diagonal;

[0115] S i is the size abnormality of the i-th region: where are the average area and perimeter of all regions respectively.

[0116] 5. The calculation of the multi-scale depth feature matrix is specifically expressed as follows:

[0117] F d ={F d1 ,F d2 ,…,F dK};

[0118] where, F dk is the depth feature matrix at the k-th scale:

[0119] F dk =W k ·I k +b k ;

[0120] In the formula, I k is the input feature at the k-th scale; W k is the weight matrix; b k is the bias term; K is the number of scale layers.

[0121] 6. The calculation of the fused feature vector is specifically expressed as follows:

[0122]

[0123] In the formula, F fusion is the fused feature vector; α m ,α t ,α s ,α a ,α d are the weight coefficients of each feature matrix, calculated through the following attention mechanism:

[0124]

[0125] where, w iis a learnable parameter vector.

[0126] Description of the acquisition methods for all features and parameters:

[0127] 1. Acquisition of morphological features:

[0128] (a) Region segmentation uses an adaptive threshold method, and the threshold T is calculated by the OTSU algorithm;

[0129] (b) Boundary extraction uses an edge tracking algorithm with 8-neighborhood connectivity determination;

[0130] (c) Area and perimeter are calculated by pixel statistics, and roundness is derived from geometric relationships.

[0131] Morphological features comprehensively describe the geometric characteristics of exosomes, and regular and irregular morphologies can be effectively distinguished through the roundness index.

[0132] 2. Acquisition of texture features:

[0133] (a) Gray-level co-occurrence matrices are calculated in four directions: 0°, 45°, 90°, and 135°;

[0134] (b) Matrix elements are normalized to ensure that the sum of probabilities is 1;

[0135] (c) Eigenvalues are extracted from the matrix by statistical methods.

[0136] The texture feature matrix contains multi-scale and multi-directional gray-level distribution information, improving the accuracy of morphological recognition.

[0137] 3. Acquisition of statistical features:

[0138] (a) Gray-level value statistics are performed in the original image space;

[0139] (b) Fast algorithms are used for the calculation of each order moment to reduce the computational amount;

[0140] (c) Eigenvalues are normalized to eliminate the scale influence.

[0141] The statistical feature matrix describes the detailed features of the gray-level distribution through high-order moments, helping to distinguish different types of exosomes.

[0142] 4. Acquisition of abnormal region features:

[0143] (a) Edge blurriness is calculated by the Sobel operator to obtain the gradient;

[0144] (b) Overlap detection is achieved through the region growing algorithm;

[0145] (c) Integrity and size abnormality are calculated based on statistics.

[0146] Special indicators are designed for the abnormal region features, which can effectively identify and quantify various abnormal situations. Feature fusion adopts the attention mechanism, realizes the adaptive weight allocation of different features, and improves the classification accuracy.

[0147] 7. The calculation of the scale mapping equation is specifically expressed as follows:

[0148]

[0149] In the formula, M(i, s) is the multi-scale feature mapping matrix; I is the input image; s k is the k-th scale parameter; s0 is the reference scale; λ k is the scale weight coefficient, satisfying is the scale transformation operator: T k (I) = I * G(x, y, σ k ), where G(x, y, σ k ) is the Gaussian kernel of scale σ k ; R(M enh ) is the feature enhancement feedback term; β is the feedback weight; ∈1 is the error term.

[0150] 8. The calculation of the feature enhancement equation is specifically expressed as follows:

[0151]

[0152] In the formula, E(M) is the enhanced feature matrix; M is the input feature matrix; is the Laplacian operator; is the time derivative term; ω1, ω2, ω3 are the weight coefficients, satisfying α1, α2 are the smoothing parameters; O(C fb ) is the contrast feedback term; γ is the feedback coefficient; ∈2 is the error term.

[0153] 9. The calculation of the contrast optimization equation is specifically expressed as follows:

[0154]

[0155] In the formula, C(E) is the optimized contrast feature matrix; E is the input feature matrix; μ l , σ l are the local mean and standard deviation; μ g , σ g are the global mean and standard deviation; is the gradient magnitude; η1, η2, η3 are the weight coefficients, satisfying δ is a small constant to prevent division by zero; N(D fb ) is the noise suppression feedback term; θ is the feedback coefficient; ∈3 is the error term.

[0156] 10. The calculation of the noise suppression equation is specifically expressed as follows:

[0157]

[0158] Where D(C) is the final feature matrix after noise reduction; C is the input feature matrix; μ n , σ n are noise statistical parameters; is the gradient amplitude; κ is the edge preservation threshold; ρ is the detail preservation parameter; φ1, φ2, φ3 are weight coefficients, satisfying M(S fb ) is the scale mapping feedback term; ξ is the feedback coefficient; ∈4 is the error term.

