An organoid ATP analysis method and system based on multi-instance learning
Through the deep network model learned by multiple examples, the learning unit characteristics of the organoid ATP annotation data set are extracted, and the ATP analysis model is trained, which solves the low-precision problem under the influence of redundant objects, and achieves higher-precision ATP analysis and trend prediction.
Patent Information
- Application Number
- CN202410707272.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-03
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-06-03
AI Technical Summary
In the existing organoid ATP analysis methods, redundant objects have a greater impact on the prediction results, resulting in the training model having low ATP detection accuracy and being unable to effectively analyze the trend of ATP change.
Using a deep network model with multi-example learning, the ATP annotation data set is obtained, the learning unit is determined, and the features are extracted based on the convolutional neural network and attention mechanism are extracted, and the ATP analysis model is trained in combination with the full-connection network to improve the robustness of the model. and precision.
It improves the robustness and accuracy of the ATP analysis model, can more accurately analyze the ATP value in tumor organ images, and enhances the ability to predict the ATP change trend.
Smart Images

Figure CN118570170B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep learning technology, and in particular to an organoid ATP analysis method and system based on multi-example learning. Background Art
[0002] Adenosine triphosphate (ATP) is a high-energy phosphate compound and the most direct source of energy in living organisms. ATP can react with luciferase to release fluorescence. Based on this principle, the fluorescence reaction of the sample can be measured by bioluminescence, and this can be used as an assessment of cell viability. However, this method is an endpoint method, so it is impossible to continuously detect ATP to analyze its changing trend, which will cause inaccurate ATP measurement. By manually annotating samples and combining deep learning technology, automatic scoring of organoid viability can be achieved, and the predicted results are highly correlated with the ATP value of the bioluminescence method. Although there is great potential for using deep learning models to evaluate organoid viability values, there is still a large gap from the ATP measurement values in biology.
[0003] The invention patent with application number: CN202210868607.6 discloses an artificial intelligence-based organoid ATP analysis method and system, which includes: collecting images of organoid culture wells one by one to obtain a series of organoid images; splicing all organoid images located in the same well plate into a panoramic image, thereby obtaining a series of organoid panoramic images; using the CTG luminescence method to detect the organoid vitality value of the organoid culture well plate to obtain the vitality value of the organoid in each well plate; making the obtained series of organoid panoramic images and the vitality values of the organoids in all well plates into a training set of the organoid ATP detection model, and using the training set to train the organoid ATP detection model; using the trained organoid ATP detection model to perform ATP vitality value detection and analysis on the acquired organoid images. This method and system are used to monitor the organoid culture process, and data acquisition is easy and intuitive, without the need to lyse the organoids, and the analysis process is intelligent, high-precision, simple and efficient.
[0004] However, there are redundant objects in the series of organoid panoramas that interfere with ATP prediction. If they are directly used as training data, the redundant objects will have a greater impact on the prediction results, and the accuracy of the trained model in subsequent ATP detection will also be low.
[0005] In view of this, there is an urgent need for an organoid ATP analysis method and system based on multi-instance learning to at least address the above-mentioned deficiencies. Summary of the invention
[0006] One of the purposes of the present invention is to provide an organoid ATP analysis method and system based on multi-instance learning, determine the learning units in the acquired organoid ATP annotation data set, introduce a multi-instance learning deep network model to learn the characteristics of each learning unit, and train the ATP analysis model. The ATP analysis model performs ATP analysis on tumor organ images according to the input needs, determines the analysis results, improves the robustness of the ATP analysis model, and makes the analysis results more accurate.
[0007] An embodiment of the present invention provides an organoid ATP analysis method based on multi-instance learning, comprising:
[0008] Step 1: Obtain the organoid ATP annotation dataset;
[0009] Step 2: Determine the learning unit based on the organoid ATP annotation dataset;
[0010] Step 3: Based on the deep network model of multi-instance learning, determine the ATP analysis model according to the learning unit;
[0011] Step 4: Obtain target analysis image;
[0012] Step 5: Input the target analysis image into the ATP analysis model to obtain the analysis results.
[0013] Preferably, step 1: obtaining an organoid ATP annotation dataset comprises:
[0014] Collect tumor organoid images;
[0015] The ATP activity of tumor organoid images was detected and labeled based on the bioluminescence method to obtain an organoid ATP annotation dataset.
