Method for identifying Mycobacterium tuberculosis using fluorescence microscopy images of sputum samples

Abstract

The present invention discloses a method for identifying Mycobacterium tuberculosis in fluorescence microscopic images of sputum samples, comprising: performing multi-focal plane scanning on the sputum sample using an optical microscope to obtain a continuous fluorescence slice image sequence; processing the fluorescence slice image sequence using a three-dimensional volume reconstruction algorithm to generate three-dimensional fluorescence volume data containing spatial coordinates and fluorescence intensity information; converting the bacterial structure in the three-dimensional fluorescence volume data into a graph model, wherein nodes correspond to bacterial feature points and carry spatial coordinates and fluorescence parameters, and edges connect adjacent feature points and reflect spatial relationships; and performing feature analysis on the graph model using a graph neural network. In the present invention, high recognition accuracy is achieved, and the use of three-dimensional imaging and graph neural network technology can comprehensively obtain multi-dimensional information of the bacteria, construct a three-dimensional model and convert it into a graph model for feature analysis, effectively solving the problem of two-dimensional viewing angle limitation and greatly improving the accuracy of Mycobacterium tuberculosis identification.

Description

Technical Field

[0001] The present invention belongs to the field of biometric identification technology, and in particular relates to a method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples. Background Art

[0002] Identifying Mycobacterium tuberculosis in sputum samples is a critical step in microbial testing. Existing technologies often rely on traditional two-dimensional image recognition methods, acquiring sputum sample images through single-focal-plane fluorescence microscopy and then processing and identifying them using traditional convolutional neural networks. This method is relatively simple to operate and can be used as a preliminary screening tool in settings with limited equipment.

[0003] However, practical applications still face pressing challenges. On the one hand, traditional two-dimensional images cannot fully capture the three-dimensional spatial structure of bacteria, making it difficult to accurately identify overlapping or morphologically complex bacteria, resulting in low recognition accuracy and a high risk of missed detections. On the other hand, differences in device parameters (such as excitation wavelength and imaging resolution) lead to poor data compatibility and difficulty in standardizing results. However, the present invention, through innovative technical means, achieves significant improvements in improving recognition accuracy and enhancing device compatibility. Therefore, a method for identifying Mycobacterium tuberculosis using fluorescence microscopy images of sputum samples is proposed. Summary of the Invention

[0004] To overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for identifying Mycobacterium tuberculosis in sputum sample fluorescence microscopy images, which solves the problem in the prior art that two-dimensional recognition is difficult to distinguish complex bacteria and equipment differences affect detection uniformity.

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

[0006] The method for identifying Mycobacterium tuberculosis by fluorescence microscopy of sputum samples includes:

[0007] S10: Sputum samples were scanned at multiple focal planes using an optical microscope to obtain a sequence of continuous fluorescent slice images;

[0008] S11: Processing the fluorescence slice image sequence using a three-dimensional volume reconstruction algorithm to generate three-dimensional fluorescence volume data including spatial coordinates and fluorescence intensity information;

[0009] S12: Convert the bacterial structure in the 3D fluorescence volume data into a graph model, where nodes correspond to bacterial feature points and carry spatial coordinates and fluorescence parameters, and edges connect adjacent feature points and reflect spatial relationships;

[0010] S13: Perform feature analysis on the graph model using a graph neural network, and output an identification result of Mycobacterium tuberculosis.

[0011] Preferably, in step S10, the multifocal plane scanning adopts a light sheet fluorescence microscope or a confocal microscope, and the axial spacing of the acquired fluorescence slice images meets the resolution requirement of three-dimensional structure reconstruction, and the axial spacing between adjacent slices is 50-500 nm.

[0012] Preferably, before step S10, the sputum sample is fluorescently stained before scanning, and multi-wavelength fluorescent labeling technology is used to differentially label different biological structures, and the difference in fluorescence wavelengths for labeling different structures is greater than 30 nm to avoid signal crosstalk.

[0013] Preferably, in step S11, the three-dimensional volume reconstruction algorithm includes a synchronous iterative reconstruction technique, a filtered back projection algorithm or a deep learning reconstruction method, and performs denoising and resolution enhancement on the fluorescent slice image before reconstruction.

[0014] Preferably, in step S12, the node features of the graph model include at least one of coordinates, fluorescence intensity, and local curvature, and the edge weight is determined by the spatial distance between nodes, the difference in fluorescence intensity, or the morphological similarity.

