High-throughput fluorescence imaging system for multi-standard detection

Abstract

The invention relates to the technical field of fluorescence imaging detection, and discloses a high-flux fluorescence imaging system for multi-standard detection. A fluorescence image acquisition module of the system is used for capturing a fluorescence image sequence of a plurality of samples; the multi-dimensional feature extraction module constructs a dynamic feature model based on the fluorescence image sequence, extracts feature parameters of the current fluorescence image in real time, and outputs theoretical feature values through the dynamic feature model; the abnormal sample analysis module performs multi-dimensional difference analysis on the theoretical characteristic value and the actual detection characteristic value to generate a sample-level difference index; and the adaptive imaging control module configures imaging parameters according to the difference index. According to the system, high-throughput imaging, comprehensive feature extraction, accurate anomaly analysis and self-adaptive parameter adjustment of a multi-standard detection sample are achieved through the synergistic effect of all the modules, and the image quality, feature extraction integrity and system adaptability in the multi-standard detection process are effectively improved; the method is suitable for life science research, clinical diagnosis and other multi-standard detection scenes.

Description

Technical Field

[0001] This invention relates to the field of fluorescence imaging detection technology, specifically a high-throughput fluorescence imaging system for multi-mark detection. Background Technology

[0002] In life science research, clinical diagnostics, and drug development, multi-label detection technology has become an important means of obtaining rich biological information because it can simultaneously analyze multiple target components in a sample. Among them, fluorescence imaging technology, with its advantages of high sensitivity and high specificity, is widely used in multi-label detection processes. By imaging and analyzing target molecules with different fluorescent labels, qualitative and quantitative studies of multiple components in a sample can be achieved.

[0003] With the increasing demands of research, higher requirements are being placed on the throughput and accuracy of multi-target detection. Traditional fluorescence imaging systems are gradually revealing many limitations when dealing with the multi-target detection of large numbers of samples. In the image acquisition stage, most systems use fixed parameters for batch imaging, making it difficult to flexibly adjust according to the differences in fluorescence characteristics of different samples. This results in insufficient fluorescence signal acquisition for some samples, or image quality being affected by excessive background interference. At the same time, traditional systems have low image acquisition efficiency, often requiring a lot of time to process high-throughput samples, making it difficult to meet the needs of rapid detection.

[0004] In feature extraction, existing technologies mostly rely on static feature analysis, extracting only static parameters such as fluorescence intensity and distribution from a single frame of image, ignoring the dynamic changes of fluorescence signals over time. However, fluorescence signals in biological samples often have dynamic characteristics. Dynamic features such as the fluorescence decay rate and fluorescence intensity change trend of different target components contain important biological information. Static feature extraction methods cannot fully capture this key information, resulting in insufficient completeness and accuracy of feature parameters, which affects the reliability of subsequent analysis results.

[0005] Traditional methods for analyzing anomalous samples typically rely on threshold judgments based on a single feature parameter, lacking multi-dimensional difference comparisons. When complex interfering factors or interactions between multiple target components exist in a sample, anomaly detection based on a single feature is prone to misjudgment or missed detection, failing to accurately identify truly anomalous samples and thus affecting the effectiveness of the overall detection results. Furthermore, the imaging parameters of existing systems are fixed once set and cannot be adjusted in real time according to the actual detection conditions of the sample. When encountering samples with weak fluorescence signals, complex backgrounds, or anomalous features, it is difficult to improve detection performance by optimizing imaging parameters, limiting the system's adaptability to complex sample detection. Summary of the Invention

[0006] The purpose of this invention is to provide a high-throughput fluorescence imaging system for multi-mark detection, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a high-throughput fluorescence imaging system for multi-mark detection, the system comprising:

[0008] Fluorescence image acquisition module: used to capture fluorescence image sequences of multiple samples;

[0009] Multi-dimensional feature extraction module: Constructs a dynamic feature model based on the fluorescence image sequence, extracts the feature parameters of the current fluorescence image in real time, and outputs theoretical feature values ​​through the dynamic feature model;

[0010] Anomaly sample analysis module: Performs multi-dimensional difference analysis between the theoretical feature values ​​and the actual detected feature values ​​to generate sample-level difference indicators;

[0011] Adaptive imaging control module: Configures imaging parameters according to the difference index.

[0012] Preferably, the multi-dimensional feature extraction module specifically includes:

[0013] Historical data feature mining unit: performs multi-scale decomposition processing on historical fluorescence image data;

[0014] Feature optimization unit: Calculates the feature energy distribution of the output processed by the historical data feature mining unit;

[0015] Dynamic feature model construction unit: inputs the output of the feature optimization unit into the feature prediction network;

[0016] Real-time feature calculation unit: Inputs the real-time acquired fluorescence image data into the dynamic feature model generated by the dynamic feature model construction unit to obtain theoretical feature values.

[0017] Preferably, the feature optimization unit specifically includes:

[0018] Feature energy distribution calculation subunit: calculates the feature energy distribution spectrum vector of the feature vector output by the historical data feature mining unit;

[0019] Mean vector calculation subunit: calculates the positional mean vector of the set of characteristic energy distribution spectrum vectors;

[0020] Span factor calculation subunit: Calculates the span factor between the mean vector and each characteristic energy distribution spectrum vector;

[0021] Feature selection subunit: Based on the comparison between the span factor and the preset threshold, determine the set of selected feature vectors.