[0016] Preferably, step 2: determining the learning unit according to the organoid ATP annotation dataset comprises:
[0017] Obtaining example construction strategies in the learning unit; example construction strategies include: constructing examples based on rules and automatically constructing examples based on models;
[0018] Construct strategies based on examples and identify learning units.
[0019] An organoid ATP analysis method based on multi-instance learning provided by an embodiment of the present invention further includes:
[0020] Before determining the learning units, resampling of the organoid ATP annotation dataset was performed;
[0021] The learning unit is determined according to the resampled data obtained by resampling.
[0022] Preferably, step 3: determining the ATP analysis model based on the deep network model of multi-instance learning and the learning unit, including:
[0023] A feature extractor based on a convolutional neural network determines the target features according to the learning units;
[0024] Determine example weights based on the attention mechanism;
[0025] Determine the aggregated features based on the example weights and the target features;
[0026] Based on the regression network of the fully connected network, the ATP analysis model is trained according to the aggregated features.
[0027] Preferably, the ATP analysis model is trained based on the regression network of the fully connected network according to the aggregated features, including:
[0028] Get the network output of the regression network;
[0029] Re-encode the network output and obtain the re-encoded target vector;
[0030] According to the re-encoding process, the ATP encoding loss is determined; the ATP encoding loss includes: the binary cross entropy loss of the integer part and the first mean square error loss of the fractional part;
[0031] According to the predicted contributions of different vector dimensions of the target vector, the second mean square error loss after ATP normalization is obtained;
[0032] Determine the loss function based on the ATP coding loss and the second mean square error loss;
[0033] The loss function is optimized based on the Adam optimizer to obtain the ATP analysis model after optimization.
[0034] Preferably, the ATP activity of the tumor organoid image is detected and labeled based on the bioluminescence method to obtain the organoid ATP annotation dataset, including:
[0035] Obtain bioluminescence assay specifications;
[0036] Determine a standard behavior sequence based on bioluminescence detection specifications and tumor organoid images; the standard behavior sequence includes: multiple standard behavior processes;
[0037] Obtain the testing process data of the staff based on bioluminescence method;
[0038] According to the detection process data, obtain the detection behavior process;
[0039] Determine a matching feature vector pair according to a first behavior label of the detection behavior process and a second behavior label of the standard behavior process;
[0040] Determine the first vector angle of the matching feature vector pair;
[0041] Determine the first behavior label or the second behavior label corresponding to the matching feature vector pair, and use it as the third behavior label;
[0042] Get the angle fusion weight corresponding to the label in the third line, and determine the angle of the second vector according to the angle fusion weight and the angle of the first vector;
[0043] If the angle sum of the second vector angles is less than or equal to a preset target threshold, the organoid ATP annotation data set is obtained based on the detection results of the corresponding detection process data.
[0044] Preferably, obtaining the angle fusion weight corresponding to the third line label includes:
[0045] Get the fourth line as the historical detection behavior of the label;
[0046] Determine the detection records of historical detection behaviors, parse the detection records, and obtain the detection accuracy;
[0047] According to the correlation between historical detection behaviors and detection accuracy, the correlation degree of the correlation relationship is quantified to obtain a correlation degree value;
[0048] Determine the preselected angle fusion weight according to the correlation degree value and the preset angle fusion weight allocation interval;
[0049] The third line label and the fourth line label are matched, and a pre-selected angle fusion weight corresponding to the fourth line label that meets the label matching is determined as the angle fusion weight.
[0050] An organoid ATP analysis method based on multi-instance learning provided by an embodiment of the present invention further includes:
[0051] Step 6: Visualize the analysis results according to the analysis results and analysis purpose;
[0052] Among them, step 6: according to the analysis results and analysis purposes, the analysis results are visualized, including:
[0053] Analyze the purpose and obtain the display type, which includes static display and dynamic display;
[0054] If the display type is static display, the analysis results are displayed statically;
[0055] If the display type is dynamic display, the correlation analysis results of the associated images are obtained;
[0056] Based on the correlation analysis results and analysis results, predict the future development trend of ATP;
[0057] Generate trend reports based on future ATP development trends;
[0058] Parse trend reports and obtain report semantics;
[0059] According to the analysis purpose, the analysis conversation record is obtained, and the analysis model is trained based on the analysis conversation record;
[0060] Input the report semantics into the analysis model to obtain the output analysis semantics;
[0061] Determine the experimental guidance strategy based on the analysis semantics and the preset guidance strategy library;
[0062] The experimental guidance strategy is displayed in a preset blank display window at the trend end display point of the trend report.