[0015] Preferably, in step S13, the graph neural network includes a graph convolutional network, a graph attention network or a graph autoencoder, which aggregates node neighborhood features through a message passing mechanism and outputs classification results.

[0016] Preferably, the method further includes performing dynamic analysis on the three-dimensional fluorescence volume data of the continuous time series:

[0017] The trajectory prediction algorithm is used to predict bacterial movement. The trajectory prediction algorithm includes the following steps:

[0018] S21: Define the bacterial state vector as a six-dimensional vector of three-dimensional coordinates and velocity ,in, is the three-dimensional coordinate of the kth frame (unit: μm), is the velocity in each direction (unit: μm / s);

[0019] S22: The bacterial structures in adjacent time frames are associated through a cross-frame matching algorithm to establish the spatiotemporal continuity of the bacterial movement trajectory. The state of the previous frame is recursively extrapolated through the state transfer matrix F to obtain the predicted value of the current frame state: , the state transfer matrix F is: ,in is the time interval between adjacent frames;

[0020] S23: Based on microscopic observation data (observation coordinates) and the observation matrix H are used to modify the predicted state to obtain the estimated value of the ring frame state: , the observation matrix , used to extract the position component in the state vector;

[0021] Extract kinematic parameters and combine them with a time series model to distinguish between dead bacteria and live bacteria, wherein the kinematic parameters include at least one of trajectory curvature, displacement speed, and direction change rate.

[0022] Preferably, the trajectory prediction algorithm includes Kalman filtering, particle filtering or a deep learning prediction model, and the cross-frame matching algorithm includes the Hungarian algorithm, the minimum cost flow algorithm or an association method based on feature matching.

[0023] Preferably, after the graph neural network analysis, a visual map reflecting the contribution of bacterial feature points to the recognition results is generated to assist in verifying the accuracy of the recognition results.

[0024] Preferably, a database containing microscope equipment parameters is established, and a transfer learning algorithm is used to adapt image data collected by different devices. The equipment parameters include at least one of the excitation light wavelength, imaging resolution, and scanning interval.

[0025] The technical effects and advantages of the method for identifying Mycobacterium tuberculosis by fluorescence microscopy of sputum samples of the present invention are as follows:

[0026] 1. This invention has high recognition accuracy and adopts three-dimensional imaging and graph neural network technology. It can comprehensively obtain multi-dimensional information of bacteria, construct a three-dimensional model and convert it into a graph model for feature analysis, effectively solving the limitation of two-dimensional viewing angle, greatly improving the recognition accuracy of Mycobacterium tuberculosis and reducing missed detections.

[0027] 2. This invention has a wide range of applicable scenarios. Through the combination of different technical means, it can meet the needs of various scenarios such as screening in high-incidence areas of tuberculosis, clinical diagnosis in hospitals, and research in scientific research institutions. It can effectively deal with complex background samples, detection of tiny low-concentration bacteria, and research on bacterial movement characteristics.

[0028] 3. This invention has strong dynamic analysis capabilities and can dynamically analyze three-dimensional fluorescence volume data of continuous time series. It uses trajectory prediction algorithms and kinematic parameters to predict bacterial movement and distinguish between dead and live bacteria, providing an important basis for research related to Mycobacterium tuberculosis and drug development.

[0029] 4. This invention has good equipment compatibility. It establishes an equipment parameter database and applies a transfer learning algorithm to effectively solve the data incompatibility problem caused by differences in microscope equipment between different medical institutions, improve the consistency and comparability of test results, and promote data sharing and collaboration.

[0030] 5. This invention provides intuitive auxiliary verification. After graph neural network analysis, it generates a visual map that intuitively displays the contribution of bacterial feature points to the recognition results, making it easier for scientific research and clinical personnel to assist in verifying the accuracy of the recognition results and improving diagnostic reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The figure is a flow chart of the method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples proposed by the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0033] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus. In the absence of further restrictions, the elements defined by the sentence "include..." do not exclude the presence of other identical elements in the process, method, article or apparatus that includes the elements.