[0022] Preferably, the feature energy distribution calculation subunit specifically includes:

[0023] Feature distribution energy cooperative representation subunit: Calculates the feature distribution energy cooperative representation vector between each eigenvector and other eigenvectors;

[0024] Feature distribution energy synergy factor calculation subunit: calculates the feature distribution energy synergy factor vector of the feature distribution energy synergy representation vector, and generates the feature energy distribution spectrum vector.

[0025] Preferably, the abnormal sample analysis module specifically includes:

[0026] Time-domain cumulative deviation calculation unit: performs sliding comparison between theoretical and measured feature values ​​using a preset time window to generate a time-domain deviation vector;

[0027] Frequency domain energy shift detection unit: performs frequency domain decomposition on the fluorescence spectral components of theoretical and measured eigenvalues ​​to construct a frequency domain shift vector;

[0028] Structural similarity evaluation unit: Evaluates the structural similarity of images based on a matching algorithm and generates a similarity vector;

[0029] Difference index generation unit: integrates and processes the time-domain deviation vector, frequency-domain offset vector, and similarity vector to output the difference index.

[0030] Preferably, the frequency domain energy shift detection unit specifically includes extracting energy distribution characteristics within a preset sensitive frequency band interval, calculating the energy density ratio, and extracting the energy shift index.

[0031] Preferably, the structural similarity evaluation unit uses a distance algorithm for position matching.

[0032] Preferably, the abnormal sample analysis module further includes a sample location modeling unit: constructing a location relationship map based on the sample location information;

[0033] Anomaly diffusion simulation unit: maps the difference index to a positional relationship map and performs anomaly diffusion inference;

[0034] Probability distribution generation unit: Statistically counts the frequency of anomalies and generates an anomaly probability distribution map;

[0035] Physical region positioning unit: Performs cluster analysis on the anomaly probability distribution map to identify anomalous clustering areas.

[0036] Preferably, the abnormal diffusion simulation unit specifically includes calculating node attenuation factors, capturing cross-regional correlation features, and simulating abnormal diffusion paths.

[0037] Preferably, the adaptive imaging control module specifically includes performing actions such as increasing imaging resolution and enabling real-time tracking mode when the anomaly probability value exceeds a preset threshold.

[0038] The test involves applying multi-band optical stimulation, including injecting a sweep test signal, calculating the theoretical response spectrum, obtaining the measured response spectrum, calculating the offset, and marking abnormal samples.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] This high-throughput fluorescence imaging system for multi-marker detection demonstrates significant advantages in the process through the synergistic effect of its various modules. The fluorescence image acquisition module focuses on capturing fluorescence image sequences from multiple samples, overcoming the limitations of traditional single-sample or single-frame image acquisition. It can acquire continuous fluorescence image information from a large number of samples at once, providing a rich data foundation for subsequent multi-dimensional analysis. This high-throughput image acquisition method effectively addresses the detection needs of large numbers of samples, reduces the time cost of sample detection, and improves the efficiency of the overall detection process.

[0041] The multi-dimensional feature extraction module constructs a dynamic feature model based on fluorescence image sequences, overcoming the limitations of traditional static feature extraction. This module not only extracts feature parameters from the current fluorescence image in real time but also outputs theoretical feature values ​​through the dynamic feature model, achieving comprehensive capture of fluorescence signals in both temporal and spatial dimensions. The construction of the dynamic feature model considers the dynamic changes in fluorescence signals, uncovering dynamic feature information that traditional static analysis cannot obtain. This results in more comprehensive and in-depth feature parameter extraction, providing richer feature data for subsequent sample analysis.

[0042] The anomaly analysis module generates sample-level difference indicators by performing multi-dimensional difference analysis between theoretical and actual detection feature values, overcoming the limitations of traditional single-threshold judgment. Multi-dimensional difference analysis compares and evaluates samples from multiple feature dimensions, enabling more accurate identification of anomalies. This approach avoids misjudgments caused by single-feature anomalies, improves the reliability of anomaly identification, and helps to promptly identify problematic samples in multi-standard detection, reducing interference in subsequent analyses.

[0043] The adaptive imaging control module configures imaging parameters based on differences in indicators, enabling dynamic adjustment of the imaging process. Traditional systems with fixed imaging parameters struggle to adapt to the varying fluorescence characteristics of different samples, while this module can specifically optimize imaging parameters based on the actual detection conditions of the sample. For samples with weak fluorescence signals, signal acquisition can be enhanced by adjusting parameters such as exposure time and excitation intensity; for samples with significant background interference, filtering parameters can be optimized to reduce interference. This adaptive parameter configuration significantly improves image quality for different samples, ensuring that the fluorescence characteristics of various samples are clearly captured. This enhances the adaptability and flexibility of the entire multi-standard detection process, allowing the system to maintain stable detection performance even when faced with complex and diverse samples. Attached Figure Description