[0063] An embodiment of the present invention provides an organoid ATP analysis system based on multi-instance learning, comprising:
[0064] The dataset acquisition subsystem is used to obtain the organoid ATP annotation dataset;
[0065] The learning unit determination subsystem is used to determine the learning unit based on the organoid ATP annotation dataset;
[0066] The multi-instance learning subsystem is used to determine the ATP analysis model based on the deep network model of multi-instance learning according to the learning unit;
[0067] A target analysis image acquisition subsystem is used to acquire target analysis images;
[0068] The analysis subsystem is used to input the target analysis image into the ATP analysis model to obtain the analysis results.
[0069] The beneficial effects of the present invention are:
[0070] The present invention determines the learning units in the acquired organoid ATP annotated data set, introduces a deep network model of multi-instance learning to learn the characteristics of each learning unit, and trains the ATP analysis model. The ATP analysis model performs ATP analysis on tumor organ images according to the input needs, determines the analysis results, improves the robustness of the ATP analysis model, and makes the analysis results more accurate.
[0071] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the present application documents.
[0072] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0074] Figure 1 Schematic diagram of an organoid ATP analysis method based on multi-instance learning in an embodiment of the present invention;
[0075] Figure 2 Schematic diagram of an organoid ATP analysis system based on multi-instance learning in an embodiment of the present invention. DETAILED DESCRIPTION
[0076] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0077] The embodiment of the present invention provides an organoid ATP analysis method based on multi-instance learning, such as Figure 1 As shown, including:
[0078] Step 1: Obtain an organoid ATP annotation dataset; wherein the organoid ATP annotation dataset includes a plurality of organoid images with measured ATP values, and the tumor organs in the organoid images are in a culture well plate;
[0079] Step 2: Determine the learning unit based on the organoid ATP annotation dataset; the learning unit is: the image of each well in the culture well plate;
[0080] Step 3: Based on the deep network model of multi-instance learning, the ATP analysis model is determined according to the learning unit; the deep network model is based on the Pytorch deep learning framework, and the model is optimized on the Nvi dia GTX 3090 GPU using the Adam optimizer, with the batch size set to 15, the learning rate set to 0.002, and the decay rate set to 0.1 every 10 rounds. A total of 200 epochs are trained, and all training data need to be preprocessed, only the images of the culture well area are retained, and then the captured image blocks are resampled to a specific size, and the normalized images are used as network input;
[0081] Step 4: Obtain target analysis images; wherein the target analysis images are: images of tumor organs that require ATP analysis;
[0082] Step 5: Input the target analysis image into the ATP analysis model to obtain the analysis result, wherein the analysis result is: the ATP value of the organ in each culture well plate of the target analysis image.
[0083] The working principle and beneficial effects of the above technical solution are:
[0084] This application determines the learning units in the acquired organoid ATP annotation data set, introduces a deep network model of multi-instance learning to learn the characteristics of each learning unit, and trains the ATP analysis model. The ATP analysis model performs ATP analysis on tumor organ images according to the input needs and determines the analysis results, thereby improving the robustness of the ATP analysis model and making the analysis results more accurate.
[0085] In one embodiment, step 1: obtaining an organoid ATP annotation dataset comprises:
[0086] Collect tumor organoid images; the tumor organoid images are: Cytat i on 5 image acquisition results of tumor organoids in 384-well plates, the image magnification is 40 times, each well is collected once, no stitching is required, and images are collected at 13 different focal length settings, each focal length interval is 54.8m, the image resolution is 1992×1992, and the coverage area is about 12mm 2 ;
[0087] The ATP activity of tumor organoid images was detected and labeled based on the bioluminescence method to obtain an organoid ATP annotation dataset.