[0034] Example 1

[0035] refer to Figure 1 This embodiment provides a method for identifying Mycobacterium tuberculosis using fluorescence microscopy images of sputum samples, which is used for tuberculosis screening. The specific implementation includes:

[0036] Implementation scenario: This system is suitable for preliminary screening in areas with a high incidence of tuberculosis. In large-scale sample testing, it can quickly and accurately identify Mycobacterium tuberculosis, providing an important basis for subsequent diagnosis.

[0037] Implementation steps:

[0038] For sample preparation and multi-focal plane scanning, clinical sputum samples were selected and multi-wavelength fluorescence labeling was used. Mycobacterium tuberculosis was labeled with a 488nm fluorescence wavelength, while background biological structures were labeled with a 520nm fluorescence wavelength, with a 32nm difference between the two to minimize signal crosstalk. Multi-focal plane scanning was performed using a confocal microscope, with an axial spacing of 200nm between adjacent slices to meet the resolution requirements for 3D structural reconstruction. Sequences of continuous fluorescence slice images were acquired. The microscope was calibrated before scanning to ensure the accuracy of imaging parameters. During dynamic analysis, a cross-frame matching algorithm (such as the Hungarian algorithm) was used to correlate the same bacteria in adjacent frames to ensure the spatiotemporal continuity of the motion trajectory.

[0039] For 3D volume reconstruction, the acquired fluorescence slice images are first subjected to median filtering for denoising, followed by resolution enhancement using bicubic interpolation. Simultaneous iterative reconstruction technology (SIRT) is then used to reconstruct the processed image sequence into 3D volume data, generating 3D fluorescence volume data containing spatial coordinates and fluorescence intensity information. During the reconstruction process, the number of SIRT algorithm iterations is adjusted based on image quality, typically set to 10-20.

[0040] A graph model is constructed to transform the bacterial structure in 3D fluorescence volume data into a graph model. Nodes are selected as bacterial feature points, each carrying 3D spatial coordinates, fluorescence intensity, and local curvature information. Edges connect adjacent feature points, with edge weights determined by the spatial distance between nodes. A threshold segmentation algorithm is used to extract bacterial regions and identify feature points.

[0041] Graph neural network feature analysis uses a graph convolutional network (GCN) to analyze the graph model. It aggregates node neighborhood features through a message passing mechanism and ultimately outputs the recognition result for Mycobacterium tuberculosis. The GCN model was trained using a dataset containing 1,000 labeled samples, a learning rate of 0.001, and 50 epochs.

[0042] Implementation effect:

[0043] Testing on 500 sputum samples in a high-TB incidence area revealed that the proposed method achieved a 92% accuracy rate for Mycobacterium tuberculosis recognition, with an average recognition time of 15 seconds. Compared with traditional 2D image recognition methods, 3D volumetric data more comprehensively reflects bacterial structure, improving recognition accuracy by 15%, and effectively reducing missed detections due to the limited 2D viewing angle. Furthermore, median filtering and bicubic interpolation effectively removed image noise, enhancing the clarity of the reconstructed image and ensuring accurate recognition.

[0044] Example 2

[0045] This embodiment provides a method for identifying Mycobacterium tuberculosis using fluorescence microscopy images of sputum samples, which is used for the implementation of complex sample detection. The specific implementation content includes:

[0046] Application scenarios: Commonly used in hospital clinical diagnosis, especially for the detection of complex background samples. It can accurately distinguish Mycobacterium tuberculosis from other impurities and host cells, and assist doctors in formulating treatment plans.

[0047] Implementation steps:

[0048] Fluorescent staining is performed on sputum samples using three fluorescence wavelengths, labeling Mycobacterium tuberculosis (450nm), host cells (500nm), and other impurities (550nm). The wavelength difference between the three wavelengths is greater than 30nm. During the staining process, the dye concentration and staining time are strictly controlled to ensure uniform and stable staining.

[0049] Multifocal plane scanning was performed using a light-sheet fluorescence microscope, with an axial spacing of 300 nm between adjacent slices, to acquire a sequence of continuous fluorescence slice images. Light-sheet microscopy reduces photobleaching and phototoxicity, making it suitable for long-term imaging. Light intensity should be adjusted during scanning to avoid excessive light exposure that could affect sample activity.

[0050] 3D volume reconstruction: After image denoising (mean filtering) and resolution enhancement (bilinear interpolation), 3D volume reconstruction was performed using a deep learning reconstruction method (based on a U-Net network model). The U-Net model was trained using 200 sets of 3D image data, and the mean squared error (MSE) was used as the loss function.