[0159] The following explains the acquisition methods of each parameter:

[0160] 1. Acquisition of scale mapping parameters:

[0161] (a) The scale parameter s k is determined by logarithmic equidistant sampling: s k = s0·2 k / n , k = 1, 2,..., K;

[0162] (b) The Gaussian kernel parameter σ k is proportional to the scale parameter: σ k = c·s k , where c is the proportionality coefficient;

[0163] (c) The weight coefficient λ k is obtained by cross-validation optimization.

[0164] 2. Acquisition of feature enhancement parameters:

[0165] (a) The smoothing parameters α1, ɑ2 are estimated by the local variance of the image;

[0166] (b) The weight coefficient ω i is optimized by the gradient descent method;

[0167] (c) The feedback coefficient γ is adjusted by an iterative method.

[0168] 3. Acquisition of contrast optimization parameters:

[0169] (a) The local statistics are calculated by a sliding window;

[0170] (b) The global statistics are obtained by histogram analysis;

[0171] (c) The weight coefficient η i is fitted by the least squares method.

[0172] 4. Acquisition of noise suppression parameters:

[0173] (a) The noise statistical parameters are estimated through the background region.

[0174] (b) The edge and detail preservation parameters are analyzed through structural similarity.

[0175] (c) The weight coefficient φ i is optimized through grid search.

[0176] Principle description of the equation system construction:

[0177] 1. The scale mapping equation adopts a multi-scale decomposition and weighted fusion strategy, highlights the contribution of the main scale through quadratic term weights, and introduces a feedback mechanism to achieve dynamic adjustment of feature enhancement.

[0178] 2. The feature enhancement equation combines a non-linear mapping and a differential operator. The first term is used for feature enhancement, the second term is used for edge detection, the third term describes the feature change rate, and the contrast is adaptively adjusted through the feedback term.

[0179] 3. The contrast optimization equation adopts a strategy combining local and global normalization, preserves edge information through the gradient term, and the feedback term is used to suppress the influence of noise on the contrast.

[0180] 4. The noise suppression equation designs three functional terms: Gaussian suppression, edge preservation, and detail preservation, corresponding to different types of noise features respectively, and realizes collaborative optimization with the scale mapping through the feedback mechanism.

[0181] Specifically, the principle of the present invention is as follows: The automatic exosome recognition and analysis method of the present invention is mainly proposed to solve the problems of poor image quality, insufficient feature expression, and low classification and recognition accuracy in the prior art. Its core idea is to realize rapid detection and accurate classification of exosomes by constructing an end-to-end intelligent analysis process, including image preprocessing, feature extraction, adaptive feature fusion, and deep classification.

[0182] In terms of image preprocessing, the method adopts adaptive denoising, contrast enhancement, and illumination compensation algorithms, effectively eliminating problems such as noise interference, non-uniform brightness, and low contrast in the image, laying a good foundation for subsequent analysis. Among them, Gaussian filter is used for denoising, which can balance the denoising effect and edge preservation; histogram equalization is used for contrast enhancement, improving the discriminability of the image by stretching the gray range of the target area; illumination compensation is based on the illumination model of Fourier transform, effectively eliminating the brightness difference caused by non-uniform illumination.

[0183] In terms of feature extraction, this method integrates morphological features, texture features, statistical features, and abnormal region features, comprehensively describing the geometric size, gray-scale distribution, and internal structure of exosomes. Among them, morphological features include area, perimeter, roundness, and gray-scale value, which can quantitatively describe the geometric characteristics of exosomes; texture features are based on the gray-level co-occurrence matrix, extracting indicators such as contrast, homogeneity, entropy, and correlation, reflecting the microscopic structural information of exosomes; statistical features include gray-scale mean, standard deviation, skewness, and kurtosis, which can describe the detailed distribution characteristics of the internal components of exosomes; abnormal region features quantify and characterize problems such as blurred edges, overlapping regions, incomplete regions, and abnormal sizes, helping to identify various abnormal situations. This all-round feature extraction can describe the complex characteristics of exosomes more comprehensively and accurately compared to traditional simple morphological indicators.

[0184] In terms of feature fusion, this method adopts an attention mechanism to achieve adaptive weighted fusion of various features. Specifically, first, the extracted feature matrix is standardized to eliminate the dimension difference; then, by learning adjustable attention weights, the contributions of different features in the final fusion vector are dynamically adjusted to give full play to their respective advantages. This adaptive feature fusion strategy can enhance the discriminability of feature expression and lay a solid foundation for subsequent classification and recognition.

