Flow type immune cell intelligent analysis system based on machine learning

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a flow type immune cell intelligent analysis system based on machine learning, comprising: a data preprocessing module for cleaning and feature optimization of a multi-dimensional fluorescence signal; the hybrid model marks suspected rare cell clusters through unsupervised clustering, and target cells are confirmed through supervised classification; the training optimization module ensures that the model reaches the standard and is stable; the reasoning decision module outputs a result with confidence and triggers early warning; and the iteration upgrading module realizes automatic model adaptation. According to the system, the rare cell recognition accuracy is greatly improved, the result reliability is guaranteed, the clinical immune cell analysis requirement can be met, and the practical value is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a flow cytometry intelligent analysis system based on machine learning. BACKGROUND

[0002] Flow cytometry is the core technology of immune cell analysis, which can realize the rapid sorting and characterization of cell populations through multi-dimensional fluorescence signals, and is widely used in disease diagnosis, immune monitoring and other fields. The existing flow cytometry analysis technology has obvious limitations: first, rare immune cells (usually less than 1%) are easily covered by mainstream cell signals, and traditional manual gates selection method relies on experience and is difficult to accurately identify; second, the data preprocessing is mostly simple filtering, without targeted optimization for fluorescence signal outliers, adherent cells and other interference, resulting in insufficient feature quality; third, the single machine learning model (such as only using unsupervised clustering) has poor generalization ability, and lacks continuous training optimization mechanism, and has insufficient adaptability when facing different batches of samples. These defects result in low recognition accuracy and poor result stability of the existing system for rare cells, which cannot meet the demand of clinical precise analysis of trace immune cells, and an intelligent analysis scheme integrating data optimization, mixed model, dynamic training and iterative upgrade is urgently needed. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application proposes a flow cytometry intelligent analysis system based on machine learning, aiming to solve at least one of the problems in the above background technology.

[0004] The present application provides a flow cytometry intelligent analysis system based on machine learning, comprising: A data preprocessing module configured to receive multi-dimensional fluorescence signal data output by a flow cytometer, and to clean and feature engineering process the multi-dimensional fluorescence signal data; A mixed machine learning model module comprising an unsupervised clustering unit and a supervised classification unit, the unsupervised clustering unit being configured to receive standardized feature data, analyze the abundance of each cell population and the relative difference of the mainstream population, determine which population belongs to the potential rare cell population, and mark it as a suspected rare cell cluster; the supervised classification unit is configured to receive the suspected rare cell cluster, determine whether the suspected population is the target rare cell, and output the target rare cell clustering result; A training optimization module configured to receive clustering results and external annotation data sources, determine whether the model meets the preset performance benchmark based on the comprehensive recognition performance of the model for rare cells, and determine whether the model is stable through the performance fluctuation amplitude of continuous multiple rounds of training, and perform multiple rounds of training and parameter adjustment on the model when it does not meet the standard or is unstable; The reasoning and decision module is configured to determine an output strategy according to the confidence distribution of the clustering result and the matching degree of the population abundance and the physiological and pathological conventional range, and output an analysis result with a confidence score. The iterative upgrading module is configured to evaluate the adaptability of the cumulative data evaluation model to new samples, determine whether the cumulative data volume meets the iteration requirement, and trigger model iteration optimization when the adaptability is less than a preset threshold and the data volume meets the requirement.

[0005] In some embodiments, the data preprocessing module is configured to receive multi-dimensional fluorescence signal data output by a flow cytometer, and when performing cleaning and feature engineering processing on the multi-dimensional fluorescence signal data, includes: obtaining a scatter plot of forward scattering light signals and side scattering light signals in the multi-dimensional fluorescence signal data, and presetting a first threshold value, when the signal intensity of a particle in the scatter plot is greater than or equal to the first threshold value, the particle is retained; when the signal intensity of the particle in the scatter plot is less than the first threshold value, the particle is removed; calculating the ratio of FSC width to FSC height of each cell event, and presetting a second threshold value, when the ratio of FSC width to FSC height is less than or equal to the second threshold value, the event is retained; when the ratio of FSC width to FSC height is greater than the second threshold value, the event is determined to be a adherent cell and is removed; performing Z-score standardization processing on the signal values of each fluorescence channel, and presetting a third threshold value range, when the Z-score of any fluorescence channel of a certain cell event after standardization is greater than or less than the third threshold value range, the cell event is determined to be an outlier, and the outlier is replaced by a nearest neighbor interpolation method; when the Z-score of any fluorescence channel of a certain cell event after standardization is equal to the third threshold value range, the cell event is determined to be a normal value.

[0006] In some embodiments, the data preprocessing module is configured to receive multi-dimensional fluorescence signal data output by a flow cytometer, and when performing cleaning and feature engineering processing on the multi-dimensional fluorescence signal data, further includes: obtaining a plurality of fluorescence features related to cell events detected by a flow cytometer, and calculating the mutual information value between each of the fluorescence features and a known rare cell population label, wherein the known rare cell population label is the identification of a pre-labeled rare immune cell population; when the mutual information value of the fluorescence feature is less than a preset fourth threshold value, it is determined that the fluorescence feature has no contribution to the rare immune cell population clustering, and it is removed from the feature set; when the mutual information value of the fluorescence feature is greater than or equal to the fourth threshold value, it is determined that the fluorescence feature contributes to the grouping of the rare immune cell population, and the fluorescence feature is retained from the feature set; The retained fluorescence feature is logarithmically converted and normalized.