[0088] The working principle and beneficial effects of the above technical solution are:
[0089] This application collects tumor organoid images at different focal lengths, and labels the ATP activity of training samples based on the bioluminescence method to obtain an organoid ATP annotation dataset, thereby improving the comprehensiveness and rationality of the construction of the organoid ATP annotation dataset.
[0090] In one embodiment, step 2: determining a learning unit according to the organoid ATP annotation dataset comprises:
[0091] Obtaining example construction strategies in the learning unit; example construction strategies include: constructing examples based on rules and automatically constructing examples based on models;
[0092] Construct strategies based on examples and identify learning units.
[0093] The working principle and beneficial effects of the above technical solution are:
[0094] This application introduces two example construction methods to construct examples in learning units. The method of constructing examples according to rules usually requires defining examples in advance, such as dividing an image into image blocks, each image block as an example, or treating each channel in a multi-channel image as an example. The model automatically constructs examples without manually defining example generation rules. It constructs examples through self-learning, which improves the flexibility of example construction.
[0095] The embodiment of the present invention provides an organoid ATP analysis method based on multi-instance learning, further comprising:
[0096] Before determining the learning units, resampling of the organoid ATP annotation dataset was performed;
[0097] The learning unit is determined according to the resampled data obtained by resampling.
[0098] The working principle and beneficial effects of the above technical solution are:
[0099] Since the sample sizes distributed in different ATP intervals vary greatly, direct optimization will make it difficult to fit the model to samples sparsely distributed in a specific interval. Therefore, the organoid ATP annotated dataset is resampled, taking the maximum value of ATP as ATPmax as an example. The detailed process is as follows: (1) Set the interval width Rb in, and the number of intervals divided is ATPmax / / Rb in, / / is used as an integer division operator, for example: the result of 7 / / 3 is 2. (2) Set the model training batch to B, then uniformly sample B / / (ATPmax / / Rb in) samples from each interval. (3) Combine the samples in each interval to form the samples for each iterative training. After sample resampling, the number of samples participating in the training is basically consistent in each interval, and the number of samples distributed in each interval in each training batch remains balanced, which improves the comprehensiveness of the training.
[0100] In one embodiment, step 3: determining an ATP analysis model based on a deep network model of multi-instance learning and a learning unit, includes:
[0101] A feature extractor based on a convolutional neural network determines target features according to learning units; wherein the target features are features useful for ATP estimation, such as cell morphology features, cell structure features, and texture features;
[0102] Based on the attention mechanism, determine the example weight; where the example weight represents the contribution of the example to the package;
[0103] Determine an aggregate feature according to the example weight and the target feature; wherein the aggregate feature is: an aggregate result of a result obtained by multiplying the example weight and the target feature;
[0104] Based on the regression network of the fully connected network, the ATP analysis model is trained according to the aggregated features.
[0105] The working principle and beneficial effects of the above technical solution are:
[0106] This application introduces a feature extractor of a convolutional neural network to extract target features in a learning unit. An attention mechanism is introduced to determine the contribution of an example to a package and quantify the example weight. According to the example weight, the weighted target features are aggregated to obtain aggregate features. The ATP analysis model is trained based on the aggregate features, which improves the pertinence of the ATP analysis model training.
[0107] In one embodiment, based on a regression network of a fully connected network, an ATP analysis model is trained according to aggregated features, including:
[0108] Obtain the network output of the regression network; wherein the network output is: the output data during the regression network training process;
[0109] Re-encode the network output to obtain the re-encoded target vector; when re-encoding the network output, divide the range of 0-400000 into multiple intervals according to the manually set interval width, and then divide the real ATP by the interval width, and encode the obtained integer part in binary to obtain the integer bit representation of ATP, and the decimal part is used as the decimal bit representation of ATP. For example, taking the sample with ATP as 239420 as an example, the interval width is set to 20000, and the intermediate result is 239420 / 20000=11.971. Take the integer part 11 and convert it to binary representation 1011, the decimal part is 0.971, and the encoded value is represented by a 5-dimensional vector {1,0,1,1,0.971};
[0110] According to the re-encoding process, the ATP encoding loss is determined; the ATP encoding loss includes: the binary cross entropy loss of the integer part and the first mean square error loss of the fractional part; wherein the first mean square error loss is: the mean square error loss of the fractional part during ATP encoding;
[0111] According to the predicted contributions of different vector dimensions of the target vector, the second mean square error loss after ATP normalization is obtained; wherein the second mean square error loss is: the mean square error loss after ATP normalization;
[0112] Determine a loss function according to the ATP coding loss and the second mean square error loss; wherein, when determining the loss function, the ATP coding loss and the second mean square error loss are added;
[0113] The loss function is optimized based on the Adam optimizer to obtain the ATP analysis model after optimization.
