Cell discrimination model construction method and apparatus, electronic device, and storage medium

Abstract

The present invention provides a cell discrimination model construction method and apparatus, an electronic device, and a storage medium. The method comprises: on the basis of cell fluorescence images and cell tags of sample cells, training a cell classification model constructed on the basis of a neural network; in addition to the cell classification model, further constructing a machine learning model, so that morphological and statistical features of the cells and signal points are extracted by means of feature engineering; classifying the cells in another dimension by using the morphological and statistical features, so as to supplement the cell classification model; and by fusing the machine learning model and the cell classification model, performing classification discrimination on the cells from different angles by using different features. In this way, interference caused by background noise and impurities in the cell classification discrimination process is reduced, the accuracy of cell classification discrimination is improved, the cell count requiring manual discrimination can be effectively reduced while the recall rate is guaranteed, and the subsequent discrimination efficiency is increased.

Description

Cell discrimination model construction method, device, electronic device and storage medium Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a method, device, electronic device and storage medium for constructing a cell discrimination model. Background Art

[0002] In the process of cell detection based on fluorescence in situ hybridization (FISH), in order to obtain high-quality detection results, a sample needs to accurately detect specific types of cells from tens of thousands of cells, such as circulating genetically abnormal cells (CAC). The discrimination criteria for CAC cells are based on the fluorescence expression of the cells under each channel. However, when using the fluorescence signal gain of FISH to identify CACs cells, due to factors such as the experimental environment and experimental materials, the hybridized fluorescent signal is actually background spots and impurities, which will affect the cell discrimination results, resulting in low accuracy of cell discrimination results, and a high number of specific types of cells detected. When manual judgment is performed later, a high number of cells need to be manually judged, affecting the subsequent judgment efficiency and throughput.

[0003] Summary of the Invention

[0004] The present invention provides a cell discrimination model construction method, device, electronic device and storage medium to solve the defect in the prior art that cell discrimination results are easily interfered with by factors such as experimental environment and experimental raw materials, resulting in low accuracy of cell discrimination results.

[0005] The present invention provides a method for constructing a cell discrimination model, comprising:

[0006] Based on the cell fluorescence images of sample cells and their cell labels, a cell classification model based on a neural network is trained;

[0007] determining the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, performing feature selection on the candidate features, and after obtaining selected features, constructing a machine learning model based on the selected features as a cell discrimination model;

[0008] Alternatively, the cell features and signal point features of the sample cells are determined as candidate features, and feature selection is performed on the candidate features. After obtaining the selected features, a machine learning model is constructed based on the selected features, and the cell classification model and the machine learning model are fused to obtain a cell discrimination model.

[0009] According to a method for constructing a cell discrimination model provided by the present invention, the cell classification model constructed based on a neural network is trained based on the cell fluorescence images of sample cells and their cell labels, specifically comprising:

[0010] Obtain cell fluorescence images and cell labels of different types of real sample cells in clinical samples;

[0011] Clustering the cell fluorescence images of the different types of real sample cells, and obtaining target cell clusters corresponding to the target type;

[0012] generating a synthetic image of cells of the target type based on the cell fluorescence images of each real sample cell of the target type in the target cell cluster;

[0013] The cell classification model is obtained by training based on the cell fluorescence image and its cell label and the cell composite image and its cell label.

[0014] According to a cell discrimination model construction method provided by the present invention, the cell fluorescence image of the real sample cell is obtained based on the following steps:

[0015] Obtain the contour information of each real sample cell in the cell image based on the cell segmentation model;

[0016] Based on the contour information of the real sample cells, regions corresponding to the real sample cells in the fluorescence images of each channel are cut and synthesized to obtain cell fluorescence images of the real sample cells.

[0017] According to a cell discrimination model construction method provided by the present invention, determining the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, specifically includes:

[0018] Obtaining characteristic values ​​of cell characteristics and signal point characteristics of real sample cells in clinical samples;

[0019] Determine the points in the feature space of the vector consisting of the cell features and signal point features of each real sample cell of the target type and the classification probability of the corresponding real sample cell output by the cell classification model as the feature points corresponding to the corresponding real sample cell;

[0020] Based on the feature points corresponding to each real sample cell of the target type, new feature points are sampled and generated in the feature space, and based on the new feature points, feature values ​​of cell features and signal point features of the synthetic sample cells and the classification probability of the corresponding synthetic sample cells output by the cell classification model are determined;

[0021] The characteristic values ​​of the cell features and signal point features of the real sample cells and the synthetic sample cells, and the classification probabilities of the corresponding sample cells output by the cell classification model are determined as the characteristic values ​​of the candidate features.

[0022] According to a cell discrimination model construction method provided by the present invention, determining the cell features and signal point features of the sample cells as candidate features specifically includes:

[0023] Acquire cell images containing real sample cells of the target type from cell images of different clinical samples in different fields of view, and determine contour information of each real sample cell in the cell images containing the real sample cells of the target type;

[0024] Based on the contour information of the real sample cells of the target type, characteristic values ​​of the cell characteristics and signal point characteristics of the real sample cells of the target type are obtained;

[0025] Based on the contour information of each real sample cell in the cell image containing the real sample cells of the target type, randomly sampling the contour information of the real sample cells of the non-target type; wherein the difference in number between the sampled real sample cells of the non-target type and the real sample cells of the target type satisfies a preset condition;

[0026] Based on the contour information of the non-target type of real sample cells, characteristic values ​​of cell features and signal point features of the non-target type of real sample cells are obtained;

[0027] The feature values ​​of the cell features and signal point features of the target type real sample cells and the feature values ​​of the cell features and signal point features of the non-target type real sample cells are determined as the feature values ​​of the candidate features.