[0007] In some embodiments, the unsupervised clustering unit is configured to receive standardized feature data, determine which groups belong to potential rare cell groups by analyzing the relative difference in abundance of each cell group from the mainstream group, and mark as suspected rare cell clusters, including: The preprocessed multi-dimensional fluorescence signal data is preliminarily grouped using a density peak clustering algorithm, and the clustering parameters of the density peak clustering algorithm are set, including: The clustering effect is determined by the silhouette coefficient, and when the silhouette coefficient is less than a preset fifth threshold value, the cutoff distance of the density peak clustering algorithm is adjusted; When the silhouette coefficient is greater than or equal to the fifth threshold value, the cutoff distance of the density peak clustering algorithm is not adjusted; The proportion of the number of cell events in each cluster obtained by clustering is counted, and when the proportion of the number of cell events of a certain cluster is less than a preset sixth threshold value, the cluster is marked as the suspected rare cell cluster; When the proportion of the number of cell events of a certain cluster is greater than or equal to the sixth threshold value, it is determined that the cluster is a mainstream cell cluster, and the grouping result of the mainstream cell cluster is output.

[0008] In some embodiments, the supervised classification unit is configured to receive the suspected rare cell cluster, determine whether the suspected group is a target rare cell, and output the target rare cell grouping result, including: The suspected rare cell cluster is identified using a lightweight convolutional neural network model; The number of cell events in the suspected rare cell cluster is obtained, and when the number of cell events is less than a preset number threshold, it is determined that the sample size is insufficient, and synthetic samples are generated by the SMOTE algorithm, and the feature values of the synthetic samples need to fall within the 95% confidence interval of the original sample features; When the number of cell events is greater than or equal to the number threshold, the suspected rare cell cluster is used as a training set for training of the lightweight convolutional neural network model; The rare cell recall rate and precision of the model are calculated on the validation set, and when the rare cell recall rate is less than a preset seventh threshold value, the number of hidden layers of the lightweight convolutional neural network is increased, and the learning rate is adjusted; When the rare cell recall rate is greater than or equal to the seventh threshold value, and the precision is less than a preset eighth threshold value, a Dropout layer is added to an output layer of the lightweight convolutional neural network; Otherwise, no adjustment is made to the output layer of the lightweight convolutional neural network.

[0009] In some embodiments, the training optimization module is configured to receive the clustering results and an external annotation data source, determine whether the model meets a preset performance benchmark based on the comprehensive identification performance of the model on the rare cells, and determine whether the model is stable by the performance fluctuation amplitude of continuous multiple rounds of training. When the model does not meet the benchmark or is not stable, the model is trained and the parameters are adjusted for multiple rounds, including: The mixed machine learning model module is trained using the training set, and the rare cell F1-score is calculated on the validation set; When the rare cell F1-score is greater than or equal to a preset ninth threshold value, it is determined that the mixed machine learning model module meets the benchmark, and the test set is verified; When the rare cell F1-score is less than the ninth threshold value, iterative optimization is entered.

[0010] In some embodiments, the training optimization module is configured to receive the clustering results and an external annotation data source, determine whether the model meets a preset performance benchmark based on the comprehensive identification performance of the model on the rare cells, and determine whether the model is stable by the performance fluctuation amplitude of continuous multiple rounds of training. When the model does not meet the benchmark or is not stable, the model is trained and the parameters are adjusted for multiple rounds, including: The features of the rare cell identification error samples in the validation set are analyzed, and the model is retrained after adding new features when the error samples are concentrated in the fluorescence signal overlap area; When the error samples are randomly distributed, the batch size of the lightweight convolutional neural network model is adjusted; The above optimization training is repeatedly performed until the rare cell F1-score on the validation set is greater than or equal to the ninth threshold value, and the rare cell F1-score fluctuation of continuous three rounds of iteration is less than or equal to a preset fluctuation threshold value.

[0011] In some embodiments, the inference decision module is configured to determine an output strategy according to the confidence distribution of the clustering results and the matching degree of the population abundance and the physiological and pathological conventional range, and output an analysis result with a confidence score, including: A trained density peak clustering model is used to preliminarily cluster the input cell events, and the suspected rare cell clusters are labeled; The suspected rare cell clusters are input into the lightweight convolutional neural network model, and three types of results of target rare cells, non-target rare cells, and uncertainty are output. wherein, when the probability output by the lightweight convolutional neural network model is greater than or equal to a preset first probability threshold, the target rare cell is determined; when the probability output by the lightweight convolutional neural network model is less than or equal to a preset second probability threshold, the non-target rare cell is determined; when the probability output by the lightweight convolutional neural network model is greater than the second probability threshold and less than the first probability threshold, the uncertainty is determined.

[0012] In some embodiments, the inference decision module is configured to determine an output strategy according to the confidence distribution of the clustering result and the matching degree of the population abundance and the physiological and pathological conventional range, and when outputting the analysis result with a confidence score, the inference decision module further comprises: when the proportion of the number of the target rare cells in the target sample is less than a preset first proportion threshold, and the proportion of the uncertain result is greater than a preset second proportion threshold, it is determined that the rare cell content in the sample is extremely low, the clustering result is low in reliability, and a first warning signal is issued; otherwise, the first warning signal is not issued; when the proportion of the number of the target rare cells in the target sample is greater than a preset third proportion threshold, the original fluorescence signal of the sample is rechecked, and if the original fluorescence signal is abnormal, a second warning signal is issued; otherwise, the second warning signal is not issued; when the proportion of the number of the target rare cells in the target sample is greater than a preset third proportion threshold, the original fluorescence signal of the sample is rechecked, and if the original fluorescence signal is abnormal, a third warning signal is issued; otherwise, the third warning signal is not issued.

[0013] In some embodiments, the iterative upgrading module is configured to evaluate the adaptability of the model to new samples by accumulating data, and determine whether the accumulated data quantity meets the iteration requirement, when the adaptability is less than a preset threshold and the data quantity meets the requirement, trigger the model iteration optimization, comprising: when the number of artificially annotated samples accumulated is greater than or equal to a preset tenth threshold, and the model judgment error rate in the new sample is greater than a preset error rate threshold, trigger the model iteration optimization; otherwise, the model iteration optimization is not triggered.