[0044] Figure 1 This is a timing diagram of the high-throughput fluorescence imaging system for multi-mark detection described in this invention;

[0045] Figure 2 A flowchart for the multi-dimensional feature extraction module;

[0046] Figure 3 The flowchart for the feature optimization unit;

[0047] Figure 4 This is a flowchart of the abnormal sample analysis module. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 This invention provides a high-throughput fluorescence imaging system for multi-mark detection, the system comprising:

[0050] This high-throughput fluorescence imaging system is used for multi-label detection of multiple samples. The system operation begins with the fluorescence image acquisition module. This module is equipped with a high-sensitivity camera and a multi-channel filter wheel, which can rapidly and continuously capture fluorescence images of multiple samples placed on the stage according to a preset time sequence or trigger signal, generating a fluorescence image sequence containing time-series information.

[0051] The captured fluorescence image sequences are input into a multi-dimensional feature extraction module. The core function of this module is to construct a feature model that describes the normal dynamic changes of the sample based on historically accumulated fluorescence image sequence data. During real-time operation, this module receives the currently acquired fluorescence image data, performs real-time calculations using the constructed dynamic feature model, and outputs the theoretical feature values ​​corresponding to the current image. These theoretical feature values ​​represent the characteristic state that the sample should exhibit at the current moment under the model's prediction.

[0052] The anomaly analysis module receives theoretical feature values ​​from the multi-dimensional feature extraction module and actual detected feature values ​​(measured feature values) of the current image acquired by the fluorescence image acquisition module. This module performs multi-dimensional and multi-angle difference analysis on the theoretical and measured feature values. This analysis goes beyond a simple comparison of a single indicator; it comprehensively calculates the degree of deviation from the theoretical and measured feature values ​​across multiple feature dimensions (such as time-domain variation, frequency-domain energy, and structural morphology). Based on these deviation calculations, the module generates a quantified sample-level difference index, which comprehensively reflects the degree of anomaly of the current sample relative to the model prediction.

[0053] The adaptive imaging control module receives differential indicators from the abnormal sample analysis module. Based on the specific values ​​or levels of these differential indicators, this module dynamically adjusts the imaging parameters of the fluorescence image acquisition module. Adjustment strategies may include changing exposure time, gain, excitation light intensity, filter switching strategy, scan resolution, scan speed, or imaging area. The aim is to enable more detailed and in-depth imaging observation of suspected abnormal samples, or to optimize imaging efficiency for normal samples.

[0054] Example 1: See Figure 2 This demonstrates that the implementation of the multi-dimensional feature extraction module involves multiple collaborative units that work together to complete the entire process from historical data learning to real-time feature prediction. The historical data feature mining unit first processes the historical fluorescence image dataset stored in the system. This unit reads the historical image sequence corresponding to the stage coordinates and timestamps, performs data cleaning, and removes invalid frames caused by equipment failure or environmental interference. The cleaned and valid historical images are then sent to the multi-scale decomposition processing stage. This processing employs a combined spatial and frequency domain analysis method, simultaneously extracting features at different scales for each image. In the spatial domain, local texture analysis based on a sliding window is used, with multiple gradient levels set for the window size to capture sample structural information from microscopic to macroscopic levels. In the frequency domain, the image undergoes multi-level sub-band decomposition to separate different frequency components, focusing on mid-to-high frequency information related to the fluorescence characteristics of biomarkers. The decomposition results at each scale generate corresponding feature maps, which together constitute the initial feature vector set describing the multi-scale characteristics of the sample.

[0055] The feature optimization unit receives a set of multi-scale feature vectors from the historical data feature mining unit. The core task of this unit is to optimize and filter these raw features, removing redundant information and retaining the most representative feature components. Internally, the unit performs a feature energy distribution calculation process. This process analyzes the energy contribution of each dimension element within each feature vector and calculates its energy distribution characteristics. Specifically, each feature vector is normalized to eliminate dimensional differences. Then, the square of each dimension's value is calculated to obtain the energy distribution spectrum. Further analysis of the central tendency and dispersion of this energy spectrum identifies the principal component regions where energy is concentrated. For multiple feature vectors of each sample, the unit also analyzes the correlation of energy distribution between different feature vectors and calculates a feature distribution energy collaborative representation vector. This vector quantifies the correlation strength of different feature dimensions in energy distribution. Based on these collaborative representation vectors, a feature distribution energy collaborative factor vector is calculated. This factor vector comprehensively reflects the collaborative effect and importance weight of each feature dimension in the overall energy distribution. Finally, the historical feature data of each sample is transformed into a set of optimized feature energy distribution spectrum vectors, which more centrally represent the key characteristics of the sample.

[0056] The dynamic feature model building unit uses the feature energy distribution spectrum vector output by the feature optimization unit as training data input. This unit constructs and trains a feature prediction network model. Before model training, the input data undergoes time series alignment and segmentation to ensure temporal continuity. The feature prediction network employs a recurrent neural network structure including memory units. The input layer receives the feature energy distribution spectrum vectors of the current and historical moments, the hidden layer contains multiple processing units to capture the temporal dependencies of dynamic feature changes, and the output layer predicts the theoretical feature energy distribution spectrum vector for the next moment. During training, the network parameters are iteratively adjusted using an optimization algorithm, with the objective function being to minimize the difference between the predicted output and the actual historical data. The trained network possesses the ability to infer future feature changes based on historical feature sequences, forming a dynamic feature model. This model encapsulates the multi-scale feature evolution patterns of samples under normal conditions.