[0185] In terms of classification and recognition, this method constructs a deep learning model, including a feature extraction module, an attention mechanism module, and a classification module. Compared with traditional machine learning methods, the deep learning model can adaptively learn the complex pattern features of exosomes, greatly improving the accuracy of classification. At the same time, this method also designs a preliminary screening model for rapid credibility assessment of exosome regions, further enhancing the robustness of the algorithm.

[0186] In addition, during the process of feature extraction and fusion, this method also introduces adaptive mechanisms such as scale mapping, feature enhancement, contrast optimization, and noise suppression. Scale mapping highlights the main scale features through multi-scale decomposition and weighted fusion; feature enhancement combines non-linear mapping and differential operators to achieve adaptive enhancement of local structures; contrast optimization adopts a strategy combining local and global normalization to improve the contrast between exosomes and the background; noise suppression designs three functional terms: Gaussian suppression, edge preservation, and detail preservation, effectively suppressing random noise. These adaptive optimization mechanisms can fully exploit the complex features of exosomes and further enhance the robustness and analysis performance of the algorithm.

[0187] The following provides a specific embodiment 1 of the present invention. The specific implementation methods of each step in this embodiment 1 are described in detail as follows:

[0188] The specific implementation method of step S10 is as follows:

[0189] First, collect the fluorescence microscopy image I(x, y) of exosomes. To improve the accuracy and stability of subsequent processing, it is necessary to preprocess the collected original image. The preprocessing includes three sub-steps: denoising, contrast enhancement, and illumination compensation.

[0190] 1. Denoising processing: Use a Gaussian filter to denoise the image, and the filtering formula is where σ is the standard deviation of the Gaussian kernel, and by adjusting this parameter, the denoising effect and edge retention can be balanced.

[0191] 2. Contrast enhancement: Use the histogram equalization algorithm to enhance the contrast of the image. First, calculate the gray histogram h(g) of the image, where g ∈ [0, L - 1] is the number of gray levels. Then calculate the cumulative histogram and linearly map it to the range of [0, L - 1] to obtain the enhanced pixel value where N is the total number of pixels in the image.

[0192] 3. Illumination compensation: Use the method of estimating and subtracting the illumination model based on Fourier transform for illumination compensation. First, calculate the spectrum of the image through Fourier transform, then estimate the illumination distribution function L(u, v) of the image, which reflects the frequency characteristics of non-uniform illumination. Finally, divide by this function to eliminate the illumination effect and obtain the compensated image

[0193] After the above preprocessing steps, the noise of the original image is suppressed, and the contrast and brightness distribution are optimized, creating good conditions for subsequent exosome region segmentation and feature extraction.

[0194] The specific implementation manner of step S20 is as follows:

[0195] Perform adaptive threshold segmentation and morphological processing on the preprocessed image I compensated (x, y) to obtain the region of exosomes.

[0196] 1. Adaptive threshold segmentation: Use the OTSU algorithm to automatically calculate the optimal threshold T. The basic idea of the OTSU algorithm is to find the threshold that maximizes the between-class variance, and its calculation formula is where ω0, ω1, μ0, μ1 are the probabilities and means of the foreground and background respectively. Use this optimal threshold to perform binary segmentation on the image to obtain the segmentation result I binary (x, y).

[0197] 2. Morphological processing: Perform morphological processing on the binary result I binary(x, y) performs opening and closing operations. The opening operation can remove noise and small fragments, and the formula is where and represent erosion and dilation operations respectively, and B is the structuring element. The closing operation can fill small holes in the region, and the formula is Through these morphological operations, the boundary and shape of the exosome region can be further optimized.

[0198] 3. Region labeling: The segmentation result I close (x, y) is region-labeled by using the connected component analysis method. Through the 8-neighborhood connectivity judgment, adjacent exosome regions can be segmented into independent target blocks, laying a foundation for subsequent feature extraction.

[0199] After these processes, the exosome regions in the original image have been accurately extracted, providing reliable input data for subsequent in-depth analysis and recognition.

[0200] The specific implementation manner of step S30 is as follows:

[0201] The extracted exosome regions are subjected to feature extraction by using a parallel processing graphics processing unit (GPU), including three types of features: morphological features, texture features, and statistical features. At the same time, the central processing unit (CPU) is started to extract the features of the abnormal regions. These feature matrices will be combined into a region feature set to support subsequent classification and recognition.