[0051] Build a graph model and perform feature analysis. Node features include coordinates and fluorescence intensity, and edge weights are determined by the difference in fluorescence intensity between nodes. Use a graph autoencoder (GAE) to analyze the graph model and output recognition results. During GAE model training, set the hidden layer dimension to 64, use Adam as the optimizer, and use a learning rate of 0.005.

[0052] Implementation effect:

[0053] Testing on 150 clinical sputum samples from a hospital revealed that this method clearly distinguished structures with different fluorescent markers, reduced signal crosstalk by 80%, and increased recognition accuracy to 93%. It demonstrated particularly robustness for bacterial identification against complex backgrounds, providing strong support for subsequent accurate diagnosis. The U-Net-based reconstruction method effectively restored image details, while the GAE model accurately extracted bacterial features, achieving high-precision recognition.

[0054] Example 3

[0055] This embodiment provides a method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples, which is used for the implementation of microbial detection. The specific implementation content includes:

[0056] Implementation scenario: This method is suitable for basic research on Mycobacterium tuberculosis in scientific research institutions, especially for the detection of tiny bacterial structures and low-concentration bacteria, which helps to deeply explore the characteristics of bacteria.

[0057] Implementation steps:

[0058] Sample processing and scanning: Sputum samples were fluorescently stained, with the labeled Mycobacterium tuberculosis at 500 nm and the background at 540 nm (40 nm differential). High-resolution image sequences were acquired using a confocal microscope with an axial spacing of 50 nm. To improve the quality of imaging of microscopic bacteria, the exposure time was appropriately extended.

[0059] For 3D volume reconstruction, the image is first subjected to non-local mean filtering for denoising, followed by resolution enhancement using a super-resolution convolutional neural network. Finally, 3D reconstruction is performed using the filtered back projection (FBP) algorithm. The super-resolution network is trained using low-resolution images and their corresponding high-resolution images as training data pairs, and a perceptual loss function is used to optimize the network parameters.

[0060] The graph model is constructed with node features including coordinates and local curvature, and edge weights are determined by morphological similarity. Morphological features are extracted using morphological algorithms to assist in determining edge weights.

[0061] Graph neural network analysis uses a graph attention network (GAT) to perform feature analysis and output recognition results. When training the GAT model, the number of attention heads is set to 4, the learning rate is set to 0.002, and the number of training rounds is set to 80.

[0062] Implementation effect:

[0063] In a test of 50 sputum samples containing small amounts of Mycobacterium tuberculosis selected by a research institute, the reconstructed 3D images using this embodiment achieved higher resolution, clearer details, and enhanced recognition of microbial structures, reaching 90% accuracy—a 10% improvement over traditional reconstruction algorithms. This allows for earlier detection of low-concentration bacteria. The combination of a super-resolution convolutional neural network and the FBP algorithm significantly improved image quality, while the GAT model effectively focused on microbial features, enabling accurate identification.

[0064] Example 4

[0065] This embodiment provides a method for identifying Mycobacterium tuberculosis in sputum sample fluorescence microscopy images for use with anti-tuberculosis drugs. The specific implementation includes:

[0066] Implementation scenario: Used to study the motility characteristics and survival status of Mycobacterium tuberculosis, providing a theoretical basis for the development of new anti-tuberculosis drugs, such as analyzing the inhibitory effect of drugs on bacterial motility.

[0067] Implementation steps:

[0068] Dynamic analysis preparation: Based on Example 1, dynamic analysis of continuous time series three-dimensional fluorescence volume data is added. The bacterial state vector is defined as a six-dimensional vector of three-dimensional coordinates (x_k, y_k, z_k) and velocity (v_x, v_y, v_z).

[0069] Trajectory prediction uses a Kalman filter. The state transition matrix F is calculated based on the time interval between adjacent frames (set to 0.1s), and the position component of the observation matrix H is extracted. The previous frame's state is recursively extrapolated using the state transition matrix. The predicted state is modified based on the microscope observation data to obtain an estimated state for the current frame. During the calculation process, the process noise covariance and observation noise covariance of the Kalman filter are adjusted based on actual conditions.