[0028] According to a cell discrimination model construction method provided by the present invention, the characteristic value of the cell characteristic of any real sample cell is determined from a cell image containing the any real sample cell based on the contour information of the any real sample cell;

[0029] The characteristic value of the signal point feature of any real sample cell is determined based on the contour information of the any real sample cell, by using a signal point detection model to detect the signal points corresponding to the any real sample cell area in the fluorescence image of each channel, and obtaining the signal point area of ​​the any real sample cell in the fluorescence image of each channel.

[0030] According to a cell discrimination model construction method provided by the present invention, the cell classification model and the machine learning model are integrated to obtain a cell discrimination model, which specifically includes:

[0031] A fusion layer is added after the cell classification model and the machine learning model to obtain the cell discrimination model; wherein the fusion layer is used to determine the average value or maximum value between the classification probabilities of the cells to be classified outputted by the cell classification model and the machine learning model respectively.

[0032] The present invention also provides a cell discrimination model construction device, comprising:

[0033] A cell classification model building unit is used to train a cell classification model based on a neural network based on the cell fluorescence images of sample cells and their cell labels;

[0034] A model fusion unit is used to determine the cell characteristics and signal point characteristics of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate characteristics, perform feature selection on the candidate characteristics, obtain the selected characteristics, and then construct a machine learning model based on the selected characteristics as a cell discrimination model; or to determine the cell characteristics and signal point characteristics of the sample cells as candidate characteristics, perform feature selection on the candidate characteristics, obtain the selected characteristics, and then construct a machine learning model based on the selected characteristics, and fuse the cell classification model and the machine learning model to obtain a cell discrimination model.

[0035] The present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any of the cell discrimination model construction methods described above is implemented.

[0036] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described cell discrimination model construction methods.

[0037] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned cell discrimination model construction methods.

[0038] The cell discrimination model construction method, device, electronic device and storage medium provided by the present invention are based on the cell fluorescence images of sample cells and their cell labels to train a cell classification model constructed based on a neural network. In addition to the cell classification model, a machine learning model is also constructed to extract the morphological and statistical features of cells and signal points through feature engineering, and use the morphological and statistical features to classify cells from another dimension to supplement the cell classification model. By fusing the machine learning model and the cell classification model, cells can be classified and discriminated from different angles using different features, thereby reducing the interference caused by background spots and impurities in the cell classification and discrimination process, improving the accuracy of cell classification and discrimination, and effectively reducing the number of manually interpreted cells while ensuring the recall rate, thereby improving the subsequent interpretation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] FIG1 is a schematic diagram of a process for constructing a cell discrimination model provided by the present invention;

[0041] FIG2 is a schematic diagram of the process of the cell classification model training method provided by the present invention;

[0042] FIG3 is a schematic diagram of a flow chart of a sample construction method provided by the present invention;

[0043] FIG4 is a second flow chart of the sample construction method provided by the present invention;

[0044] FIG5 is a schematic structural diagram of a cell discrimination model construction device provided by the present invention;

[0045] FIG6 is a schematic structural diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0047] FIG1 is a flow chart of a method for constructing a cell discrimination model provided by the present invention. As shown in FIG1 , the method includes:

[0048] Step 110 , training a cell classification model based on a neural network based on the cell fluorescence images of the sample cells and their cell labels;

[0049] Step 120, determining the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, performing feature selection on the candidate features, and after obtaining selected features, constructing a machine learning model based on the selected features as a cell discrimination model;

[0050] Alternatively, in step 130, the cell features and signal point features of the sample cells are determined as candidate features, feature selection is performed on the candidate features, and after obtaining the selected features, a machine learning model is constructed based on the selected features, and the cell classification model and the machine learning model are integrated to obtain a cell discrimination model.

[0051] Specifically, microscope scanning data of clinical samples (such as blood samples) are collected, and the initial cell classification data of the clinical samples are obtained through AI models. Among them, the current AI models include cell segmentation models (such as Mask-rcnn models) and signal point detection models (such as Yolo models), and the number of signal points of each cell under each channel is obtained, and the cells are initialized and classified based on this. The interpreter interprets the interpreted cells in the initial classification, confirms the true target type cells (such as CAC cells) in the sample, obtains the cell labels of each sample cell, and uses this to sample and divide the data to generate a sample set for constructing subsequent cell classification models and machine learning models. Among them, the sample sets used to construct cell classification models and machine learning models are generated independently.

[0052] Based on the cell fluorescence images and cell labels of each sample cell in the sample set corresponding to the cell classification model and their cell labels (i.e., labels indicating the cell type of the sample cell), the cell classification model constructed based on the neural network is trained. Among them, the cell fluorescence image of the sample cell is a superimposed composite image corresponding to the sample cell area in the fluorescence imaging of each channel. In some embodiments, the cell classification model can be implemented using a convolutional neural network, such as DenseNet-121. Taking DenseNet-121 as an example, in the cell classification model, the input image is first enhanced, including scaling and random cropping; then, the data enhancement result is processed based on a large-scale convolution and a pooling layer; then, after processing based on several consecutive sub-modules (i.e., Dense Block and Transition Layer), a pooling layer and a fully connected layer are used to output the corresponding classification results.