[0202] 1. Extraction of morphological features: including area A i , perimeter P i , circularity C i and gray value G i Four indicators. The area and perimeter are obtained by pixel statistical counting, and the circularity is calculated according to the geometric relationship as The gray value uses the average value of the pixels within the region These morphological features can form a feature matrix

[0203] 2. Extraction of texture features: Based on the gray-level co-occurrence matrix P(k, l), contrast μ y )P(k, l), homogeneity entropy and correlation Four texture features are extracted. They form the texture feature matrix

[0204] 3. Extraction of statistical features: including gray mean standard deviation Skewness and kurtosis Four indicators. They form the statistical feature matrix

[0205] 4. Abnormal region feature extraction: For the possible edge-blurred regions overlapping regions incomplete regions and size-abnormal regions Four types of abnormal situations, respectively extract the corresponding quantitative indicators. They form the abnormal region feature matrix

[0206] Through GPU parallel acceleration, the feature extraction of a large number of exosome regions can be efficiently completed. These feature matrices F m , F t , F s , F a Comprehensively reflect the morphology, texture and statistical characteristics of exosomes, laying a solid foundation for subsequent classification and recognition.

[0207] The specific implementation manner of step S40 is as follows:

[0208] Input the region feature sets F m , F t , F s , F a extracted in step S30 into the pre-trained preliminary screening model to obtain the credibility score matrix of each region. Then, according to the set credibility threshold α, the exosome regions with high credibility are screened out.

[0209] 1. Preliminary screening model training: Collect a large number of exosome fluorescence microscopy image samples, perform manual annotation, extract region features, and construct a training set and a validation set. Then define the structure of the preliminary screening model, including a feature input layer, multiple convolutional layers, a fully connected layer and an output layer. Use these data to perform end-to-end supervised learning training on the model and optimize the loss function where y i is the true label, is the model output, until the model performance converges.

[0210] 2. Region credibility evaluation: Input the region feature sets extracted in step S30 into the trained preliminary screening model to obtain the credibility score matrix of each region The credibility here reflects the probability that this region is an exosome.

[0211] 3. Threshold Screening: Set a suitable confidence threshold α, and eliminate the regions with scores lower than this threshold, only retaining the regions of exosomes with high confidence. This step can effectively remove some false positive regions introduced by noise or segmentation errors, and retain the true exosome targets.

[0212] Through the training and application of the preliminary screening model, a large number of exosome regions can be quickly and preliminarily evaluated, laying a foundation for subsequent in-depth analysis and accurate classification, and improving the robustness and reliability of the entire recognition process.

[0213] The specific implementation manner of step S50 is as follows:

[0214] Input the regions of exosomes with high confidence screened in step S40 into the feature enhancement equation set for in-depth feature extraction, and finally obtain the multi-scale depth feature matrix F d ={F d1 , F d2 , …, F dK}. The feature enhancement equation set includes a scale mapping equation, a feature enhancement equation, a contrast optimization equation, and a noise suppression equation. These four equations achieve the adaptive optimization of features through cascading and feedback.

[0215] 1. Scale Mapping Equation: Among them, T k (I) = i * G(x, y, σ k ) is a scale transformation operator, implemented by Gaussian convolution; G(x, y, σ k ) is a Gaussian kernel function; λ k is a scale weight coefficient, β is a feedback weight; ∈1 is an error term. This equation highlights the contribution of the main scale through multi-scale decomposition and weighted fusion, and introduces a feedback mechanism to achieve dynamic adjustment of feature enhancement.

[0216] 2. Feature Enhancement Equation:

[0217] Among them, is the Laplace operator; is the time derivative term; ω i is a weight coefficient, α1, α2 are smoothing parameters; C fb is a contrast feedback term; γ is a feedback coefficient; ∈2 is an error term. This equation combines nonlinear mapping and differential operators to achieve the enhancement of local structural features.

[0218] 3. Contrast Optimization Equation: Among them, μ l , σ l are the local mean and standard deviation; μg , σ g are the global mean and standard deviation; is the gradient magnitude; η i is the weight coefficient, δ is a small constant to prevent division by zero; D fb is the noise suppression feedback term; θ is the feedback coefficient; ∈3 is the error term. This equation adopts a strategy combining local and global normalization, improving the contrast between exosomes and the background.

[0219] 4. Noise suppression equation: where, μ n , σ n are the noise statistical parameters; is the gradient magnitude; κ is the edge-preserving threshold; ρ is the detail-preserving parameter; φ i is the weight coefficient, S fb is the scale mapping feedback term; ξ is the feedback coefficient; ∈4 is the error term. This equation designs three functional terms: Gaussian suppression, edge preservation, and detail preservation, achieving effective suppression of different types of noise.

[0220] Through the cascaded processing of these feature enhancement equations, the multi-scale structural features of exosomes can be fully exploited, enhancing the detail information, improving the contrast, suppressing noise interference, and finally obtaining a high-quality deep feature matrix F d .