[0014] Compared with the prior art, the beneficial effects of the present application are that: through the hybrid model of unsupervised clustering and supervised classification confirmation, combined with the SMOTE algorithm to supplement samples, both the rare cells are avoided to be covered by the mainstream cells and the small sample training problem is solved, the recall rate and the precision rate of the rare cells are greatly improved, and the needs of clinical microcell analysis are met; the data preprocessing module is targeted to remove impurities, adherent cells and optimize features, and provides high-quality data for subsequent analysis; the training optimization module ensures that the model meets the standard and is stable through multiple rounds of training and parameter adjustment; the reasoning decision module outputs the results combined with the confidence and the physiological and pathological range, and the warning mechanism is matched to further reduce the risk of misjudgment; the iterative upgrade module is based on cumulative data evaluation adaptability, automatically triggers optimization, does not need manual re-development, prolongs the system life cycle, reduces the use cost, and can flexibly adapt to different batches and different types of sample analysis needs.

[0015] The above general description and the following detailed description are only exemplary and explanatory, but not limiting the present disclosure.

[0016] Other features and aspects of the present disclosure will be more clearly understood from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the specific embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings based on these drawings without creative labor.

[0018] Figure 1 The function block diagram of the machine learning-based intelligent flow cytometry cell analysis system provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to convey the scope of the present disclosure to those skilled in the art. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0020] Reference Figure 1 As shown in the drawings, the machine learning-based intelligent flow cytometry cell analysis system according to the embodiments of the present application comprises: a data preprocessing module configured to receive multidimensional fluorescence signal data output by a flow cytometer, and perform cleaning and feature engineering processing on the multidimensional fluorescence signal data; a hybrid machine learning model module including an unsupervised clustering unit and a supervised classification unit, the unsupervised clustering unit being configured to receive standardized feature data, determine which populations belong to potential rare cell populations by analyzing the abundance of each cell population and the relative difference of the main population, and mark as suspected rare cell clusters; the supervised classification unit being configured to receive the suspected rare cell clusters, determine whether the suspected populations are target rare cells, and output target rare cell clustering results; a training optimization module configured to receive clustering results and external annotation data sources, determine whether the model meets the preset performance benchmark based on the comprehensive recognition performance of the model on rare cells, and determine whether the model is stable by the performance fluctuation amplitude of continuous multiple rounds of training, and perform multiple rounds of training and parameter adjustment on the model when the model does not meet the benchmark or is not stable; an inference decision module configured to determine an output strategy according to the confidence distribution of the clustering results and the matching degree of the population abundance and the physiological and pathological conventional range, and output an analysis result with a confidence score; an iterative upgrade module configured to evaluate the adaptability of the model to new samples by accumulating data, determine whether the accumulated data quantity meets the iteration requirement, and trigger model iteration optimization when the adaptability is less than a preset threshold and the data quantity meets the requirement.

[0021] In some embodiments, the data preprocessing module is configured to receive multidimensional fluorescence signal data output by a flow cytometer, and perform cleaning and feature engineering processing on the multidimensional fluorescence signal data, including: obtaining a scatter plot of forward scatter light signals and side scatter light signals in the multidimensional fluorescence signal data, and presetting a first threshold value, when the signal intensity of a particle in the scatter plot is greater than or equal to the first threshold value, the particle is retained; when the signal intensity of the particle in the scatter plot is less than the first threshold value, the particle is removed; calculating the ratio of FSC width to FSC height of each cell event, and presetting a second threshold value, when the ratio of FSC width to FSC height is less than or equal to the second threshold value, the event is retained; when the ratio of FSC width to FSC height is greater than the second threshold value, the event is determined to be a clumped cell and is removed; performing Z-score standardization processing on the signal values of each fluorescence channel, and presetting a third threshold value range, when the Z-score of any fluorescence channel of a cell event after standardization is greater than or less than the third threshold value range, the cell event is determined to be an outlier, and the outlier is replaced by a nearest neighbor interpolation method; When the Z-score of any fluorescence channel of a certain cell event after normalization equals to the third threshold range, the cell event is determined as normal value.

[0022] It should be understood that the first threshold is set to screen cells by the scatter plot of forward scatter light (FSC, reflecting cell size) and side scatter light (SSC, reflecting cell internal complexity); The first threshold = 200 (set according to: the FSC mean value of normal immune cells is about 1000, and the FSC of non-cell particles (such as impurities) is usually < 200, and 1 / 5 of the mean value is taken as the screening standard); The signal intensity ≥ 200 is determined as a cell and retained; and < 200 is determined as a non-cell particle and rejected; The ratio of FSC width (FSC-W) to FSC height (FSC-H) of each cell event is calculated, and whether it is a sticking cell (multiple cells overlapping by the instrument) is determined by the second threshold; The second threshold = 1.2 (set according to: the FSC-W / FSC-H ratio of a single cell is usually between 1.0-1.2, and more than 1.2 is a multiple cell sticking); The ratio ≤ 1.2 is determined as a single cell and retained; and > 1.2 is determined as a sticking cell and rejected; The Z-score standardization (eliminating the difference in signal range of different channels) is performed on each fluorescence channel signal, and whether it is an abnormal value is determined by the third threshold range; The third threshold range = ± 3 (set according to: in statistics, Z-score values outside ± 3 are extreme abnormal values, and the occurrence probability is < 0.3%); The Z-score within ± 3 is determined as a normal value and retained; and > 3 or < -3 is determined as an abnormal value and replaced by the "nearest neighbor interpolation method" (taking the mean value of the characteristics of normal cells in the same sample).