[0053] In some embodiments, as shown in FIG2 , the cell classification model constructed based on a neural network is trained based on the cell fluorescence images of the sample cells and their cell labels, specifically including:

[0054] Step 210 , obtaining cell fluorescence images and cell labels of different types of real sample cells in the clinical sample;

[0055] Step 220 , clustering the cell fluorescence images of the different types of real sample cells, and obtaining target cell clusters corresponding to the target type;

[0056] Step 230 , generating a target type cell composite image based on the cell fluorescence images of each real sample cell of the target type in the target cell cluster;

[0057] Step 240 : Based on the cell fluorescence image and its cell label and the cell composite image and its cell label, the cell classification model is trained.

[0058] Specifically, because a clinical sample contains only a few or a dozen cells of the target type (e.g., CAC cells), while non-target cells can number in the hundreds, there is a serious sample imbalance. This imbalance can affect the training of cell classification models and reduce classification accuracy in practical applications. To address this sample imbalance during cell classification model training, a sample set for the cell classification model can be generated using the KMeans SMOTE method.

[0059] Here, cell fluorescence images and cell labels of different types of real sample cells in clinical samples can be obtained. The cell fluorescence images of real sample cells in clinical samples are obtained by cutting and synthesizing the areas corresponding to the real sample cells in each channel fluorescence image based on the contour information of each real sample cell in the cell image obtained by the cell segmentation model. Both the cell image and the channel fluorescence image are microscope images collected by a microscope under different channels. In some embodiments, the cell image can be a microscope image under the Dapi channel, while the channel fluorescence images are microscope images under the Red, Green, Aqua, and Gold channels.

[0060] Based on the pixel values ​​of the pixel points in the cell fluorescence images of different types of real sample cells, the cell fluorescence images of different types of real sample cells are clustered to obtain multiple cell clusters, and the target cell cluster corresponding to the target type is obtained from them. Among them, according to the cell labels of the real sample cells in each cell cluster, the cell cluster with the most real sample cells of the target type can be used as the target cell cluster corresponding to the target type. Based on the cell fluorescence images of each real sample cell of the target type in the target cell cluster, a cell composite image of the target type is generated based on the oversampling technology. For example, based on the pixel values ​​of the same position in the cell fluorescence images of multiple real sample cells of the target type in the target cell cluster, a cell composite image of the target type can be obtained by fusion, and a cell label can be set for the cell composite image. A sample set is constructed based on the cell fluorescence images and cell labels of each real sample cell and the cell composite image and cell label generated in the above steps, so that a balanced sample set of each type of sample can be obtained, so that a cell classification model can be trained based on the sample set to improve the classification accuracy of the cell classification model.

[0061] To further improve the accuracy of cell classification, a machine learning model was constructed in addition to the cell classification model. This model uses feature engineering to extract the morphological and statistical characteristics of cells and signal points. These morphological and statistical characteristics are then used to classify cells from another dimension, supplementing the cell classification model. By integrating the machine learning model with the cell classification model, cells can be classified and identified from different perspectives using different features, thereby reducing the interference caused by background noise and impurities during the cell classification process and improving the accuracy of cell classification.

[0062] Specifically, there are two ways to integrate machine learning models and cell classification models:

[0063] (1) Determine the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, perform feature selection on the candidate features, and after obtaining the selected features, construct a machine learning model based on the selected features as a cell discrimination model.