[0221] The specific implementation of step S60 is as follows:

[0222] Adopt an attention fusion mechanism to adaptively weight and fuse the morphological feature F m , texture feature F t , statistical feature F s , abnormal region feature F a extracted in step S30, and the multi-scale depth feature F d obtained in step S50 to construct a comprehensive fusion feature vector F fusion .

[0223] 1. Feature matrix normalization: Standardize various feature matrices to eliminate the differences in dimension and magnitude. The formula is where μ i , σ i are the mean and standard deviation of the feature matrix respectively.

[0224] 2. Attention mechanism calculation: Use a learnable parameter vector w i to calculate the attention weight of each type of feature In this way, the contribution of different features in the fused features can be dynamically adjusted according to their importance.

[0225] 3. Feature fusion: Multiply each type of normalized feature matrix by the corresponding attention weight, and then concatenate them to form the final fused feature vector. This fused feature comprehensively describes the multi-faceted features of exosomes, laying a foundation for subsequent accurate classification.

[0226] The attention fusion mechanism can adaptively adjust the weight of different features in the final feature vector according to their importance, improving the robustness and discriminability of feature representation, and creating good input conditions for the subsequent training of the deep classifier.

[0227] The specific implementation of step S70 is as follows:

[0228] Input the fused feature vector F obtained in step S60 fusion into the pre-trained deep classifier model to output the classification probability matrix of exosomes. Perform result screening according to the set confidence threshold β.

[0229] 1. Training of the deep classifier model: First, a large number of exosome samples of different types need to be collected to make a standard training dataset. Then, define the model structure of the deep classifier, including a feature extraction module, an attention mechanism module, and a classification module. Use the training dataset to perform end-to-end supervised learning training on the model and optimize the cross-entropy loss function. Until the model performance converges.

[0230] 2. Exosome classification prediction: Input the fused feature vector F constructed in step S60 fusion into the trained deep classifier model, and the probability matrix of each exosome region belonging to different categories can be output. Here, the categories can be subtypes of exosomes, such as exosomes, microvesicles, metabosomes, etc.

[0231] 3. Result screening: According to the preset confidence threshold β, screen the classification results. Only the exosome regions with classification probability are finally determined as valid targets. This step can further improve the accuracy and reliability of recognition.

[0232] Through the training and application of the deep learning classifier, accurate recognition of different types of exosomes can be achieved. The input of the fused feature vector and the setting of a reasonable confidence threshold β are the keys to ensuring classification performance.

[0233] To better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention:

[0234] Background introduction:

[0235] Exosomes are membrane vesicle-like nanoparticles secreted by cells and play important roles in cell-to-cell communication, immune regulation, and disease diagnosis. In recent years, human-derived exosomes have received extensive attention because they carry characteristic molecules of host cells and show great application potential in fields such as tumors, neurodegenerative diseases, and cardiovascular diseases. Accurately detecting and analyzing the characteristics of human-derived exosomes is crucial for deeply understanding their physiological functions and disease-related mechanisms.

[0236] This embodiment will adopt the automatic exosome recognition and analysis method based on microscopic imaging proposed by the present invention to rapidly detect and accurately classify human-derived exosome samples. Through technological innovations such as image preprocessing, multi-dimensional feature extraction, adaptive feature fusion, and deep classification, this method realizes the intelligent analysis of exosomes and provides strong support for biomedical research.

[0237] Sample collection and labeling:

[0238] Plasma samples from liver cancer patients, Parkinson's disease patients, and healthy subjects were collected in the experiment, and exosome samples were separated using differential centrifugation. The separated exosome samples were co-incubated and labeled with the fluorescent dyes CD9-FITC, CD81-PE, and Annexin V-APC to label exosome subpopulations derived from tumors, the nervous system, and apoptosis, respectively. After steps such as washing, centrifugation, and resuspension, finally, fluorescently labeled exosome samples were prepared for subsequent fluorescence microscopy analysis.

[0239] Image acquisition and preprocessing:

[0240] The labeled exosome samples were placed on a glass slide, and imaging was performed using a confocal fluorescence microscope (Zeiss LSM 880). During acquisition, a 63× oil immersion objective was used, the scanning area was 512×512 pixels, and the grating interval was 0.07 μm. For each sample, 20 fields of view were randomly collected, obtaining a total of 400 fluorescence microscopy images.

[0241] To improve the accuracy and stability of subsequent processing, it is necessary to preprocess the collected original images. First, a Gaussian filter is used to denoise the images, with the filter kernel size set to 7×7 and the standard deviation σ = 1.5. Then, the histogram equalization algorithm is used to enhance the contrast of the images, stretching the gray scale range to [0, 255]. Finally, based on the illumination model estimation and subtraction method using Fourier transform, illumination compensation is performed on the images to eliminate the influence of non-uniform illumination. After the above preprocessing, the noise of the original images is suppressed, and the contrast and brightness distribution are optimized, creating good conditions for subsequent exosome region segmentation and feature extraction.