[0057] The real-time feature calculation unit operates during system online operation. This unit directly interfaces with the fluorescence image acquisition module, receiving newly captured fluorescence image data in real time. For each newly input real-time image, the unit first performs the same multi-scale decomposition processing procedure as the historical data feature mining unit, generating a real-time multi-scale feature vector for the image. Subsequently, this vector is fed into the feature optimization unit to perform the same optimization process, including normalization, energy distribution calculation, and co-factor calculation, ultimately generating a real-time feature energy distribution spectrum vector. This real-time spectrum vector, together with historical spectrum vectors from several previous time points, constitutes a short time-series sequence. This sequence is input into the feature prediction network trained by the dynamic feature model construction unit. Based on the current and historical inputs, the network performs forward computation, outputting a predicted value of the theoretical feature energy distribution spectrum vector for the next time step (i.e., the current time step). This predicted value is the theoretical feature value, representing the theoretical feature state that the sample should exhibit in the current historical context. This theoretical feature value is output to the anomaly sample analysis module as a benchmark reference value for subsequent difference analysis. The entire real-time calculation process is completed within strict time constraints, ensuring the timeliness of the system response. This unit also includes a caching mechanism to temporarily store the most recent feature sequences to support the network’s temporal inference needs.

[0058] Example 2: See Figure 3 The implementation of the feature optimization unit involves the collaborative operation of multiple sub-units, performing in-depth processing and filtering on the original feature vector set output by the historical data feature mining unit. The feature energy distribution calculation sub-unit first receives the feature vector set from the historical data feature mining unit, which contains multi-scale feature vectors from multiple samples at different time points. The core task of this sub-unit is to calculate the feature energy distribution spectrum vector for each feature vector. The processing begins with a global analysis of all feature vectors within the set. For each feature vector in the set, this sub-unit calculates its feature distribution energy collaborative representation vector with every other feature vector in the set. This calculation focuses on analyzing the relative energy distribution patterns between feature vectors, evaluating the similarity, complementarity, or repulsion of the energy distribution of a specific feature vector with other vectors in the feature space. This analysis does not rely on a single distance metric but examines the overall distribution relationship in the multi-dimensional feature space. Based on the calculated feature distribution energy collaborative representation vector, this sub-unit further calculates the feature distribution energy collaborative factor vector. The calculation of this factor vector involves statistical analysis of the collaborative representation vector, extracting quantitative indicators that characterize the consistency and uniqueness of the energy distribution of the feature vector in the overall set. Finally, the sub-unit outputs the feature energy distribution spectrum vector corresponding to each feature vector. This spectrum vector is a comprehensive representation of the energy distribution characteristics of the feature vector and its cooperative relationship in the set.

[0059] The mean vector calculation subunit receives all feature energy distribution spectral vectors output by the feature energy distribution calculation subunit. Its task is to calculate the statistical central tendency of this set of spectral vectors. Specifically, it aggregates the feature energy distribution spectral vectors of all samples. This subunit aligns the vectors according to their dimensional positions. For each dimensional index position of the spectral vector, it calculates the average value of all samples at that position. For example, if the spectral vector has N dimensions, the calculation process iterates through dimensions 1 to N, collecting the values ​​of all samples in each dimension and calculating their arithmetic mean. After iterating through all dimensions, a mean vector with the same dimensions as a single feature energy distribution spectral vector is generated. This mean vector represents the overall average pattern or baseline pattern of the feature energy distribution of all samples in the historical data.

[0060] The span factor calculation subunit operates on the feature energy distribution spectral vector of each sample. This subunit receives two inputs: the feature energy distribution spectral vector of a specific sample and the global mean vector generated by the mean vector calculation subunit. The core function of this subunit is to calculate the degree of difference between the sample's spectral vector and the global mean vector, quantifying this difference as the span factor. The calculation process involves comparing the numerical differences between the two vectors in corresponding dimensions. This subunit employs a comprehensive difference measurement method that considers not only the absolute differences in each dimension but also the pattern and distribution of differences between dimensions. The result is either a scalar value or a low-dimensional vector, collectively referred to as the span factor. The magnitude of the span factor directly reflects the degree to which the feature energy distribution pattern of the sample deviates from the overall average distribution pattern. A larger span factor indicates that the sample's feature distribution has higher uniqueness or anomaly, while a smaller span factor indicates that its distribution pattern is close to the overall average.