[0064] Here, the cell features of the sample cells, the signal point features, and the classification probabilities of the corresponding sample cells output by the cell classification model are all used as candidate features. That is, the candidate features include the cell features, the signal point features, and the classification probabilities output by the cell classification model (i.e., the probabilities that the corresponding cells belong to each type). The cell features of the sample cells include the morphological and statistical characteristics of the cells, such as area, signal intensity, roundness, and elongation. The optional cell features may include multiple of the following features. The embodiment of the present invention does not specifically limit which features are used as cell features: 'nucleus_pixels_area' (nucleus pixel area), 'dapi_mean_intensity' (DAPI average intensity), 'dapi_intensity_std' (DAPI intensity standard deviation), 'dapi_total_intensity' (DAPI total intensity), 'nucleus_roundness' (nucleus roundness), 'nucleus_elongation' (nucleus elongation), 'nucleus_average_diameter' (nucleus average diameter), 'num_signals' (total number of signal points), 'gr_signals_total_intensity' (total intensity of Green signal points), 'gr_signals_total_area_pixels' (total area of ​​Green signal points), 'number_of_gr_signals' (number of Green signal points), 'gr_signal_intensity_to_nucleus_pixel_area' (Green signal intensity to nucleus pixel area), 'gr_signal_pixel_area_to_nucleus_pixel_area' (Green signal pixel area to nucleus pixel area), 'gr_signal_intensity_to_nucleus_intensity' (Green signal intensity to nucleus intensity), 'gr_signal_intensity_to_signal_pixel_area' (Green signal intensity to signal pixel area), 'background_intensity_of_gr_signals' (Green signal background intensity), 'r_signals_total_intensity' (Red signal total intensity), 'r_signals_total_area_pixels' (Red signal total area), 'number_of_r_signals' (Red signal number),'r_signal_intensity_to_nucleus_pixel_area' (Red signal intensity to nucleus pixel area), 'r_signal_pixel_area_to_nucleus_pixel_area' (Red signal pixel area to nucleus pixel area), 'r_signal_intensity_to_nucleus_intensity' (Red signal intensity to nucleus intensity), 'r_signal_intensity_to_signal_pixel_area' (Red signal intensity to signal pixel area), 'background_intensity_of_r_signals' (Background intensity of Red signal points), 'aq_signals_total_intensity' (Total intensity of Aqua signal points), 'aq_signals_total_area_pixels' (Total area of ​​Aqua signal points), 'number_of_aq_signals' (Number of Aqua signal points), 'aq_s signal_intensity_to_nucleus_pixel_area' (Aqua signal point intensity to nucleus pixel area), 'aq_signal_pixel_area_to_nucleus_pixel_area' (Aqua signal point pixel area to nucleus pixel area), 'aq_signal_intensity_to_nucleus_intensity' (Aqua signal intensity to nucleus intensity), 'aq_signal_intensity_to_signal_pixel_area' (Aqua signal point intensity to signal point pixel area), 'background_intensity_of_aq_signals' (Aqua signal point background intensity), 'gd_signals_total_intensity' (Gold signal point total intensity), 'gd_signals_total_area_pixels' (Gold signal point total area), 'number_of_gd_signals' (Number of Gold signal points), 'gd_signal_intensity_to_nucleus_pixel_area' (Gold signal point intensity to nucleus pixel area), 'gd_signal_pixel_area_to_nucleus_pixel_area' (Gold signal point pixel area to nucleus pixel area),'gd_signal_intensity_to_nucleus_intensity' (Gold signal intensity to nucleus intensity), 'gd_signal_intensity_to_signal_pixel_area' (Gold signal point intensity to signal point pixel area), 'background_intensity_of_gd_signals' (Gold signal point background intensity), 'gr_255_area_pixels' (Green pixel value is 2 55), 'r_255_area_pixels' (area of ​​Red pixel value 255), 'aq_255_area_pixels' (area of ​​Aqua pixel value 255), 'gd_255_area_pixels' (area of ​​Gold pixel value 255), 'gr_255_area_pixels_to_nucleus_pixel_area' (area of ​​Green pixel value 255 to nucleus pixel area), 'r_255_area_pixels_ to_nucleus_pixel_area' (the ratio of the area of ​​Red pixel value 255 to the pixel area of ​​the cell nucleus), 'aq_255_area_pixels_to_nucleus_pixel_area' (the ratio of the area of ​​Aqua pixel value 255 to the pixel area of ​​the cell nucleus), 'gd_255_area_pixels_to_nucleus_pixel_area' (the ratio of the area of ​​Gold pixel value 255 to the pixel area of ​​the cell nucleus), 'cell_score' (cell segmentation score output by the cell segmentation model), 'nucleus_perimeter' (cell perimeter), 'nucleus_aspect_ratio' (cell aspect ratio), 'nucleus_extent' (cell extent), 'nucleus_solidity' (cell solidity), 'nucleus_angle' (cell orientation), 'nucleus_ma_s' (cell short axis), 'nucleus_ma_l' (cell long axis), 'nucleus_sharpness' (cell sharpness).

[0065] The signal point features of the sample cell include the morphological and statistical features of the fluorescent signal points in each channel within the cell, such as the target detection score of the fluorescent signal points in each channel (that is, the probability of each fluorescent signal point output by the signal point detection model), signal point intensity, signal point area, signal point perimeter, etc., including: 'angle' (signal point direction), 'aspect_ratio' (signal point aspect ratio), 'bbox_h' (signal point detection bounding box height), 'bbox_w' (signal point detection bounding box width), 'extent' (signal point range), 'intensity' (signal intensity), 'intensity_std' (signal intensity standard deviation), 'ma_l' (signal point long axis), 'ma_s' (signal point short axis), 'max_value' (signal point intensity maximum), 'mean_intensity' (average signal intensity), 'min_value' (signal point intensity minimum), 'pixels_area' (signal pixel area), 'pixels_diameter' (signal point pixel diameter), 'score' (signal point detection score output by the signal point detection model), 'solidity' (signal point solidity); in addition, the signal point features can also include the distribution of the above-mentioned morphological features and statistical features (such as signal point intensity, signal point area, signal point perimeter, etc.) under each channel, such as minimum value, maximum value, median, mean value, and variance.

[0066] It should be noted that the characteristic value of the cell feature of any real sample cell is determined from the cell image containing the real sample cell based on the contour information of the real sample cell. The characteristic values ​​of some cell features need to be determined by combining the signal point detection model with the signal point detection to obtain the signal point area of ​​the real sample cell; the characteristic value of the signal point feature of any real sample cell is determined by detecting the signal points corresponding to the real sample cell area in the fluorescence image of each channel based on the contour information of the real sample cell using the signal point detection model, and obtaining the signal point area of ​​the real sample cell in the fluorescence image of each channel.