[0242] Table 1 lists the key parameters of image preprocessing:

[0243] Table 1 Key Parameters of Image Preprocessing

[0244] Step Parameter Denoising processing Filter kernel size: 7×7 Standard deviation: σ = 1.5 Contrast enhancement Target gray range: [0, 255] Lighting compensation Lighting model: Based on Fourier transform

[0245] Region Segmentation and Feature Extraction:

[0246] As Figure 2 shown, for the preprocessed images, the OTSU algorithm is first used to automatically calculate the optimal threshold T * = 45, and this threshold is used to binarize the images to obtain the preliminary segmentation results of the exosome regions. Then, opening and closing operations are performed on the binarized results to remove noise points and fill small holes, further optimizing the boundaries and shapes of the exosome regions. Finally, the connected component analysis method is used to label the regions of the segmentation results, splitting adjacent exosome regions into independent target blocks, laying the foundation for subsequent feature extraction.

[0247] For the extracted exosome regions, the method of the present invention combines morphological features, texture features, statistical features, and abnormal region features for comprehensive description. Specifically as follows:

[0248] 1. Morphological Features:

[0249] Area A i : Obtained by counting pixel points, with an average of pixels 2 , and a standard deviation of σ A = 2.89×10 2 pixels 2 .

[0250] Perimeter P i : Obtained by counting boundary pixel points, with an average of pixels and a standard deviation of σ P = 39 pixels.

[0251] Roundness C i : Calculated according to geometric relationships, with an average of The standard deviation is σ C = 0.11

[0252] The gray value G i : Calculated by the average value of the pixels within the region, the average value is The standard deviation is σ G = 21

[0253] 2. Texture features:

[0254] The contrast H i1 : The average value is The standard deviation is

[0255] The homogeneity H i2 : The average value is The standard deviation is

[0256] The entropy H i3 : The average value is The standard deviation is

[0257] The correlation H i4 : The average value is The standard deviation is

[0258] 3. Statistical features:

[0259] The gray mean μ i : The average value is The standard deviation is σ μ = 21

[0260] The standard deviation σ i : The average value is The standard deviation is σ σ = 7

[0261] The skewness s i : The average value is The standard deviation is σ s = 0.18

[0262] The kurtosis k i : The average value is The standard deviation is σ k = 0.86

[0263] 4. Abnormal region features:

[0264] The edge blur degree E i : The average value is The standard deviation is σ E = 4.7

[0265] The overlap degree Oi : The average value is and the standard deviation is σ O = 0.12.

[0266] Integrity I i : The average value is and the standard deviation is σ I = 0.05.

[0267] Size abnormality S i : The average value is and the standard deviation is σ S = 0.15.

[0268] These feature matrices comprehensively describe the characteristics of exosomes such as geometric size, gray-scale distribution, and internal structure, laying a solid foundation for subsequent classification and recognition.

[0269] Preliminary screening and deep classification:

[0270] To quickly evaluate the credibility of exosome regions, the method of the present invention constructs a preliminary screening model. First, the above-mentioned extracted regional feature set is input into the model. After being processed by the feature input layer, multiple convolutional layers, fully connected layers, and output layers, a credibility scoring matrix for each region is finally obtained Then, a credibility threshold α = 0.8 is set, and regions with scores lower than this threshold are excluded, and only exosome regions with high credibility are retained.

[0271] For the exosome regions with high credibility after screening, the method of the present invention further uses deep learning technology for accurate classification and recognition. First, the model structure of the deep classifier is defined, including a feature extraction module, an attention mechanism module, and a classification module. Then, a large number of exosome samples from different sources (tumor, nervous system, and normal) are collected to make a standard training data set. The cross-entropy loss function is used to perform end-to-end supervised learning training on the model until the performance converges. Finally, the screened fusion feature vectors are input into the trained deep classifier, and a probability matrix of each exosome region belonging to different categories can be output A confidence threshold β = 0.9 is set, and only exosome regions with classification probabilities exceeding this threshold are finally determined as effective targets.

[0272] The following table lists the key parameters of the preliminary screening model and the deep classifier:

[0273] Model Parameter Preliminary screening model Confidence threshold α = 0.8 Depth classifier Confidence threshold β = 0.9 Loss function: Cross entropy

[0274] Result analysis and discussion:

[0275] Using the above method, automatic detection and classification analysis were performed on exosome samples from liver cancer patients, Parkinson's disease patients, and healthy subjects. The following table summarizes the analysis results:

[0276]

[0277] As can be seen from the results, the method of the present invention can quickly detect and accurately classify human-derived exosomes from different sources. For liver cancer patients, the detected exosomes mainly originated from tumors, accounting for as high as 73.7%; for Parkinson's disease patients, the detected exosomes mainly originated from the nervous system, accounting for 80.8%; while for healthy subjects, exosomes mainly originated from the nervous system and the apoptosis process. These analysis results are highly consistent with the existing biomedical research results, verifying the effectiveness and reliability of the method of the present invention.