[0114] The working principle and beneficial effects of the above technical solution are:
[0115] Since the output range of the network is 0-400000, direct optimization of ATP will lead to instability in the training process. Therefore, this application obtains the network output of the regression network, re-encodes the network output, and determines the re-encoded target vector. According to the re-encoding process, the ATP encoding loss and the ATP normalization loss are determined. The loss function is calculated based on the ATP encoding loss and the ATP normalization loss, and then the loss function is optimized using the Adam optimizer to obtain the ATP analysis model, thereby improving the stability of the training process.
[0116] In one embodiment, the ATP activity of tumor organoid images is detected and labeled based on bioluminescence to obtain an organoid ATP annotation dataset, including:
[0117] Obtain bioluminescence detection specifications; wherein the bioluminescence detection specifications are: standard bioluminescence detection specification process;
[0118] Determine a standard behavior sequence based on bioluminescence detection specifications and tumor organoid images; the standard behavior sequence includes: multiple standard behavior processes; the standard behavior process is the result of dividing all behavior processes according to process types, and the process types include: sample preparation, detection condition setting, and data recording;
[0119] Acquire the testing process data of the staff performing the test based on the bioluminescence method; wherein the testing process data is: the testing management record of the staff performing the test based on the bioluminescence method, such as: the monitoring video of the laboratory during the testing process;
[0120] According to the detection process data, the detection behavior process is obtained; wherein the detection behavior process is: a process obtained by sorting the detection behaviors of the staff according to the order of detection time;
[0121] Determine a matching feature vector pair according to the first behavior label of the detection behavior process and the second behavior label of the standard behavior process; wherein the matching feature vector pair includes: when the labels of the first behavior label and the second behavior label match each other, a vector pair consisting of a first behavior feature description vector constructed based on the detection behavior process and a second behavior feature description vector constructed based on the standard behavior process respectively;
[0122] Determine a first vector angle of the matching feature vector pair; wherein the first vector angle is: a vector angle formed by the first line feature description vector and the second line feature description vector;
[0123] Determine the first behavior label or the second behavior label corresponding to the matching feature vector pair, and use it as the third behavior label;
[0124] Get the angle fusion weight corresponding to the third line label, and determine the second vector angle according to the angle fusion weight and the first vector angle; where the angle fusion weight is: angle conversion coefficient, the value range of the angle fusion weight is [1,1.5]. The greater the influence of the behavior corresponding to the third line label on the detection result, the greater the corresponding angle fusion weight. When determining the second vector angle, the angle fusion weight is multiplied by the first vector angle.
[0125] If the angle sum of the second vector angles is less than or equal to a preset target threshold, then based on the detection results of the corresponding detection process data, an organoid ATP annotation data set is obtained, wherein the preset target threshold is manually preset.
[0126] The working principle and beneficial effects of the above technical solution are:
[0127] The accuracy of the annotation of the organoid ATP annotation data set greatly affects the training quality of the final analysis model. Therefore, it is extremely important to obtain an accurately annotated organoid ATP annotation data set. This application introduces bioluminescence detection specifications to determine a standardized and standard standard behavior sequence. The standard behavior sequence includes multiple standard behavior processes, such as: the behavior process of sample preparation, the behavior process of detection condition setting, and the behavior process of data recording. The detection process data of the staff based on the bioluminescence method is obtained, and the detection behavior process in which the detection behavior is sorted in the order of detection time is determined according to the detection process data. The first behavior label of the detection behavior process and the second behavior label of the standard behavior process are matched. When the labels of the first behavior label and the second behavior label are consistent, the first behavior feature description vector constructed based on the detection behavior process and the second behavior feature description vector constructed based on the standard behavior process constitute a matching feature vector pair. Determine the first vector angle of the first behavior feature description vector and the second behavior feature description vector in the matching feature vector pair, determine the corresponding third behavior label of the matching feature vector pair, and obtain the angle fusion weight corresponding to the third behavior label. Multiply the angle fusion weight and the first vector angle to obtain the second vector angle. The sum of the angles of the second vectors is calculated to obtain the angle sum value. If the angle sum value is less than or equal to the preset target threshold, the organoid ATP annotation data set is determined based on the detection results corresponding to the corresponding detection process data, and the quality of the training data is higher.