[0067] Subsequently, feature engineering is used to select the candidate features to obtain selected features. Feature engineering can perform feature selection based on the characteristic values ​​of each sample cell corresponding to each type of candidate feature, thereby selecting discriminative features as selected features. It can be seen that the selected features are some features among cell features, signal point features, and classification probabilities. An embedded feature selection method can be used for feature selection, such as a tree model. In some embodiments, a random forest model can be used for feature selection. After obtaining the selected features, a machine learning model can be constructed based on the selected features as a cell discrimination model. Specifically, a random forest model can be selected as the machine learning model, and model training is performed based on the characteristic values ​​of each selected feature corresponding to each sample cell. To prevent overfitting, each decision tree constructed in the random forest model is pre-pruned, i.e., the growth of the decision tree is stopped after reaching the set maximum depth. The machine learning model constructed in the above manner is the cell discrimination model. It can be seen that this fusion method incorporates the output results of the cell classification model into the construction process of the machine learning model.

[0068] Similar to the cell classification model, the machine learning model construction process also faces the problem of sample imbalance, which can also lead to insufficient classification accuracy. Therefore, to address the sample imbalance problem during machine learning model construction, the sample set for the machine learning model can be generated based on the SMOTE method.

[0069] In some embodiments, as shown in FIG3 , determining the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, specifically includes:

[0070] Step 310 , obtaining characteristic values ​​of cell characteristics and signal point characteristics of real sample cells in the clinical sample;

[0071] Step 320 , determining the points in the feature space of the vector consisting of the cell features and signal point features of each real sample cell of the target type and the classification probability of the corresponding real sample cell output by the cell classification model as the feature points corresponding to the corresponding real sample cells;

[0072] Step 330: Based on the feature points corresponding to each real sample cell of the target type, new feature points are sampled and generated in the feature space, and the feature values ​​of the cell features and signal point features of the synthetic sample cells and the classification probability of the corresponding synthetic sample cells output by the cell classification model are determined based on the new feature points.

[0073] Step 340 : Determine the feature values ​​of the cell features and signal point features of the real sample cells and the synthetic sample cells, and the classification probabilities of the corresponding sample cells output by the cell classification model as the feature values ​​of the candidate features.

[0074] Specifically, after obtaining the characteristic values ​​of various cell features and signal point features corresponding to each real sample cell in the clinical sample, a vector is constructed, which is composed of the characteristic values ​​of each cell feature and signal point feature corresponding to each real sample cell and the classification probability of the corresponding cell output by the cell classification model. The point in the feature space composed of the characteristic values ​​of the cell features and signal point features of each real sample cell of the target type and the classification probability of the corresponding cell output by the cell classification model is determined as the feature point corresponding to the corresponding real sample cell. Based on the characteristic points corresponding to each real sample cell of the target type, new feature points are sampled and generated in the feature space using oversampling technology. In this case, a feature point corresponding to a real sample cell of the target type can be arbitrarily selected, and the feature point corresponding to the real sample cell of the target type closest to the feature point is determined. The new feature point is sampled between the two feature points. Based on the vector corresponding to the sampled new feature points, the characteristic values ​​of the various cell features and signal point features corresponding to the synthetic sample cell and the classification probability of the corresponding synthetic sample cell output by the cell classification model are determined. The characteristic values ​​of the cell features and signal point features of the above-mentioned real sample cells and synthetic sample cells, as well as the classification probabilities of the corresponding sample cells output by the cell classification model can be used as characteristic values ​​of each candidate feature for feature selection and training of the machine learning model.

[0075] (2) Determine the cell features and signal point features of the sample cells as candidate features, perform feature selection on the candidate features, and after obtaining the selected features, construct a machine learning model based on the selected features, and fuse the cell classification model and the machine learning model to obtain a cell discrimination model.

[0076] Here, the cell features and signal point features of the sample cells are used as candidate features. That is, the candidate features include cell features and signal point features. Subsequently, feature engineering is used to perform feature selection on the candidate features to obtain selected features. Among them, the feature selection method is similar to the fusion method (1) and will not be repeated here. After obtaining the selected features, a machine learning model can be constructed based on the selected features, and the method of constructing the machine learning model based on the selected features is also similar to the fusion method (1) and will not be repeated here. After obtaining the machine learning model, the cell classification model and the machine learning model can be fused to obtain a cell discrimination model. In some embodiments, a fusion layer can be added after the cell classification model and the machine learning model to obtain a cell discrimination model. That is, the cell discrimination model includes a cell classification model, a machine learning model and a fusion layer. Among them, the fusion layer is used to determine the average value or maximum value between the classification probabilities of the cells to be classified outputted by the cell classification model and the machine learning model respectively. Specifically, the cell classification model and the machine learning model will classify the cells to be classified respectively, and obtain the classification probabilities of the cells to be classified output by the cell classification model and the machine learning model respectively. Then the fusion layer will calculate the average or maximum value between the above two classification probabilities as the classification result for the cells to be classified output by the cell discrimination model.

[0077] It should be noted that in order to solve the sample imbalance problem in the process of building the machine learning model, the method of generating the sample set of the machine learning model in the fusion method (2) is different from that of the cell classification model and the fusion method (1).