[0070] Kinematic parameter extraction and differentiation: Extract kinematic parameters such as trajectory curvature, displacement velocity, and directional change rate, and use them in conjunction with a time series model to distinguish between live and dead bacteria. Kinematic parameters are calculated using a sliding window algorithm, and the time series model is trained using a long short-term memory (LSTM) network.

[0071] Implementation effect:

[0072] Fifty bacteria were selected for trajectory analysis and prediction, achieving an average prediction error of 0.5 μm / s and an accuracy rate of 85%. Kinematic parameters were used to effectively distinguish between live and dead bacteria, with an accuracy rate of 88%, providing important insights into the survival and movement characteristics of Mycobacterium tuberculosis. The Kalman filter accurately predicted bacterial trajectories, while the LSTM model effectively distinguished between live and dead bacteria based on kinematic parameters.

[0073] Example 5

[0074] This embodiment provides a method for identifying Mycobacterium tuberculosis using fluorescence microscopy images of sputum samples, which is used to implement consistency in detection results. The specific implementation includes:

[0075] Implementation scenario:

[0076] It is suitable for data sharing and collaboration among different medical institutions, solving data incompatibility problems caused by equipment differences and achieving consistency and comparability of test results.

[0077] Implementation steps:

[0078] Device parameter adaptation: A database containing microscope parameters (excitation wavelength, imaging resolution, and scanning interval) is established, and a transfer learning algorithm is used to adapt image data collected by different devices. For example, two different confocal microscope models: Device A has an excitation wavelength of 488 nm, an imaging resolution of 0.2 μm, and a scanning interval of 100 nm; Device B has an excitation wavelength of 514 nm, an imaging resolution of 0.3 μm, and a scanning interval of 200 nm. During transfer learning, the large amount of annotated data collected by Device A serves as the source domain data, while the data from Device B serves as the target domain data, allowing fine-tuning of the pre-trained model parameters.

[0079] For sample processing and recognition, two devices were used to collect fluorescent slice images of sputum samples. After transfer learning adaptation, 3D volume reconstruction, graph model construction, and graph neural network analysis were performed according to the steps in Example 1 to output recognition results. During the adaptation process, the effects of different transfer learning methods (such as those based on adversarial training and feature alignment) were compared to select the optimal solution.

[0080] Implementation effect:

[0081] 200 sputum samples were collected using devices A and B, respectively. After transfer learning adaptation, the recognition accuracy of data collected by devices A and B reached 91% and 90%, respectively, significantly improving compared to 85% and 83%, respectively, without adaptation. This embodiment effectively addresses the issue of data discrepancies between different devices, improving the versatility and adaptability of the method. The transfer learning algorithm successfully eliminates the impact of device differences, enabling accurate recognition of data collected by different devices.

[0082] Comparative Example 1

[0083] This comparative example provides commercially available traditional PLA coated paper, and the specific implementation contents include:

[0084] Application scenarios: Traditional two-dimensional image recognition methods are often used in some primary medical institutions with limited detection equipment and technology as a preliminary and simple screening method, but they have certain limitations.

[0085] Implementation steps: Use traditional two-dimensional image recognition methods to perform single focal plane fluorescence microscopy imaging of sputum samples, and directly process and recognize the two-dimensional images. The steps are as follows:

[0086] The sputum samples were fluorescently stained (same as in Example 1), and single focal plane fluorescence images were acquired using a confocal microscope.

[0087] After denoising and enhancing the two-dimensional images, a traditional convolutional neural network was used for recognition and classification. The convolutional neural network consists of three convolutional layers, two pooling layers, and one fully connected layer. The training used a cross-entropy loss function and a learning rate of 0.01.

[0088] Implementation effect:

[0089] In a test of the same 500 samples, the 2D image recognition method achieved an accuracy rate of only 77%, with a high missed detection rate (15%). Because 2D images cannot provide information about the three-dimensional spatial structure of bacteria, they struggle to accurately identify overlapping or complex morphologies, resulting in significantly lower recognition performance than the 3D processing method of the present invention. Furthermore, traditional convolutional neural networks are susceptible to noise and interference when processing complex background images, resulting in reduced recognition accuracy.

[0090] Experimental Example 1

[0091] This experiment is used for clinical sample screening experiments.

[0092] Purpose of the experiment:

[0093] Compare the recognition effects of Examples 1-5 and the comparative example in large-scale clinical sample screening.