[0278] As Figure 3 shown, the distribution characteristics of exosomes in different sample sources (liver cancer patients, Parkinson's disease patients, healthy subjects) are presented. In the figure, three different colored columns are used to represent exosomes from different sources (tumor-derived, apoptosis-related, nervous system-derived), and the vertical axis represents the exosome ratio (%). This figure intuitively shows the differences in exosome sources under different pathological states.

[0279] Compared with the existing exosome analysis technology based on fluorescence microscopy imaging, the method of the present invention has the following advantages:

[0280] 1. The image preprocessing is more intelligent, which can effectively eliminate problems such as noise interference, non-uniform brightness, and low contrast, laying a good foundation for subsequent analysis.

[0281] 2. The feature extraction is more comprehensive, integrating multi-dimensional features such as morphology, texture, statistics, and abnormal regions, and can more accurately describe the geometric size, microscopic structure, and internal components of exosomes.

[0282] 3. The feature fusion adopts an adaptive attention mechanism, giving full play to the advantages of various features and enhancing the discriminability of feature expression.

[0283] 4. The classification and recognition adopt deep learning technology, which can adaptively learn the complex pattern features of exosomes and greatly improve the accuracy of classification.

[0284] 5. The entire analysis process is highly automated, realizing the rapid detection and classification of a large number of exosome samples and meeting the high-throughput requirements of biomedical research.

[0285] It should be noted that the detailed explanations of the variables involved in the present invention are as follows:

[0286] 1. Related to preprocessing:

[0287] i(x, y): Pixel value of the input image; σ: Standard deviation of the Gaussian filter; L: Number of gray levels of the image; N: Total number of pixels of the image; F(u, v): Spectrum of the image; L(u, v): Illumination distribution function of the image.

[0288] 2. Related to region segmentation:

[0289] I binary (x, y): Binary segmentation result; I open (x, y): Result of opening operation; I close (x, y): Result of closing operation; T * : Optimal threshold calculated by the OTSU algorithm; ω0, ω1: Probabilities of the foreground and background; μ0, μ1: Gray means of the foreground and background; Between-class variance.

[0290] 3. Related to feature extraction:

[0291] A i : Area of the i-th region; P i : Perimeter of the i-th region; C i : Circularity of the i-th region; G i : Average gray value of the i-th region; F m : Morphological feature matrix; P(k, l): Gray-level co-occurrence matrix; H ij : The j-th texture feature of the i-th region; F t : Texture feature matrix; μ i , σ i , s i , k i : Statistical feature of the i-th region; F s : Statistical feature matrix; E i , O i , I i , S i : Feature of the abnormal region; F a : Abnormal region feature matrix.

[0292] 4. Related to depth features:

[0293] F d1 , F s2 , …, F dK : Depth features at different scales; F d : Multi-scale depth feature matrix; s k : The k-th scale parameter; s0: Reference scale; λ k : Scale weight coefficient; T k (I): Scale transformation operator; G(x, y, σ k ): Gaussian kernel function; R(Fenh ): Feature enhancement feedback term; β: Feedback weight; ∈1: Error term; F enh : Enhanced feature matrix; Laplacian operator; Time derivative term; ω i : Weight coefficient of the feature enhancement equation; α1, α2: Smoothing parameters; C fb : Contrast feedback term; γ: Feedback coefficient; ∈2: Error term of the enhancement equation; μ l , σ l : Local mean and standard deviation; μ g , σ g : Global mean and standard deviation; Gradient magnitude; η i : Weight coefficient of the contrast optimization equation; δ: Small constant to prevent division by zero; D fb : Noise suppression feedback term; θ: Feedback coefficient; ∈3: Error term of the contrast optimization equation; μ n , σ n : Noise statistical parameters; Gradient magnitude of the optimized feature; v: Edge-preserving threshold; ρ: Detail-preserving parameter; φ i : Weight coefficient of the noise suppression equation; S fb : Scale mapping feedback term; ξ: Feedback coefficient of the noise suppression equation; ∈4: Error term of the noise suppression equation.

[0294] 5. Feature fusion related:

[0295] Normalization result of the feature matrix; w i : Learnable parameter vector in the attention mechanism; α i : Attention weights of various features; F fusion : Fused feature vector.