[0061] The feature selection subunit performs feature filtering based on the output of the span factor calculation subunit. This subunit receives the span factor values ​​of all samples. Internally, it sets a preset threshold, which can be determined based on historical data analysis or system configuration, to distinguish between normal distributions and significant deviations. The subunit compares the span factor of each sample with this preset threshold. The comparison result is used to decide whether to retain the sample's feature vector. Specifically, if a sample's span factor is greater than the preset threshold, it is determined that the sample's feature energy distribution pattern has sufficient uniqueness or potential information value, and its corresponding feature vector is selected to be retained in the subsequent processing set. Conversely, if a sample's span factor is less than or equal to the preset threshold, it is determined that the sample's feature energy distribution pattern is highly consistent with the overall average pattern, and its information may be redundant or insufficiently representative; therefore, its corresponding feature vector is marked as a removable item and does not enter the subsequent process. After traversing, comparing, and filtering all samples, the subunit outputs a new set of feature vectors. This set is a subset of the original set, containing only the feature vectors of samples determined to have significant feature distribution characteristics (i.e., span factors exceeding the threshold). This filtered set of feature vectors is passed to the dynamic feature model building unit as the foundational data for training the dynamic feature model. The filtering process aims to focus on feature data that better represents sample diversity or potential anomaly patterns, thereby optimizing the data quality for model training.

[0062] Example 3: See Figure 4This demonstrates the implementation of the anomaly sample analysis module, which involves multiple analysis units working collaboratively to perform multi-dimensional difference analysis on theoretical and measured feature values. The time-domain cumulative deviation calculation unit is responsible for capturing the difference patterns in feature values ​​over time. This unit sets a fixed-length time window that slides in steps across the time series. For each window slide, the time period covered includes multiple consecutive time points. At each time point, this unit acquires the theoretical feature values ​​output by the multi-dimensional feature extraction module and the corresponding measured feature values ​​provided by the fluorescence image acquisition module. This unit performs a comparison operation, calculating the difference between the two in each feature dimension. The difference calculation can be an absolute difference, a relative difference, or other defined difference measurement functions. For each time point within the window, a difference value vector is calculated, where each element corresponds to the difference in a feature dimension. Subsequently, this unit performs cumulative processing on the difference value vectors of all time points within the window. The cumulative processing is not a simple summation and averaging, but considers the temporal order and trends. For example, it weights and accumulates the difference values ​​across the time series within the window, giving higher weight to recent differences; or it calculates the standard deviation of the difference values ​​to reflect the degree of fluctuation. Finally, each window slide outputs a time-domain bias vector, which comprehensively reflects the deviation and change pattern of the measured feature values ​​relative to the theoretical predictions over a specific time period. This vector contains multiple components, each corresponding to the cumulative bias calculation result of a feature dimension.

[0063] The frequency-domain energy shift detection unit focuses on analyzing the differences in fluorescence spectral characteristics in the frequency domain. This unit receives fluorescence spectral data contained within theoretical and measured characteristic values. This spectral data is typically represented as a sequence of intensity variations with wavelength or frequency. The unit performs frequency-domain transformation operations, such as Fast Fourier Transform, on both the theoretical and measured spectral sequences, converting the spectral data from the wavelength domain to the frequency domain. In the frequency domain, spectral information is represented as the energy distribution across different frequency components. The unit presets one or more sensitive frequency bands, typically determined based on known excitation / emission characteristics or historical data analysis of the target fluorescent marker in the sample, encompassing the frequency range most sensitive to changes in sample state. Within these sensitive frequency bands, the unit extracts the energy distribution characteristics of the theoretical and measured spectra. This includes calculating the energy value at each frequency point, as well as the total energy or average energy density across the entire band. Based on these energy distribution characteristics, the unit calculates the ratio of the energy density of the theoretical spectrum to that of the measured spectrum within the sensitive frequency band. This ratio calculation can be for the total energy across the entire band or for specific sub-bands or key frequency points within the band. The energy density ratio directly reflects the relative energy changes of the measured spectrum relative to the theoretical spectrum at key frequency components. To more accurately quantify the degree of shift, this unit further calculates an energy shift index. This index is calculated using the following formula:

[0064]

[0065] Where: ζ represents the calculated energy shift index, Ω represents the set of all discrete frequency points within the preset sensitive frequency band interval, and E m (k) represents the energy value of the measured spectrum at frequency point k, E t w(k) represents the energy value of the theoretical spectrum processed at frequency point k, and w(k) represents the weighting coefficient of frequency point k (which can be set according to the importance of the frequency or historical statistics). This formula calculates the weighted average of the absolute energy difference between the measured spectrum and the theoretical spectrum at all frequency points within the sensitive frequency band, with the weights determined by w(k). Finally, the unit outputs a frequency domain offset vector containing information such as the energy offset index.