[0023] In some specific embodiments, the data preprocessing module is configured to receive multi-dimensional fluorescence signal data output by a flow cytometer, and when performing cleaning and feature engineering processing on the multi-dimensional fluorescence signal data, further comprising: Obtaining a plurality of fluorescence characteristics related to cell events detected by a flow cytometer, and calculating the mutual information value between each of the fluorescence characteristics and a known rare cell population label, wherein the known rare cell population label is the identification of a pre-labeled rare immune cell population; When the mutual information value of the fluorescence characteristic is less than a preset fourth threshold, it is determined that the fluorescence characteristic has no contribution to the rare immune cell population clustering, and it is excluded from the feature set; when the mutual information value of the fluorescence feature is greater than or equal to the fourth threshold value, it is determined that the fluorescence feature contributes to the grouping of the rare immune cell population, and is retained from the feature set; The retained fluorescence features are subjected to logarithmic conversion and normalization processing.

[0024] It should be understood that the flow cytometer usually detects 10-15 fluorescence features (such as CD4, CD25, Foxp3, etc.), but not all features are valuable for rare cell grouping. By calculating the mutual information value (measuring the degree of association between the feature and the label, the higher the value, the stronger the association) of each fluorescence feature with the "known rare cell population label" (artificially labeled target rare cells such as Treg cells), a fourth threshold value (such as 0.1) is set: when the mutual information value is greater than or equal to 0.1, it is determined that it is an effective feature that contributes to the grouping (such as CD25 and Treg strong association), and is retained; <0.1, it is determined to be a redundant feature (such as a non-specific feature related to cell size), and is removed. This step reduces the feature dimension and allows the model to focus on the key signals that can distinguish rare cells; The original fluorescence signal is usually "right-skewed distribution" (main cell signal is strong, and rare cell signal is weak, such as the mean value of CD25 signal of Treg is 100, and that of conventional T cell is 1000), and direct input into the model will cause the rare cell feature to be ignored. Therefore, the retained features are first subjected to logarithmic conversion (log10(x+1)), which reduces the gap between strong and weak signals (such as 100→2.004, 1000→3.000, the gap is reduced from 10 times to 1.04 times), and amplifies the weak signal of the rare cell; then the converted features are normalized to the interval [0, 1] to eliminate the signal magnitude difference between different channels (such as CD4 signal range 100-10000, Foxp350-5000), so as to avoid the model ignoring the weak features of the rare cell due to "high weight of strong signal features".

[0025] In some specific embodiments, the unsupervised clustering unit is configured to receive standardized feature data, and when it is determined that some groups belong to potential rare cell groups by analyzing the abundance of each cell group and the relative difference from the mainstream group, and marked as suspected rare cell clusters, it includes: The preprocessed multi-dimensional fluorescence signal data is subjected to preliminary grouping by using a density peak clustering algorithm, and the clustering parameters of the density peak clustering algorithm are set, including: The clustering effect is judged by the silhouette coefficient, and when the silhouette coefficient is less than a preset fifth threshold value, the cutoff distance of the density peak clustering algorithm is adjusted; When the silhouette coefficient is greater than or equal to the fifth threshold value, the cutoff distance of the density peak clustering algorithm is not adjusted; The number proportion of cell events in each cluster obtained by statistical clustering, when the number proportion of cell events of a certain cluster is less than a preset sixth threshold, the cluster is marked as the suspected rare cell cluster; When the number proportion of cell events of a certain cluster is greater than or equal to the sixth threshold, the cluster is determined as a mainstream cell cluster, and the clustering result of the mainstream cell cluster is output.

[0026] It should be understood that the role of the unsupervised clustering unit is to "preset-free explore cell population", and potential rare cell clusters are preliminarily screened from all cells to narrow the range for subsequent fine identification. The traditional method needs to manually preset the number of clusters, and unknown rare cells are easy to be missed. The density peak clustering (DPC) algorithm does not need to preset the number, and is more suitable for exploratory clustering. The core parameter of the DPC algorithm is the "cut-off distance dc", which directly affects the calculation of cell density and clustering results. The clustering quality is judged by the silhouette coefficient (an index for evaluating clustering effect, taking value [-1, 1], the closer to 1, the higher the similarity within the cluster and the greater the difference between clusters), and the fifth threshold (such as 0.5) is set: when the silhouette coefficient is <0.5, it means that the clusters overlap seriously (such as T cell and B cell cluster cross), and the dc needs to be gradually increased (each time +0.1), and the density and clustering results are recalculated; when the silhouette coefficient is ≥0.5, the clustering effect is acceptable, and the adjustment of dc is stopped. This step ensures that the mainstream cell population (such as T cells and B cells) is accurately clustered first, laying a foundation for subsequent screening of rare cells. The biological characteristics of rare immune cells determine that their proportion is usually <5% (such as Treg accounting for 1%-3% of peripheral blood T cells), so by counting the number proportion of cells in each cluster, the sixth threshold (such as 5%) is set: when the proportion of a certain cluster is <5%, it is marked as a "suspected rare cell cluster" (may contain target rare cells) and sent to the supervised classification unit for further identification; when the proportion is ≥5%, it is determined as a mainstream cell cluster (such as T cells accounting for 60%), and the clustering result is directly output. This step quickly locks the potential rare cells through "abundance difference", avoiding that the characteristics of rare cells are covered by the mainstream cell population.

[0027] In some specific embodiments, the supervised classification unit is configured to receive the suspected rare cell cluster, determine whether the suspected population is a target rare cell, and output a target rare cell clustering result, including: A lightweight convolutional neural network model is used to identify the suspected rare cell cluster; The number of cell events in the suspected rare cell cluster is obtained, and when the number of cell events is less than a preset number threshold, it is determined that the sample size is insufficient, then a synthetic sample is generated through the SMOTE algorithm, and the feature value of the synthetic sample needs to fall within the 95% confidence interval of the original sample feature; when the number of the cell events is greater than or equal to the number threshold, using the suspected rare cell cluster as a training set for training of the lightweight convolutional neural network model; calculating a rare cell recall rate and a rare cell precision rate of the model on the verification set, when the rare cell recall rate is less than a preset seventh threshold, increasing a number of hidden layers of the lightweight convolutional neural network, and adjusting a learning rate; when the rare cell recall rate is greater than or equal to the seventh threshold, and the rare cell precision rate is less than a preset eighth threshold, adding a Dropout layer to an output layer of the lightweight convolutional neural network; otherwise, not adjusting the output layer of the lightweight convolutional neural network.