[0296] 6. Classification and recognition related:

[0297] y i : True label; Predicted probability output by the model; Loss functions of the preliminary screening model and the deep classifier; α: Confidence threshold of the preliminary screening model; β: Confidence threshold of the deep classifier; Exosome classification probability matrix.

[0298] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. An automatic recognition and analysis method of exosomes based on microscopic imaging, characterized in that, Including: Collecting fluorescence microscopy images of exosomes and performing preprocessing to obtain preprocessed images; Subsequently, performing adaptive threshold segmentation and morphological processing to obtain the exosome region; Using a graphics processing unit parallel processing stream to extract features from the exosome region and calculating to obtain the morphological feature matrix, the texture feature matrix, and the statistical feature matrix, and combining them into a region feature set; inputting the region feature set into a pre-trained preliminary screening model for preliminary screening to obtain a credibility score matrix for the region; And according to a preset credibility threshold, screening high-credibility regions, inputting them into a feature enhancement equation set for deep feature extraction to obtain a multi-scale deep feature matrix; Using an attention fusion mechanism to perform adaptive weight fusion on the region feature set to construct a fusion feature vector; Inputting the fusion feature vector into a pre-trained deep classifier to obtain an exosome classification probability matrix; And performing result screening according to a confidence threshold to output an exosome recognition result.

2. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 1, wherein, The feature enhancement equation set includes a scale mapping equation, a feature enhancement equation, a contrast optimization equation, and a noise suppression equation.

3. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 2, wherein The scale mapping equation is used to construct feature representations at different scales. The inputs include the original image scale, target scale parameters, scale transformation coefficients, scale interpolation parameters, scale weight coefficients, and it receives the feedback features of the feature enhancement equation, and the output is a feature mapping matrix at multiple scale levels.

4. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 3, wherein The feature enhancement equation is used to enhance the local structural features of exosomes. The inputs include the feature mapping matrix, local structure tensor, enhancement weight matrix, structural similarity threshold, local contrast parameter, local variance threshold, neighborhood size parameter, and it receives the contrast feedback information of the contrast optimization equation, and the output is an enhanced feature matrix; The contrast optimization equation is used to enhance the contrast between exosomes and the background. The inputs include the enhanced feature matrix, contrast threshold parameter, brightness equalization coefficient, gray-scale stretching factor, global contrast gain, local contrast gain, edge enhancement coefficient, and it receives the noise evaluation feedback of the noise suppression equation, and the output is an optimized contrast feature matrix; The noise suppression equation is used to suppress random noise in the feature extraction process. The inputs include the optimized contrast feature matrix, noise evaluation parameters, noise threshold, smoothing coefficient, edge preservation parameter, detail preservation coefficient, noise variance estimation, and it feeds back the noise suppression result to the scale mapping equation, and the output is a denoised final feature matrix.

5. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 4, wherein, The edge-blurred region is a region where the gray-scale value gradient is less than a preset threshold, and the value range of the preset threshold is 15 to 45.

6. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 5, wherein, The overlapping region is a region where multiple regions have local overlaps, and the value range of the overlap rate threshold is 0.2 to 0.

6.

7. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 6, wherein The incomplete region is an incomplete region located at the image edge, and the value range of the edge distance threshold is 5 to 20 pixels.

8. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 7, characterized in that, The size-abnormal region is a region where the area or perimeter significantly deviates from the normal range, and the value range of the area deviation threshold is plus or minus 30%, and the value range of the perimeter deviation threshold is plus or minus 25%.

9. The automatic recognition and analysis method of exosomes based on microscopic imaging according to claim 8, wherein, The classification labels include pure exosomes, degraded exosomes, exosome aggregates, and non-exosome particles.

10. The method for automatic recognition and analysis of exosomes based on microscopic imaging according to claim 9, wherein, The preliminary screening model is used to preliminarily evaluate and screen the exosome region. The input is the region feature set, and the output is the region credibility score matrix. The structure includes a feature input layer, multiple convolutional layers, a fully connected layer, and an output layer; The steps for obtaining the training data set of the preliminary screening model include: collecting exosome fluorescence microscopy image samples, performing image annotation, extracting features, and constructing a training set and a validation set; The training steps of the preliminary screening model include: data preprocessing, model parameter initialization, iterative training, model validation, and model saving; The deep classifier is used to accurately classify exosomes. The input is the fusion feature vector, and the output is the exosome classification probability matrix. The structure includes a feature extraction module, an attention mechanism module, and a classification module; The steps for obtaining the training data set of the deep classifier include: collecting multi-type exosome samples, making a standard data set, feature extraction, data augmentation, and data set division; The training steps of the deep classifier include: data standardization, model structure definition, loss function optimization, batch training, and model evaluation.

Images