[0066] The structural similarity evaluation unit is responsible for analyzing the differences in spatial structure between fluorescence images. This unit compares and matches the image structural information implied or reconstructed by theoretical feature values ​​with the measured fluorescence images. The matching process employs a matching algorithm. This algorithm first extracts key structural feature points or feature region descriptors from the theoretical feature information, and simultaneously extracts corresponding feature points or descriptors from the measured images. Feature point extraction can employ methods such as corner detection, blob detection, or deep learning-based feature extractors. Subsequently, the unit performs feature point matching operations to find the correspondence between theoretical feature points and feature points in the measured images. The matching process needs to consider the similarity of feature descriptors and spatial geometric constraints. After establishing the correspondence between feature points, the unit uses a distance algorithm to evaluate the accuracy of position matching. For each successfully matched pair of feature points, its actual positional distance in image space is calculated. This distance can be Euclidean distance or other defined distance metrics. The unit collects the distance values ​​of all successfully matched point pairs. Based on these distance values, the unit calculates a comprehensive similarity metric. This calculation can be a statistical processing of all distance values, such as calculating the average distance, median distance, or the proportion of matched point pairs with a distance less than a certain threshold. A larger distance value indicates a greater deviation in the location of the matching points and a lower similarity in the image structure; conversely, a smaller distance value indicates a higher similarity. Ultimately, this unit generates a similarity vector containing one or more quantized values ​​that reflect the overall degree of matching or difference between the theoretically predicted image structure and the measured image structure in terms of spatial morphology.

[0067] The difference index generation unit is a crucial step in integrating multi-dimensional analysis results. This unit receives time-domain deviation vectors from the time-domain cumulative deviation calculation unit, frequency-domain offset vectors (including energy offset indices, etc.) from the frequency-domain energy offset detection unit, and similarity vectors from the structural similarity assessment unit. These vectors represent the quantitative results of differences across different analytical dimensions (time domain, frequency domain, spatial structure). This unit performs integration processing, fusing these vectors to generate a single sample-level difference index. The integration processing strategy can be multi-step. First, each input vector may be normalized to eliminate differences in units and numerical ranges between different dimensions. Second, weighting coefficients are assigned to each vector or key components within a vector based on the importance of each analytical dimension for anomaly detection. These weighting coefficients can be determined based on expert experience or historical data analysis. Then, weighted summation, weighted averaging, or other fusion functions are used to combine the normalized and weighted data from each dimension. The design goal of the fusion function is to ensure that the generated difference index comprehensively and evenly reflects the overall deviation of the sample from theoretical predictions across multiple analytical dimensions. Ultimately, the unit outputs a scalar value or a low-dimensional vector as a sample-level difference index. This index is the primary basis for the subsequent adaptive imaging control module's decision-making; a higher value generally indicates a greater likelihood of the sample being abnormal.

[0068] Example 4: This demonstrates how the spatial analysis function of the abnormal sample analysis module is implemented through the collaboration of multiple units. The sample position modeling unit first acquires the physical position information of the samples on the stage. This information typically comes from the layout configuration file or the stage coordinate system at the time of sample loading. For example, in a typical layout of a 96-well plate, the samples are arranged in an 8-row (AH) 12-column (1-12) matrix. This unit constructs a positional relationship map based on this coordinate data. In the map, each sample is considered a node, and the node attributes include its row and column coordinates. The connection relationships between nodes are defined based on a preset adjacency rule: the samples with the closest physical distance are automatically connected; for example, each sample node establishes bidirectional connection edges with its adjacent sample nodes in the four directions above, below, left, and right. For samples at edge positions, the number of connections may be less than four. This map is stored in memory as a graph data structure, with nodes storing coordinates and unique identifiers, and edges storing connection relationships.

[0069] The anomaly diffusion simulation unit receives sample-level difference indices output by the difference index generation unit. This unit assigns these difference index values ​​to the corresponding sample nodes in the location relationship map. See Table 1 for a sample of the difference index data obtained in one imaging analysis cycle.

[0070] Table 1: Some data of the difference indicators obtained in a single imaging analysis cycle are as follows.

[0071]

[0072]

[0073] This unit performs anomaly propagation simulation on the graph. The simulation process models the effect of anomaly states propagating from nodes with high dissimilarity indices to their connected nodes. First, a node attenuation factor is calculated. This factor simulates the intensity attenuation of anomaly states as they propagate between nodes. The attenuation factor is set based on the physical distance between nodes or the properties of the connecting edges. For example, the attenuation factor between directly adjacent nodes is set to 0.7, meaning that the intensity of the anomaly state is attenuated to 70% of the original node when it propagates to its direct neighbor. For nodes that are not directly adjacent but connected by short paths, the attenuation factor is further reduced based on the path length and the number of intermediate nodes. Simultaneously, this unit captures cross-regional association features. This involves analyzing the impact of long-distance connections or highly connected nodes in the graph on anomaly propagation. For example, a node located in a central position and connecting multiple rows and columns may become a hub for anomaly propagation, even if its initial dissimilarity index is not high. Based on the attenuation factor and cross-regional association features, this unit simulates the anomaly propagation path. The simulation starts from a node with a high dissimilarity index and calculates the potential impact strength of its anomaly state on each of its neighboring nodes. The impact strength is equal to the dissimilarity index value of the source node multiplied by the corresponding attenuation factor. If the calculated impact strength exceeds a preset propagation threshold, the neighboring node is marked as being affected by the anomalous diffusion, and its impact strength is recorded. This process iterates: newly marked nodes (whose impact strength exceeds the threshold) become new source nodes, continuing to propagate the anomalous state to their unmarked or less affected neighbors. The simulation continues until no new nodes are marked or their impact strength falls below the threshold. After the simulation, some nodes in the graph are marked as being affected by the anomalous diffusion of the initial high-difference samples.