[0094] Experimental methods:

[0095] A total of 1,500 sputum samples were collected from three hospitals and randomly divided into six groups. The methods described in Examples 1-5 and the comparative example were used to identify Mycobacterium tuberculosis. The recognition accuracy, missed detection rate, and average recognition time were recorded for each group. During the experiment, the concentration distribution and sample type of each group were similar to ensure comparability of the experimental results.

[0096] Experimental data: The comparison of the recognition parameters of Examples 1-5 and Comparative Example 1 is shown in Table 1.

[0097]

[0098] Experimental results: Examples 1-5 all achieved higher recognition accuracy than the control examples, with lower missed detection rates. While some examples experienced longer average recognition times, these were acceptable and significantly improved detection accuracy, making them more suitable for clinical sample screening. Example 2's superiority in handling complex background samples enabled it to perform exceptionally well in clinical sample screening. Through device adaptation, Example 5 demonstrated stable performance across data collected by different devices.

[0099] Experimental Example 2

[0100] This experiment provides detection experiments of samples with different concentrations.

[0101] Purpose of the experiment:

[0102] The detection capabilities of Examples 1-5 and the comparative example in samples of Mycobacterium tuberculosis at different concentrations were investigated.

[0103] Experimental methods:

[0104] Five groups of sputum samples of different concentrations (high concentration, medium-high concentration, medium concentration, medium-low concentration, and low concentration) were prepared, with 30 samples in each group. They were tested using the methods of Examples 1-5 and the comparative example, and the recognition accuracy was recorded. During sample preparation, the sample concentration was controlled by gradient dilution and calibration was performed using standard strains to ensure the accuracy of the sample concentration.

[0105] Experimental data: Effect of Mycobacterium tuberculosis concentration on detection experiments in Examples 1-5 and Comparative Example 1.

[0106]

[0107] Experimental results: Across samples of varying concentrations, Examples 1-5 achieved significantly higher recognition accuracy than the control examples, with a particularly pronounced advantage in low-concentration samples. This demonstrates the method's enhanced detection capabilities and stability. Example 3, leveraging high-resolution reconstruction and a graph attention network, demonstrated excellent performance in low-concentration sample detection. Example 4, through dynamic analysis, provided additional insight into viable bacteria detection in low-concentration samples.

[0108] Comparison of Examples 1-5 with Comparative Example 1 from the perspectives of application scenarios, implementation steps, and implementation effects clearly demonstrates the advantages of the method of the present invention. Specific details are as follows:

[0109] Application Scenarios: Examples 1-5 each focus on a specific application scenario. Example 1 is suitable for preliminary screening in high-incidence areas of tuberculosis; Example 2 is commonly used in hospital clinical diagnosis, especially for samples with complex backgrounds; Example 3 is suitable for research institutions to detect microbial structures and low-concentration bacteria; Example 4 focuses on studying the motility and survival of Mycobacterium tuberculosis; and Example 5 is used to resolve data incompatibility issues caused by equipment differences between different medical institutions. In contrast, traditional two-dimensional image recognition methods are only suitable for preliminary and simple screening in primary care institutions with limited detection equipment and technology.

[0110] Implementation steps: Examples 1-5 are all based on three-dimensional imaging and graph neural network technology, and realize the identification of Mycobacterium tuberculosis through multi-focal plane scanning, three-dimensional volume reconstruction, graph model construction, graph neural network analysis and other steps. Some embodiments also involve dynamic analysis or equipment parameter adaptation, and each step adopts different technology and algorithm combinations; the control example adopts the traditional two-dimensional image recognition method, and only processes and recognizes the two-dimensional images of single focal plane fluorescence microscopy imaging. The steps are simple and the technical means are traditional.

[0111] Implementation Results: In terms of recognition accuracy, Examples 1-5 were significantly higher than the control example. In experiments with samples of varying concentrations, they showed a clear advantage, particularly in detecting low-concentration samples. In terms of missed detection rates, Examples 1-5 were lower than the control example. While the control example took a shorter average recognition time, Examples 1-5 significantly improved detection accuracy while maintaining an acceptable average recognition time. Furthermore, the examples demonstrated unique advantages in image denoising, detail recovery, complex background processing, and device adaptation. However, due to the limitations of two-dimensional images, the control example struggled to accurately identify overlapping or complex bacterial forms and was susceptible to noise interference.