[0028] It should be understood that the role of the supervised classification unit is to finely identify the "suspected rare cell cluster" and confirm whether it is the target rare cell; the unsupervised clustering can only screen potential clusters by proportion, and the characteristics of the rare cell and other low-abundance cells (such as impurities) may overlap, and a supervised model (lightweight CNN) is needed to capture subtle differences; The amount of rare cell sample is small (often <100), and direct training will cause model overfitting (unable to generalize to new samples). A preset number threshold (such as 100) is set: when the number of suspected cluster cells is <100, synthetic samples are generated by SMOTE algorithm (new samples are generated in the original sample feature space by interpolation, and the feature values need to fall within the 95% confidence interval of the original sample to avoid generating false features), and supplemented to 100; when the number is ≥100, it is directly used as a training set. This step ensures the amount of training data, so that the model can learn the stable characteristics of the rare cell.

[0029] The advantage of the lightweight CNN is to capture the nonlinear correlation between features (such as the synergistic expression of CD25 and Foxp3), but the structure needs to be adjusted according to the performance. Through the "rare cell recall rate" (miss rate) and "precision rate" (false detection rate) on the verification set: set the seventh threshold (such as 80%, recall rate) and the eighth threshold (such as 70%, precision rate), when the recall rate is <80% (missed more), increase the CNN hidden layer (from 2 layers to 3 layers) and reduce the learning rate (from 0.01 to 0.001), to improve the learning ability of the model; when the recall rate is ≥80% but the precision rate is <70% (false detection is more), add a Dropout layer (dropoutrate=0.3) to the output layer to suppress overfitting; if both are up to standard, no adjustment is needed. This step ensures that the model can "catch" the rare cell (high recall rate) and "judge accurately" (high precision rate) through targeted optimization.

[0030] In some specific embodiments, the training optimization module is configured to receive clustering results and external labeled data sources, determine whether the model has reached a preset performance benchmark based on the model's comprehensive recognition performance of rare cells, and determine whether the model is stable by the performance fluctuation range of continuous training rounds. When the benchmark is not met or the model is unstable, the module performs multiple rounds of training and parameter adjustment, including: The hybrid machine learning model module is trained using the training set, and the rare cell F1-score is calculated on the validation set. When the F1-score of the rare cell is greater than or equal to the preset ninth threshold, the hybrid machine learning model module is deemed to have met the standard and enters the test set for verification. When the F1-score of the rare cell is less than the ninth threshold, iterative optimization begins.

[0031] It should be understood that the benchmarking of the training optimization module is to ensure that the model performance meets the actual needs. A model trained only once may have insufficient accuracy or overfitting, and a comprehensive index is needed to evaluate whether it is "usable". The core logic uses the "rare cell F1-score" as the evaluation standard: the F1-score is the harmonic mean of recall (R) and precision (P) (F1=2*(P*R) / (P+R)), with a value range of [0,1]. It can balance false negatives and false positives, and is more suitable for evaluating the performance of rare cell recognition (looking at recall or precision alone may be one-sided, as high recall may be accompanied by high false positives). A ninth threshold is set (e.g., 0.85): when the validation set F1-score ≥ 0.85, the model is considered to have met the standard (the overall accuracy meets the needs of scientific research / clinical practice) and enters the test set for validation (using samples not used in training to evaluate generalization ability); when the F1-score < 0.85, it is considered to have failed the standard and enters the iterative optimization stage. This step, through a clear performance benchmark, avoids the use of models with insufficient accuracy and ensures the reliability of the clustering results.

[0032] In some specific embodiments, the training optimization module is configured to receive clustering results and external labeled data sources, determine whether the model has reached a preset performance benchmark based on the model's comprehensive recognition performance of rare cells, and determine whether the model is stable by the performance fluctuation range of multiple rounds of training. When the benchmark is not met or the model is unstable, the module further includes: The characteristics of the erroneous samples in the rare cell identification set were analyzed and verified. When the erroneous samples were concentrated in the overlapping region of fluorescence signals, new features were added and the model was retrained. If the erroneous samples are randomly distributed, the batch size of the lightweight convolutional neural network model is adjusted. The above optimization training is repeatedly performed until the rare cell F1-score on the validation set is greater than or equal to the ninth threshold value, and the rare cell F1-score of three consecutive iterations fluctuates less than or equal to a preset fluctuation threshold value.

[0033] It should be understood that the purpose of iterative optimization is to continuously improve the performance until it is stable by "locating error causes and making targeted adjustments" for the substandard model; the substandard model is often due to insufficient features or training fluctuations, which need to be solved through multiple iterations; Specific working logic: First, analyze the characteristics of the "rare cell recognition error samples" in the validation set: if the errors are concentrated in the "fluorescent signal overlap area" (such as the CD25 signal intersection of Treg and conventional T cells), it indicates that the existing features are insufficient to distinguish, and new features need to be added (such as the ratio of CD25 to Foxp3, which is higher in Treg), and retraining is needed; if the error samples are randomly distributed (irregular), it indicates that the training process fluctuates greatly, and the batch size of the CNN needs to be adjusted (from 32 to 64) to reduce the sample difference of single training and improve stability. Specific working logic: First, analyze the characteristics of the "rare cell recognition error samples" in the validation set: if the errors are concentrated in the "fluorescent signal overlap area" (such as the CD25 signal intersection of Treg and conventional T cells), it indicates that the existing features are insufficient to distinguish, and new features need to be added (such as the ratio of CD25 to Foxp3, which is higher in Treg), and retraining is needed; if the error samples are randomly distributed (irregular), it indicates that the training process fluctuates greatly, and the batch size of the CNN needs to be adjusted (from 32 to 64) to reduce the sample difference of single training and improve stability.