[0074] The probability distribution generation unit performs statistical analysis on anomalous states. This unit doesn't rely solely on the results of a single analysis loop, but rather records the number of times each sample node is marked as anomalous across multiple consecutive imaging analysis loops. It counts the number of times each node exhibits an anomalous state across the total number of loops. Based on this count, the probability of that node being anomalous is calculated: Anomalous probability = Number of times that node is anomalous / Total number of imaging analysis loops. For example, assuming 10 loops were performed and sample B3 was marked as anomalous in 8 loops, its anomalous probability is 0.8. This unit iterates through all sample nodes, calculating their respective anomalous probabilities, and generating an anomalous probability distribution map covering the entire stage area. This map is a two-dimensional probability matrix, where each element corresponds to the anomalous probability value at a sample location.

[0075] The physical region localization unit receives an anomaly probability distribution map. This unit performs a clustering analysis algorithm on the probability distribution map to identify high-probability regions that are spatially clustered. The clustering algorithm employs a density-based spatial clustering method. This method scans the probability distribution map, searching for sample points whose anomaly probability values ​​exceed a set threshold, and examines the spatial proximity between these high-probability points. If a group of high-probability points is spatially close to each other, and the number of points in this group exceeds a minimum cluster size threshold, these points are identified as an anomalous cluster. The algorithm outputs all identified anomalous clusters and the range of sample locations they contain. For example, the analysis might identify one cluster containing locations B2, B3, C2, C3, and another cluster containing locations F7, F8, G7. These clusters are clearly marked on the spatial distribution map, and their boundary coordinates are recorded. The information on the identified anomalous clusters is output to guide subsequent targeted reviews or experimental operations.

[0076] Example 5: The implementation of the adaptive imaging control module focuses on dynamically adjusting the imaging strategy based on anomaly analysis results. This module continuously receives output data from the anomaly sample analysis module, including sample-level difference indices and anomaly probability distribution maps. Internally, the module sets a preset anomaly probability threshold, which serves as the decision boundary for triggering enhanced imaging operations. When the system detects that the anomaly probability value of a sample exceeds the preset threshold, the module determines that the sample has a high anomaly risk and immediately initiates the enhanced imaging strategy for that sample.

[0077] The enhanced imaging strategy comprises two core operations. The first is increasing imaging resolution. This is achieved by controlling the hardware parameters of the fluorescence image acquisition module. Specific instructions might include switching to a higher numerical aperture objective to increase optical magnification, or commanding the scanning mechanism to reduce the scan step size to increase pixel sampling density. These adjustments aim to acquire finer spatial structural information of the target sample, such as subcellular localization details of organelles or morphological features of tiny fluorescent spots. The second operation is enabling a real-time tracking mode. This mode marks the target sample as a high-priority monitoring object. In subsequent imaging cycles, the system significantly increases the imaging frequency of that sample. For example, instead of a full scan of the entire stage, the system inserts multiple additional rapid local scans of that specific sample. The real-time tracking mode may also include trigger condition settings, such as immediately triggering an imaging session when the sample's real-time characteristic value undergoes a sudden change, or setting a minimum time interval for continuous imaging to ensure intensive capture of the sample's dynamic changes. The combined application of these two operations enables the system to conduct deeper and more timely spatial and temporal observations of high-risk samples.

[0078] Furthermore, this module implements a strategy for applying multi-band optical stimulation tests. This strategy is specifically designed for samples marked as having a high probability of anomalous change. The module controls the light source system to generate a swept-frequency test signal. A swept-frequency test signal refers to a continuous or discrete stepwise variation of the wavelength of the excitation light within a preset spectral range. The rate, step size, and range of wavelength variation are pre-configured based on the sample type and marker characteristics. Simultaneously with applying the swept-frequency test signal to the target sample, the module calculates the expected fluorescence emission spectrum, i.e., the theoretical response spectrum, based on the sample's internally stored theoretical model or known fluorescence response characteristics, under the stimulation of that specific swept-frequency signal. The theoretical response spectrum predicts the fluorescence intensity distribution that should be observed at different excitation wavelengths.

[0079] Simultaneously, the fluorescence image acquisition module works concurrently to capture the actual fluorescence response signal generated by the target sample under the excitation of the swept-frequency test signal. This module acquires fluorescence images or spectral data of the corresponding band through the appropriate emission filter channel, based on the currently used excitation wavelength. The acquired raw data is processed, such as through spectral extraction or intensity integration in a specific band, to finally generate the measured response spectrum. The measured response spectrum reflects the true fluorescence response of the sample to swept-frequency excitation under actual test conditions.

[0080] The adaptive imaging control module then calculates the offset between the theoretical and measured response spectra. The offset calculation involves a comparative analysis of the two spectral sequences. Calculation methods may include: calculating the intensity difference between the two spectra at key characteristic wavelengths; calculating the correlation coefficient or similarity of the overall shapes of the two spectra; or calculating the area difference between the two spectral curves within a preset wavelength range. This offset is a quantifiable value characterizing the degree of deviation of the measured response from the theoretical prediction.