[0128] In one embodiment, obtaining the angle fusion weight corresponding to the third line label includes:
[0129] Obtaining a historical detection behavior with a fourth behavior label; wherein the fourth behavior label is: a behavior type label obtained by classifying all historical detection behaviors;
[0130] Determine the detection record of the historical detection behavior, analyze the detection record, and obtain the detection accuracy; wherein the detection record is: the bioluminescence detection record corresponding to the historical detection behavior;
[0131] According to the correlation between the historical detection behavior and the detection accuracy, the correlation degree of the correlation relationship is quantified to obtain a correlation degree value; wherein the correlation degree value indicates the influence degree of the historical detection behavior of the fourth line label on the detection accuracy;
[0132] According to the correlation degree value and the preset angle fusion weight allocation interval, the pre-selected angle fusion weight is determined; wherein, the preset angle fusion weight allocation interval is [1,1.5]; when determining the pre-selected angle fusion weight, the pre-selected angle fusion weight of the minimum correlation degree value is set to 1, and the pre-selected angle fusion weight of the maximum correlation degree value is set to 1.5, and then the wide correlation degree value interval composed of the maximum correlation degree value and the minimum correlation degree value is proportionally mapped on the angle fusion weight allocation interval. After the mapping is completed, the pre-selected angle fusion weight corresponding to each correlation degree value is determined according to the mapping point of the correlation degree value;
[0133] The third line label and the fourth line label are matched, and a pre-selected angle fusion weight corresponding to the fourth line label that meets the label matching is determined as the angle fusion weight.
[0134] The working principle and beneficial effects of the above technical solution are:
[0135] This application classifies all historical detection behaviors to obtain fourth behavior labels. At the same time, it determines the detection records that contain historical detection behaviors, and parses the detection records to obtain detection accuracy. The correlation between historical detection behaviors and detection accuracy is quantified to determine the correlation degree value. The correlation degree value is mapped to the angle fusion weight allocation interval to determine the pre-selected angle fusion weight. The third behavior label and the fourth behavior label are matched, and the pre-selected angle fusion weight of the matching fourth behavior label is determined as the angle fusion weight, which improves the accuracy of the angle fusion weight determination.
[0136] The embodiment of the present invention provides an organoid ATP analysis method based on multi-instance learning, further comprising:
[0137] Step 6: Visualize the analysis results according to the analysis results and analysis purpose; the analysis purpose is: the application purpose of the analysis results, such as: tumor microenvironment simulation, tumor drug resistance research, etc.;
[0138] Among them, step 6: according to the analysis results and analysis purposes, the analysis results are visualized, including:
[0139] Analyze the analysis purpose and obtain the display type, which includes static display and dynamic display. The static display is to directly mark and display the ATP value obtained by the analysis.
[0140] If the display type is static display, the analysis results are displayed statically;
[0141] If the display type is dynamic display, the associated analysis results of the associated images are obtained; wherein the associated images are: analysis images with the same analysis purpose within a preset time; the associated analysis results are: ATP estimation results of the associated images;
[0142] According to the correlation analysis results and the analysis results, the future ATP development trend is predicted; wherein the future ATP development trend is obtained by fitting characteristic curves based on the changes of experimental variables of the analysis images and ATP estimation results;
[0143] Generate a trend report based on the future ATP development trend; the trend report is: a visual display report of the future ATP development trend;
[0144] Parse the trend report to obtain the report semantics; wherein the report semantics is: the descriptive semantics of the report;
[0145] According to the purpose of analysis, the analysis conversation record is obtained, and the analysis model is trained based on the analysis conversation record; wherein the analysis conversation record is: the conversation record when the corresponding experimental analysis of the analysis purpose is manually performed; the analysis model is: an intelligent model that automatically determines the problem semantics of the experiment based on the extracted report semantics;
[0146] Input the report semantics into the analysis model to obtain the output analysis semantics; where the analysis semantics is: the semantics of what kind of problems exist in the experiment;
[0147] Determine the experimental guidance strategy based on the analysis semantics and the preset guidance strategy library; the preset guidance strategy library includes: multiple experimental problem semantics and corresponding guidance strategies; the experimental guidance strategy is: based on the analysis purpose, in which direction to improve the experiment, such as: how to adjust the experimental variables;
[0148] The experimental guidance strategy is displayed in a preset blank display window at the trend end display point of the trend report.