[0034] The iteration needs to meet the "double convergence condition": the F1-score of the validation set is greater than or equal to 0.85 (up to standard), and the F1-score fluctuation of three consecutive iterations is less than or equal to 0.02 (stable, to avoid accidental standard). For example, the F1-score of three consecutive iterations is 0.86, 0.87, and 0.86, and the fluctuation is less than or equal to 0.02, which is determined to be converged. This step ensures that the model not only meets the accuracy standard, but also can stably output results, avoiding group error caused by training fluctuations, through repeated optimization and stability verification.

[0035] In some specific embodiments, the inference decision module is configured to determine an output strategy according to the confidence distribution of the clustering result and the matching degree of the population abundance and the physiological and pathological conventional range, and output an analysis result with a confidence score. The trained density peak clustering model is used to preliminarily cluster the input cell events, and the suspected rare cell clusters are labeled; inputting the suspected rare cell cluster into the lightweight convolutional neural network model, outputting a target rare cell, a non-target rare cell, and an uncertain three-class result; wherein, when the lightweight convolutional neural network model outputs a probability greater than or equal to a preset first probability threshold, the target rare cell is determined; when the lightweight convolutional neural network model outputs a probability less than or equal to a preset second probability threshold, the non-target rare cell is determined; when the lightweight convolutional neural network model outputs a probability greater than the second probability threshold and less than the first probability threshold, the uncertain is determined.

[0036] It should be understood that cluster reasoning is to apply the trained model to new samples to output preliminary clustering results, providing a basis for subsequent "result reliability judgment"; the reasoning stage needs to quickly and accurately identify target rare cells from new samples; Use the trained DPC model to cluster the input new sample (such as patient peripheral blood), and based on the cluster proportion <5% standard, mark "suspected rare cell cluster" to quickly narrow down the analysis range; Input the suspected cluster into the trained lightweight CNN model, and the model outputs the probability of each cell "belonging to the target rare cell". Set the first probability threshold (such as 0.9) and the second probability threshold (such as 0.1): when the probability ≥0.9, it is determined as "target rare cell" (high confidence); ≤0.1, it is determined as "non-target rare cell" (such as other low-abundance impurities); 0.1-0.9, it is determined as "uncertain" (manual review is required). This step divides the result type by probability threshold, which not only guarantees the reliability of high-confidence results, but also leaves room for manual intervention through "uncertain" label, balancing automation and accuracy.

[0037] In some specific embodiments, the inference decision module is configured to determine the output strategy according to the confidence distribution of the clustering result and the matching degree of the population abundance and the physiological and pathological conventional range, and when outputting the analysis result with a confidence score, it further includes: When the number proportion of the target rare cell of the target sample is less than a preset first proportion threshold, and the uncertain result proportion is greater than a preset second proportion threshold, it is determined that the rare cell content in the sample is extremely low, and the clustering result is low in reliability, and a first warning signal is issued; Otherwise, no first warning signal is issued; When the number proportion of the target rare cell of the target sample is greater than a preset third proportion threshold, the original fluorescence signal of the sample is rechecked, and if the original fluorescence signal is normal, a second warning signal is issued; Otherwise, no second warning signal is issued; When the proportion of the number of the target rare cells in the target sample is greater than a preset third proportion threshold, the original fluorescence signal of the sample is rechecked, and if the original fluorescence signal is abnormal, a third early warning signal is issued; Otherwise, the third early warning signal is not issued.

[0038] It should be understood that the role of result judgment and early warning is to perform "reliability verification" on the clustering result, so as to avoid false conclusions caused by sample quality, cell proportion abnormality, etc.; the rare cell analysis result is often used for clinical diagnosis (such as immunodeficiency disease), and a false result may cause misdiagnosis; When the "proportion of target rare cells <1%" (the content is extremely low, and the statistical error is large) and the "proportion of uncertain results >30%" (the model judgment reliability is low), the first early warning ("it is suggested to increase the sample detection amount") is triggered. This is because when the rare cells in the sample are too few, the clustering result is greatly affected by random error, and more samples are needed to reduce statistical fluctuation; The normal proportion of rare cells is usually 0.5%-5%, if the proportion of a certain sample is >10% (far beyond the physiological range), "feature backtracking verification" is triggered: if the original fluorescence signal is not abnormal (the Z-score is within ±3), it indicates that it may be a pathological state (such as immune disorder), and the second early warning ("abnormal increase of rare cells, combined with clinical verification") is triggered; if the signal is abnormal (there is a signal outside the Z-score ±3), it indicates that the sample is contaminated or the instrument is malfunctioning, and the third early warning ("recollect the sample") is triggered. This step distinguishes between "true abnormality" and "false positive" by combining the physiological range and signal quality; If the above early warnings are not triggered, the complete result is output, including the main cell clustering, the rare cell clustering (including the proportion, the mean value of the fluorescence characteristics), and the confidence score (0-10 points, based on F1-score, such as F1=0.9 corresponding to 10 points). This step ensures that reliable results can be directly used for analysis, and at the same time, the result quality is directly reflected through the score.

[0039] In some specific embodiments, the iterative upgrading module is configured to evaluate the adaptability of the model to new samples by accumulating data, and to determine whether the accumulated data quantity meets the iteration requirement, and when the adaptability is less than a preset threshold and the data quantity meets the requirement, the model iteration optimization is triggered, including: When the number of artificially annotated samples accumulated is greater than or equal to a preset tenth threshold, and the model judgment error rate in the new sample is greater than a preset error rate threshold, the model iteration optimization is triggered; Otherwise, the model iteration optimization is not triggered.