[0078] Specifically, as shown in FIG4 , determining the cell features and signal point features of the sample cells as candidate features specifically includes:

[0079] Step 410, acquiring cell images containing real sample cells of the target type from cell images of different clinical samples in different fields of view, and determining contour information of each real sample cell in the cell images containing the real sample cells of the target type;

[0080] Step 420 , based on the contour information of the real sample cells of the target type, obtain the characteristic values ​​of the cell characteristics and signal point characteristics of the real sample cells of the target type;

[0081] Step 430: Based on the contour information of each real sample cell in the cell image containing the real sample cells of the target type, randomly sample the contour information of real sample cells of non-target type; wherein the difference in number between the sampled real sample cells of non-target type and the real sample cells of target type satisfies a preset condition;

[0082] Step 440 , based on the contour information of the non-target type of real sample cells, obtain the characteristic values ​​of the cell characteristics and signal point characteristics of the non-target type of real sample cells;

[0083] Step 450 : Determine the feature values ​​of the cell features and signal point features of the target type real sample cells, and the feature values ​​of the cell features and signal point features of the non-target type real sample cells as the feature values ​​of the candidate features.

[0084] Here, cell images containing target-type real sample cells are obtained from cell images of different clinical samples under different fields of view, and the contour information of each real sample cell in the cell images containing the target-type real sample cells is determined based on the cell segmentation model. On the one hand, based on the contour information of the target-type real sample cells, the method provided in the above embodiments can be used to obtain the characteristic values ​​of the cell features and signal point features of the target-type real sample cells. On the other hand, considering that the number of non-target-type real sample cells is much larger than the number of target-type real sample cells, in order to compensate for the difference in number and obtain a more balanced sample set, the contour information of several non-target-type real sample cells can be randomly sampled based on the contour information of each real sample cell in the cell images containing the target-type real sample cells. Then, based on the contour information of the non-target-type real sample cells, the characteristic values ​​of the cell features and signal point features of the non-target-type real sample cells are obtained; wherein the difference in number between the sampled non-target-type real sample cells and the target-type real sample cells satisfies a preset condition. In some embodiments, the number of sampled non-target-type real sample cells is the same as the number of target-type real sample cells. Determine the characteristic values ​​of the cell features and signal point features of the real sample cells of the above-mentioned target type, as well as the characteristic values ​​of the cell features and signal point features of the real sample cells of the above-mentioned non-target type, as the characteristic values ​​of the candidate features for feature selection and training of the machine learning model.

[0085] It should be noted that both of the above-mentioned fusion methods can achieve the effect of improving the accuracy of cell classification and discrimination, and the embodiment of the present invention does not impose any limitation on the specific fusion method to be adopted.

[0086] It can be seen that the method provided in the embodiment of the present invention is based on the cell fluorescence images of sample cells and their cell labels to train a cell classification model constructed based on a neural network, and in addition to the cell classification model, a machine learning model is also constructed to extract the morphological and statistical features of cells and signal points through feature engineering, and use the morphological and statistical features to classify cells from another dimension to supplement the cell classification model; by fusing the machine learning model and the cell classification model, cells can be classified and judged from different angles using different features, thereby reducing the interference caused by background noise and impurities in the cell classification process, improving the accuracy of cell classification, and effectively reducing the number of manually judged cells while ensuring the recall rate, thereby improving the subsequent judgment efficiency.

[0087] The cell discrimination model construction device provided by the present invention is described below. The cell discrimination model construction device described below and the cell discrimination model construction method described above can be referenced to each other.

[0088] Based on any of the above embodiments, FIG5 is a schematic structural diagram of a cell discrimination model construction device provided by the present invention. As shown in FIG5 , the device includes: a cell classification model construction unit 510 and a model fusion unit 520 .

[0089] The cell classification model construction unit 510 is used to train a cell classification model based on a neural network based on the cell fluorescence images of the sample cells and their cell labels;

[0090] The model fusion unit 520 is used to determine the cell characteristics and signal point characteristics of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate characteristics, perform feature selection on the candidate characteristics, and after obtaining the selected characteristics, construct a machine learning model based on the selected characteristics as a cell discrimination model; or, to determine the cell characteristics and signal point characteristics of the sample cells as candidate characteristics, perform feature selection on the candidate characteristics, after obtaining the selected characteristics, construct a machine learning model based on the selected characteristics, and fuse the cell classification model and the machine learning model to obtain a cell discrimination model.

[0091] The device provided by the embodiment of the present invention is based on the cell fluorescence images of sample cells and their cell labels to train a cell classification model constructed based on a neural network. In addition to the cell classification model, a machine learning model is also constructed to extract the morphological and statistical features of cells and signal points through feature engineering, and use the morphological and statistical features to classify cells from another dimension to supplement the cell classification model; by fusing the machine learning model and the cell classification model, cells can be classified and judged from different angles using different features, thereby reducing the interference caused by background noise and impurities in the cell classification process, improving the accuracy of cell classification, and effectively reducing the number of manually judged cells while ensuring the recall rate, thereby improving the subsequent judgment efficiency.

[0092] Based on any of the above embodiments, the cell classification model constructed based on the training of the cell fluorescence images of the sample cells and their cell labels may specifically include:

[0093] Obtain cell fluorescence images and cell labels of different types of real sample cells in clinical samples;

[0094] Clustering the cell fluorescence images of the different types of real sample cells, and obtaining target cell clusters corresponding to the target type;

[0095] generating a synthetic image of cells of the target type based on the cell fluorescence images of each real sample cell of the target type in the target cell cluster;

[0096] The cell classification model is obtained by training based on the cell fluorescence image and its cell label and the cell composite image and its cell label.