[0081] Significant offsets are used as an important criterion for further confirmation of anomalies. When the calculated offset exceeds a preset offset threshold, the module explicitly marks the sample as an anomalous sample. This multi-band optical stimulation test based on excitation-response characteristics provides characteristic dimensional information independent of conventional imaging. It can reveal anomalous functional responses of samples under specific optical stimuli, such as fluorophore quenching, changes in energy transfer efficiency, or alterations in environmental sensitivity, information that may not be fully obtained through static or single-band imaging. This marking information is then transmitted to other modules in the system or recorded in the results report.

[0082] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0083] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A high-throughput fluorescence imaging system for multi-mark detection, characterized in that, include: Fluorescence image acquisition module: used to capture fluorescence image sequences of multiple samples; Multi-dimensional feature extraction module: Constructs a dynamic feature model based on the fluorescence image sequence, extracts the feature parameters of the current fluorescence image in real time, and outputs theoretical feature values ​​through the dynamic feature model; Anomaly sample analysis module: Performs multi-dimensional difference analysis between the theoretical feature values ​​and the actual detected feature values ​​to generate sample-level difference indicators; Adaptive imaging control module: Configures imaging parameters according to the difference index.

2. The high-throughput fluorescence imaging system for multi-mark detection according to claim 1, characterized in that, The multi-dimensional feature extraction module specifically includes: Historical data feature mining unit: performs multi-scale decomposition processing on historical fluorescence image data; Feature optimization unit: Calculates the feature energy distribution of the output processed by the historical data feature mining unit; Dynamic feature model construction unit: inputs the output of the feature optimization unit into the feature prediction network; Real-time feature calculation unit: Inputs the real-time acquired fluorescence image data into the dynamic feature model generated by the dynamic feature model construction unit to obtain theoretical feature values.

3. A high-throughput fluorescence imaging system for multi-mark detection according to claim 2, characterized in that, The feature optimization unit specifically includes: Feature energy distribution calculation subunit: calculates the feature energy distribution spectrum vector of the feature vector output by the historical data feature mining unit; Mean vector calculation subunit: calculates the positional mean vector of the set of characteristic energy distribution spectrum vectors; Span factor calculation subunit: Calculates the span factor between the mean vector and each characteristic energy distribution spectrum vector; Feature selection subunit: Based on the comparison between the span factor and the preset threshold, determine the set of selected feature vectors.

4. A high-throughput fluorescence imaging system for multi-mark detection according to claim 3, characterized in that, The characteristic energy distribution calculation subunit specifically includes: Feature distribution energy cooperative representation subunit: Calculates the feature distribution energy cooperative representation vector between each eigenvector and other eigenvectors; Feature distribution energy synergy factor calculation subunit: calculates the feature distribution energy synergy factor vector of the feature distribution energy synergy representation vector, and generates the feature energy distribution spectrum vector.

5. A high-throughput fluorescence imaging system for multi-mark detection according to claim 1, characterized in that, The abnormal sample analysis module specifically includes: Time-domain cumulative deviation calculation unit: performs sliding comparison between theoretical and measured feature values ​​using a preset time window to generate a time-domain deviation vector; Frequency domain energy shift detection unit: performs frequency domain decomposition on the fluorescence spectral components of theoretical and measured eigenvalues ​​to construct a frequency domain shift vector; Structural similarity evaluation unit: Evaluates the structural similarity of images based on a matching algorithm and generates a similarity vector; Difference index generation unit: integrates and processes the time-domain deviation vector, frequency-domain offset vector, and similarity vector to output the difference index.

6. A high-throughput fluorescence imaging system for multi-mark detection according to claim 5, characterized in that, The frequency domain energy shift detection unit specifically includes extracting energy distribution characteristics within a preset sensitive frequency band interval, calculating the energy density ratio, and extracting the energy shift index.

7. A high-throughput fluorescence imaging system for multi-mark detection according to claim 5, characterized in that, The structural similarity evaluation unit uses a distance algorithm for position matching.

8. A high-throughput fluorescence imaging system for multi-mark detection according to claim 1, characterized in that, The abnormal sample analysis module also includes a sample location modeling unit: constructing a location relationship map based on the sample location information; Anomaly diffusion simulation unit: maps the difference index to a positional relationship map and performs anomaly diffusion inference; Probability distribution generation unit: Statistically counts the frequency of anomalies and generates an anomaly probability distribution map; Physical region positioning unit: Performs cluster analysis on the anomaly probability distribution map to identify anomalous clustering areas.

9. A high-throughput fluorescence imaging system for multi-mark detection according to claim 8, characterized in that, The abnormal diffusion simulation unit specifically includes calculating node attenuation factors, capturing cross-regional correlation features, and simulating abnormal diffusion paths.

10. A high-throughput fluorescence imaging system for multi-mark detection according to claim 1, characterized in that, The adaptive imaging control module specifically includes performing actions such as increasing imaging resolution and enabling real-time tracking mode when the anomaly probability value exceeds a preset threshold. The test involves applying multi-band optical stimulation, including injecting a sweep test signal, calculating the theoretical response spectrum, obtaining the measured response spectrum, calculating the offset, and marking abnormal samples.

Images