[0112] The above embodiments may be implemented in whole or in part through software, hardware, firmware or any other combination. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0113] Those skilled in the art will appreciate that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0114] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0115] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited to this. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0116] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples, characterized in that: include: S10: Sputum samples were scanned in multiple focal planes using an optical microscope to obtain a series of continuous fluorescent slice images; S11: Processing the fluorescence slice image sequence using a three-dimensional volume reconstruction algorithm to generate three-dimensional fluorescence volume data including spatial coordinates and fluorescence intensity information; S12: Convert the bacterial structure in the 3D fluorescence volume data into a graph model, where nodes correspond to bacterial feature points and carry spatial coordinates and fluorescence parameters, and edges connect adjacent feature points and reflect spatial relationships; S13: Perform feature analysis on the graph model using a graph neural network, and output an identification result of Mycobacterium tuberculosis.

2. The method for identifying Mycobacterium tuberculosis by fluorescence microscopy of sputum samples according to claim 1, characterized in that: In step S10, the multi-focal plane scanning adopts a light sheet fluorescence microscope or a confocal microscope, and the axial spacing of the acquired fluorescence slice images meets the resolution requirement of three-dimensional structure reconstruction, and the axial spacing between adjacent slices is 50-500 nm.

3. The method for identifying Mycobacterium tuberculosis by fluorescence microscopy of sputum samples according to claim 1, characterized in that: Prior to step S10, the sputum sample is fluorescently stained before scanning, and multi-wavelength fluorescent labeling technology is used to differentially label different biological structures. The difference in fluorescence wavelengths for labeling different structures is greater than 30 nm to avoid signal crosstalk.

4. The method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples according to claim 1, wherein: In step S11, the three-dimensional volume reconstruction algorithm includes a synchronous iterative reconstruction technique, a filtered back projection algorithm or a deep learning reconstruction method, and performs denoising and resolution enhancement on the fluorescent slice image before reconstruction.

5. The method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples according to claim 1, wherein: In step S12, the node features of the graph model include at least one of coordinates, fluorescence intensity, and local curvature, and the edge weight is determined by the spatial distance between nodes, the difference in fluorescence intensity, or the morphological similarity.

6. The method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples according to claim 1, wherein: In step S13, the graph neural network includes a graph convolutional network, a graph attention network or a graph autoencoder, which aggregates node neighborhood features through a message passing mechanism and outputs classification results.

7. The method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples according to claim 1, wherein: It also includes dynamic analysis of continuous time series of three-dimensional fluorescence volume data: The trajectory prediction algorithm is used to predict bacterial movement. The trajectory prediction algorithm includes the following steps: S21: Define the bacterial state vector as a six-dimensional vector of three-dimensional coordinates and velocity ,in, is the three-dimensional coordinate of the kth frame, and its unit is μm, is the velocity in each direction, and its unit is μm / s; S22: The bacterial structures in adjacent time frames are associated through a cross-frame matching algorithm to establish the spatiotemporal continuity of the bacterial movement trajectory. The state of the previous frame is recursively extrapolated through the state transfer matrix F to obtain the predicted value of the current frame state: , the state transfer matrix F is: ,in is the time interval between adjacent frames; S23: Based on microscopic observation data The predicted state is modified by the observation matrix H to obtain the estimated value of the ring frame state: , the observation matrix , used to extract the position component in the state vector; Extract kinematic parameters and combine them with a time series model to distinguish between dead bacteria and live bacteria, wherein the kinematic parameters include at least one of trajectory curvature, displacement speed, and direction change rate.

8. The method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples according to claim 7, wherein: In step S22, the trajectory prediction algorithm includes Kalman filtering, particle filtering or a deep learning prediction model, and the cross-frame matching algorithm includes the Hungarian algorithm, the minimum cost flow algorithm or an association method based on feature matching.

9. The method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples according to claim 1, wherein: After the graph neural network analysis, a visual map reflecting the contribution of bacterial feature points to the recognition results is generated to assist in verifying the accuracy of the recognition results.

10. The method for identifying Mycobacterium tuberculosis using fluorescence microscopic images of sputum samples according to claim 1, wherein: A database containing microscope equipment parameters is established, and a transfer learning algorithm is used to adapt image data collected by different devices. The equipment parameters include at least one of the excitation light wavelength, imaging resolution, and scanning interval.

Images