[0097] Based on any of the above embodiments, the cell fluorescence image of the real sample cells is obtained based on the following steps:

[0098] Obtain the contour information of each real sample cell in the cell image based on the cell segmentation model;

[0099] Based on the contour information of the real sample cells, regions corresponding to the real sample cells in the fluorescence images of each channel are cut and synthesized to obtain cell fluorescence images of the real sample cells.

[0100] Based on any of the above embodiments, determining the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features specifically includes:

[0101] Obtaining characteristic values ​​of cell characteristics and signal point characteristics of real sample cells in clinical samples;

[0102] Determine the points in the feature space of the vector consisting of the cell features and signal point features of each real sample cell of the target type and the classification probability of the corresponding real sample cell output by the cell classification model as the feature points corresponding to the corresponding real sample cell;

[0103] Based on the feature points corresponding to each real sample cell of the target type, new feature points are sampled and generated in the feature space, and based on the new feature points, feature values ​​of cell features and signal point features of the synthetic sample cells and the classification probability of the corresponding synthetic sample cells output by the cell classification model are determined;

[0104] The characteristic values ​​of the cell features and signal point features of the real sample cells and the synthetic sample cells, and the classification probabilities of the corresponding sample cells output by the cell classification model are determined as the characteristic values ​​of the candidate features.

[0105] Based on any of the above embodiments, determining the cell features and signal point features of the sample cells as candidate features specifically includes:

[0106] Acquire cell images containing real sample cells of the target type from cell images of different clinical samples in different fields of view, and determine contour information of each real sample cell in the cell images containing the real sample cells of the target type;

[0107] Based on the contour information of the real sample cells of the target type, characteristic values ​​of the cell characteristics and signal point characteristics of the real sample cells of the target type are obtained;

[0108] Based on the contour information of each real sample cell in the cell image containing the real sample cells of the target type, randomly sampling the contour information of the real sample cells of the non-target type; wherein the difference in number between the sampled real sample cells of the non-target type and the real sample cells of the target type satisfies a preset condition;

[0109] Based on the contour information of the non-target type of real sample cells, characteristic values ​​of cell features and signal point features of the non-target type of real sample cells are obtained;

[0110] The feature values ​​of the cell features and signal point features of the target type real sample cells and the feature values ​​of the cell features and signal point features of the non-target type real sample cells are determined as the feature values ​​of the candidate features.

[0111] Based on any of the above embodiments, the feature value of the cell feature of any real sample cell is determined from a cell image containing the any real sample cell based on the contour information of the any real sample cell;

[0112] The characteristic value of the signal point feature of any real sample cell is determined based on the contour information of the any real sample cell, by using a signal point detection model to detect the signal points corresponding to the any real sample cell area in the fluorescence image of each channel, and obtaining the signal point area of ​​the any real sample cell in the fluorescence image of each channel.

[0113] Based on any of the above embodiments, the fusion of the cell classification model and the machine learning model to obtain a cell discrimination model specifically includes:

[0114] A fusion layer is added after the cell classification model and the machine learning model to obtain the cell discrimination model; wherein the fusion layer is used to determine the average value or maximum value between the classification probabilities of the cells to be classified outputted by the cell classification model and the machine learning model respectively.

[0115] FIG6 is a schematic diagram of the structure of an electronic device provided by the present invention. As shown in FIG6 , the electronic device may include: a processor 610, a memory 620, a communication interface 630, and a communication bus 640, wherein the processor 610, the memory 620, and the communication interface 630 communicate with each other via the communication bus 640. The processor 610 may call the logic instructions in the memory 620 to execute a cell discrimination model construction method, the method comprising: training a cell classification model based on a neural network based on the cell fluorescence image of the sample cell and its cell label; determining the cell features and signal point features of the sample cell, and the classification probability of the sample cell output by the cell classification model as candidate features, performing feature selection on the candidate features, obtaining the selected features, and constructing a machine learning model based on the selected features as a cell discrimination model; or determining the cell features and signal point features of the sample cell as candidate features, performing feature selection on the candidate features, obtaining the selected features, constructing a machine learning model based on the selected features, and fusing the cell classification model and the machine learning model to obtain a cell discrimination model.

[0116] In addition, the logic instructions in the above-mentioned memory 620 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0117] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the cell discrimination model construction method provided by the above methods, the method including: based on the cell fluorescence image of the sample cells and their cell labels, training a cell classification model constructed based on a neural network; determining the cell characteristics and signal point characteristics of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, performing feature selection on the candidate features, and after obtaining the selected features, constructing a machine learning model based on the selected features as a cell discrimination model; or, determining the cell characteristics and signal point characteristics of the sample cells as candidate features, performing feature selection on the candidate features, and after obtaining the selected features, constructing a machine learning model based on the selected features, and fusing the cell classification model and the machine learning model to obtain a cell discrimination model.