[0040] It should be understood that the purpose of the iterative upgrading module is to let the model continuously evolve with new samples and new cell subtypes, so as to avoid "model obsolescence"; new subtypes of immune cells are continuously discovered, and instrument calibration and sample source changes may cause feature drift, so the model needs to be dynamically adapted; The system automatically records the "clustering result + manual review feedback" (laboratory personnel mark the true type for "uncertain" results) of each sample, accumulates "new data" for iteration; When the accumulated manually labeled samples ≥ 500 (the data volume is sufficient to support iteration), and the "model judgment error rate" (the proportion of manual review inconsistent with the model result) in new samples > 10% (the model adaptability has decreased), iteration is triggered. The threshold of 500 samples ensures that enough variation types are covered, and the 10% error rate reflects that the model performance has failed to meet the demand; Analyze error sample types: if the errors are concentrated in "new subtype rare cells" (such as a new subtype of dendritic cells), it indicates that the existing clustering does not cover new groups, and the unsupervised clustering unit needs to be returned to for re-exploration; if the errors are concentrated in "known cell feature drift" (such as a general decrease in CD4 signal after instrument calibration), it indicates that the model parameters need to be fine-tuned, and the training process is returned to for fine-tuning with new samples at a low learning rate (0.0001) and a small number of rounds (10 rounds) without retraining. This step allows the model to both recognize new subtypes and adapt to feature drift, maintaining long-term effectiveness.

[0041] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage media, etc.) having computer usable program code embodied thereon.

[0042] The application is described with reference to the flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The functions specified in one flow or multiple flows and / or blocks

[0043] These computer program instructions can also be stored in a computer readable storage medium that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a manufactured product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams.Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0044] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A machine learning-based intelligent flow cytometry system for analyzing immune cells, characterized in that, include: The data preprocessing module is configured to receive multidimensional fluorescence signal data output by a flow cytometer and perform cleaning and feature engineering processing on the multidimensional fluorescence signal data. The hybrid machine learning model module includes an unsupervised clustering unit and a supervised classification unit. The unsupervised clustering unit is configured to receive standardized feature data, analyze the relative difference between the abundance of each cell population and the mainstream population, determine which populations belong to potential rare cell populations, and mark them as suspected rare cell clusters. The supervised classification unit is configured to receive the suspected rare cell clusters, determine whether the suspected populations are target rare cells, and output the target rare cell clustering results. The training optimization module is configured to receive clustering results and external labeled data sources, judge whether the model has reached the preset performance benchmark based on the model's comprehensive recognition performance of rare cells, and judge whether the model is stable by the performance fluctuation range of multiple rounds of training. When the benchmark is not met or the model is unstable, multiple rounds of training and parameter adjustment are performed on the model. The reasoning and decision-making module is configured to determine the output strategy based on the confidence distribution of the clustering results and the matching degree between the population abundance and the conventional range of physiological and pathological conditions, and output the analysis results with confidence scores. The iterative upgrade module is configured to evaluate the model’s suitability for new samples through accumulated data, determine whether the amount of accumulated data meets the iteration requirements, and trigger model iterative optimization when the suitability is less than a preset threshold and the amount of data meets the standard.

2. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 1, characterized in that, The data preprocessing module is configured to receive multidimensional fluorescence signal data output from a flow cytometer, and the cleaning and feature engineering processing of the multidimensional fluorescence signal data includes: A scatter plot of forward-scattered light signal and side-scattered light signal in the multidimensional fluorescence signal data is obtained. A first threshold is preset. When the particle signal intensity in the scatter plot is greater than or equal to the first threshold, the particle is retained. When the particle signal intensity in the scatter plot is less than the first threshold, the particle is removed. Calculate the ratio of FSC width to FSC height for each cell event, and preset a second threshold. When the ratio of FSC width to FSC height is less than or equal to the second threshold, the event is retained. When the ratio of the FSC width to the FSC height is greater than the second threshold, the event is determined to be an adherent cell and it is removed. The signal value of each fluorescence channel is Z-score normalized. A third threshold range is preset. When the normalized Z-score of any fluorescence channel of a certain cell event is greater than or less than the third threshold range, the cell event is determined to be an outlier. The outlier is replaced by nearest neighbor interpolation. When the normalized Z-score of any fluorescence channel of a certain cell event is equal to the third threshold range, the cell event is determined to be a normal value.

3. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 2, characterized in that, The data preprocessing module is configured to receive multidimensional fluorescence signal data output from a flow cytometer, and when cleaning and feature-engineering the multidimensional fluorescence signal data, it further includes: Acquire multiple fluorescence features related to cellular events detected by flow cytometry, and calculate the mutual information value between each fluorescence feature and a known rare cell population label, wherein the known rare cell population label is an identifier of a pre-labeled rare immune cell population; When the mutual information value of the fluorescence feature is less than a preset fourth threshold, it is determined that the fluorescence feature does not contribute to the classification of rare immune cell populations and is removed from the feature set. When the mutual information value of the fluorescence feature is greater than or equal to the fourth threshold, it is determined that the fluorescence feature contributes to the classification of rare immune cell populations and is retained from the feature set. The retained fluorescence features are then subjected to logarithmic transformation and normalization.

4. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 3, characterized in that, The unsupervised clustering unit is configured to receive standardized feature data, analyze the relative differences between the abundance of each cell population and the mainstream population, determine which populations belong to potential rare cell populations, and mark them as suspected rare cell clusters, including: The preprocessed multidimensional fluorescence signal data is initially clustered using a density peak clustering algorithm. The clustering parameters of the density peak clustering algorithm are set, including: The clustering effect is judged by the silhouette coefficient. When the silhouette coefficient is less than the preset fifth threshold, the cutoff distance of the density peak clustering algorithm is adjusted. When the contour coefficient is greater than or equal to the fifth threshold, the cutoff distance of the density peak clustering algorithm is not adjusted. The percentage of cell events in each cluster obtained by statistical clustering is counted. When the percentage of cell events in a certain cluster is less than a preset sixth threshold, the cluster is marked as the suspected rare cell cluster. When the proportion of cell events in a certain cluster is greater than or equal to the sixth threshold, the cluster is determined to be the mainstream cell cluster, and the clustering result of the mainstream cell cluster is output.

5. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 4, characterized in that, The supervised classification unit is configured to receive the suspected rare cell clusters, determine whether the suspected group is the target rare cell, and output the target rare cell clustering result, including: A lightweight convolutional neural network model was used to identify the suspected rare cell clusters; The number of cell events in the suspected rare cell cluster is obtained. When the number of cell events is less than a preset threshold, it is determined that the sample size is insufficient. Then, a synthetic sample is generated by the SMOTE algorithm, and the feature value of the synthetic sample must fall within the 95% confidence interval of the original sample feature. When the number of cell events is greater than or equal to the number threshold, the suspected rare cell clusters are used as a training set for training the lightweight convolutional neural network model. The rare cell recall and precision of the model are calculated on the validation set. When the rare cell recall is less than a preset seventh threshold, the number of hidden layers of the lightweight convolutional neural network is increased, and the learning rate is adjusted. When the rare cell recall rate is greater than or equal to the seventh threshold and the precision rate is less than the preset eighth threshold, a Dropout layer is added to the output layer of the lightweight convolutional neural network. Otherwise, the output layer of the lightweight convolutional neural network is not adjusted.

6. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 5, characterized in that, The training optimization module is configured to receive clustering results and external labeled data sources, determine whether the model has reached a preset performance benchmark based on the model's comprehensive recognition performance of rare cells, and determine whether the model is stable by the performance fluctuation range of multiple rounds of training. When the benchmark is not met or the model is unstable, multiple rounds of training and parameter adjustment are performed on the model, including: The hybrid machine learning model module is trained using the training set, and the rare cell F1-score is calculated on the validation set. When the F1-score of the rare cell is greater than or equal to the preset ninth threshold, the hybrid machine learning model module is deemed to have met the standard and enters the test set for verification. When the F1-score of the rare cell is less than the ninth threshold, iterative optimization begins.

7. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 6, characterized in that, The training optimization module is configured to receive clustering results and external labeled data sources, determine whether the model has reached a preset performance benchmark based on the model's comprehensive recognition performance of rare cells, and determine whether the model is stable by the performance fluctuation range of multiple rounds of training. When the benchmark is not met or the model is unstable, the module also includes performing multiple rounds of training and parameter adjustments. The characteristics of the erroneous samples in the rare cell identification set were analyzed and verified. When the erroneous samples were concentrated in the overlapping region of fluorescence signals, new features were added and the model was retrained. If the erroneous samples are randomly distributed, the batch size of the lightweight convolutional neural network model is adjusted. Repeat the above optimization training until the F1-score of the rare cells on the validation set is greater than or equal to the ninth threshold, and the fluctuation of the F1-score of the rare cells in three consecutive iterations is less than or equal to the preset fluctuation threshold.

8. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 7, characterized in that, The reasoning and decision-making module is configured to determine the output strategy based on the confidence distribution of the clustering results and the matching degree between the population abundance and the conventional range of physiological and pathological conditions. When outputting analysis results with confidence scores, it includes: The input cell events were initially clustered using a trained density peak clustering model, and the suspected rare cell clusters were identified. The suspected rare cell clusters are input into the lightweight convolutional neural network model, which outputs three types of results: target rare cells, non-target rare cells, and uncertain results. Wherein, when the output probability of the lightweight convolutional neural network model is greater than or equal to a preset first probability threshold, it is determined to be the target rare cell; When the output probability of the lightweight convolutional neural network model is less than or equal to a preset second probability threshold, it is determined to be a non-target rare cell; When the output probability of the lightweight convolutional neural network model is greater than the second probability threshold and less than the first probability threshold, it is determined to be uncertain.

9. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 8, characterized in that, The reasoning and decision-making module is configured to determine the output strategy based on the confidence distribution of the clustering results and the matching degree between the population abundance and the conventional range of physiological and pathological conditions. When outputting analysis results with confidence scores, it also includes: When the proportion of the target rare cells in the target sample is less than a preset first proportion threshold and the proportion of uncertain results is greater than a preset second proportion threshold, it is determined that the content of rare cells in the sample is extremely low, the reliability of the clustering result is low, and a first warning signal is issued. Otherwise, the first warning signal will not be issued; When the proportion of the target rare cells in the target sample is greater than the preset third proportion threshold, the original fluorescence signal of the sample is re-examined. If the original fluorescence signal is normal, a second warning signal is issued. Otherwise, no second warning signal will be issued; When the proportion of the target rare cells in the target sample is greater than the preset third proportion threshold, the original fluorescence signal of the sample is re-examined. If the original fluorescence signal is abnormal, a third warning signal is issued. Otherwise, no third warning signal will be issued.

10. The intelligent flow cytometry immune cell analysis system based on machine learning according to claim 9, characterized in that, The iterative upgrade module is configured to evaluate the model's suitability for new samples using accumulated data, determine whether the accumulated data volume meets the iteration requirements, and trigger model iterative optimization when the suitability is less than a preset threshold and the data volume meets the standard, including: When the number of accumulated manually labeled samples is greater than or equal to the preset tenth threshold, and the model's error rate in new samples is greater than the preset error rate threshold, the model is triggered to iterate and optimize. Otherwise, model iterative optimization will not be triggered.

Images