[0118] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the above-mentioned cell discrimination model construction method, the method comprising: training a cell classification model constructed based on a neural network based on the cell fluorescence image of the sample cells and their cell labels; determining the cell characteristics and signal point characteristics of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, performing feature selection on the candidate features, and after obtaining the selected features, constructing a machine learning model based on the selected features as a cell discrimination model; or, determining the cell characteristics and signal point characteristics of the sample cells as candidate features, performing feature selection on the candidate features, and after obtaining the selected features, constructing a machine learning model based on the selected features, and fusing the cell classification model and the machine learning model to obtain a cell discrimination model.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0120] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for constructing a cell discrimination model, characterized in that: include: Based on the cell fluorescence images of sample cells and their cell labels, a cell classification model based on a neural network is trained; Determine the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, perform feature selection on the candidate features, and after obtaining the selected features, construct a machine learning model based on the selected features as a cell discrimination model; Alternatively, the cell features and signal point features of the sample cells are determined as candidate features, and feature selection is performed on the candidate features. After the selected features are obtained, a machine learning model is constructed based on the selected features, and the cell classification model and the machine learning model are fused to obtain a cell discrimination model.

2. The method for constructing a cell discrimination model according to claim 1, characterized in that: The cell classification model based on the cell fluorescence image of the sample cells and the cell label thereof is trained to be constructed based on a neural network, specifically including: Obtain cell fluorescence images and cell labels of different types of real sample cells in clinical samples; Clustering the cell fluorescence images of the different types of real sample cells, and obtaining target cell clusters corresponding to the target types; Based on the cell fluorescence images of each real sample cell of the target type in the target cell cluster, generating a synthetic image of the cell of the target type; The cell classification model is obtained by training based on the cell fluorescence image and its cell label and the cell composite image and its cell label.

3. The method for constructing a cell discrimination model according to claim 2, characterized in that: The cell fluorescence image of the real sample cell is obtained based on the following steps: Based on the cell segmentation model, the contour information of each real sample cell in the cell image is obtained; Based on the contour information of the real sample cells, the regions corresponding to the real sample cells in the fluorescence images of each channel are cut and synthesized to obtain the cell fluorescence images of the real sample cells.

4. The method for constructing a cell discrimination model according to claim 1, characterized in that: The determining of the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features specifically includes: Obtaining characteristic values ​​of cell characteristics and signal point characteristics of real sample cells in clinical samples; Determine the point in the feature space of the vector composed of the cell features and signal point features of each real sample cell of the target type and the classification probability of the corresponding real sample cell output by the cell classification model as the feature point corresponding to the corresponding real sample cell; Based on the feature points corresponding to each real sample cell of the target type, new feature points are sampled and generated in the feature space, and based on the new feature points, feature values ​​of cell features and signal point features of synthetic sample cells and classification probabilities of corresponding synthetic sample cells output by the cell classification model are determined; The feature values ​​of the cell features and signal point features of the real sample cells and the synthetic sample cells, and the classification probabilities of the corresponding sample cells output by the cell classification model are determined as the feature values ​​of the candidate features.

5. The method for constructing a cell discrimination model according to claim 1, characterized in that: The step of determining the cell features and signal point features of the sample cells as candidate features specifically includes: Acquire cell images containing real sample cells of the target type from cell images of different clinical samples in different fields of view, and determine contour information of each real sample cell in the cell images containing real sample cells of the target type; Based on the contour information of the real sample cells of the target type, characteristic values ​​of the cell characteristics and signal point characteristics of the real sample cells of the target type are obtained; Based on the contour information of each real sample cell in the cell image containing the real sample cells of the target type, the contour information of the real sample cells of the non-target type is randomly sampled. Information; wherein the difference in number between the real sample cells of the non-target type obtained by sampling and the real sample cells of the target type meets the preset condition; Based on the contour information of the real sample cells of the non-target type, characteristic values ​​of the cell characteristics and signal point characteristics of the real sample cells of the non-target type are obtained; The feature values ​​of the cell features and signal point features of the real sample cells of the target type and the feature values ​​of the cell features and signal point features of the real sample cells of the non-target type are determined as the feature values ​​of the candidate features.

6. The method for constructing a cell discrimination model according to claim 4 or 5, characterized in that: The feature value of the cell feature of any real sample cell is determined from a cell image containing the any real sample cell based on the contour information of the any real sample cell; The characteristic value of the signal point feature of any real sample cell is determined based on the contour information of the any real sample cell, by using a signal point detection model to detect the signal points in the area corresponding to the any real sample cell in each channel fluorescence image, and obtaining the signal point area of ​​the any real sample cell in the fluorescence image of each channel.

7. The method for constructing a cell discrimination model according to claim 1, characterized in that: The fusion of the cell classification model and the machine learning model to obtain a cell discrimination model specifically includes: A fusion layer is added after the cell classification model and the machine learning model to obtain the cell discrimination model; wherein the fusion layer is used to determine the average value or maximum value between the classification probabilities of the cells to be classified respectively output by the cell classification model and the machine learning model.

8. A cell discrimination model construction device, characterized in that: include: A cell classification model building unit, used for training a cell classification model based on a neural network based on cell fluorescence images of sample cells and cell labels; A model fusion unit is used to determine the cell features and signal point features of the sample cells, and the classification probability of the sample cells output by the cell classification model as candidate features, Feature selection is performed on the candidate features, and after the selected features are obtained, a machine learning model is constructed based on the selected features as a cell discrimination model; or, cell features and signal point features of sample cells are determined as candidate features, feature selection is performed on the candidate features, and after the selected features are obtained, a machine learning model is constructed based on the selected features, and the cell classification model and the machine learning model are fused to obtain a cell discrimination model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the cell discrimination model construction method as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for constructing a cell discrimination model as claimed in any one of claims 1 to 7 is implemented.

Images