[0149] The working principle and beneficial effects of the above technical solution are:
[0150] The analysis results generally serve specific medical applications, such as tumor microenvironment simulation, tumor drug resistance research, etc. In order to enhance intuitiveness, the analysis results are visualized. There are two situations in the display process. The first is that the tester only needs to view the immediate analysis results; the second is that the tester implements a series of tests, studies the ATP changes by adjusting certain variables, and manually records and compares the test records, which is very cumbersome, and there are also cases of misremembering and omissions. Therefore, the association analysis results of the associated images are introduced, and the future ATP development trend is predicted based on the association analysis results and analysis results, and a trend report is generated. The semantic analysis technology is introduced to obtain the report semantics. At the same time, the analysis model is trained according to the analysis dialogue records corresponding to the analysis purpose, and the report semantics are input into the analysis model to determine the analysis semantics. Then, according to the guidance strategy library and the analysis semantics, the experimental guidance strategy is determined and displayed in the preset blank display window at the end display point of the trend, guiding the staff to make subsequent adjustments to the experiment, which is more intelligent.
[0151] The embodiment of the present invention provides an organoid ATP analysis system based on multi-instance learning, such as Figure 2 As shown, including:
[0152] Dataset acquisition subsystem 1, used to acquire organoid ATP annotation dataset;
[0153] Learning unit determination subsystem 2, used to determine the learning unit based on the organoid ATP annotation dataset;
[0154] A multi-instance learning subsystem 3 is used for determining an ATP analysis model based on a deep network model of multi-instance learning according to a learning unit;
[0155] The target analysis image acquisition subsystem 4 is used to acquire the target analysis image;
[0156] The analysis subsystem 5 is used to input the target analysis image into the ATP analysis model to obtain the analysis result.
[0157] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A method for analyzing ATP in organoids based on multi-instance learning, characterized in that: include: Step 1: Obtain the organoid ATP annotation dataset; Step 2: Determine the learning unit based on the organoid ATP annotation dataset; Step 3: Based on the deep network model of multi-instance learning, determine the ATP analysis model according to the learning unit; Step 4: Obtain target analysis image; Step 5: Input the target analysis image into the ATP analysis model to obtain the analysis results; Step 1: Obtain the organoid ATP annotation dataset, including: Collect tumor organoid images; The ATP activity of tumor organoid images was detected and labeled based on the bioluminescence method to obtain an organoid ATP annotation dataset. Based on the bioluminescence method, ATP activity of tumor organoid images was detected and labeled to obtain an organoid ATP annotation dataset, including: Obtain bioluminescence assay specifications; Determine a standard behavior sequence based on bioluminescence detection specifications and tumor organoid images; the standard behavior sequence includes: multiple standard behavior processes; Obtain the testing process data of the staff based on bioluminescence method; According to the detection process data, obtain the detection behavior process; Determine a matching feature vector pair according to a first behavior label of the detection behavior process and a second behavior label of the standard behavior process; Determine the first vector angle of the matching feature vector pair; Determine the first behavior label or the second behavior label corresponding to the matching feature vector pair, and use it as the third behavior label; Get the angle fusion weight corresponding to the label in the third line, and determine the angle of the second vector according to the angle fusion weight and the angle of the first vector; If the angle sum of the second vector angles is less than or equal to a preset target threshold, then based on the detection results of the corresponding detection process data, an organoid ATP annotation data set is obtained; Get the angle fusion weight corresponding to the third line label, including: Get the fourth line as the historical detection behavior of the label; Determine the detection records of historical detection behaviors, parse the detection records, and obtain the detection accuracy; According to the correlation between historical detection behaviors and detection accuracy, the correlation degree of the correlation relationship is quantified to obtain a correlation degree value; Determine the preselected angle fusion weight according to the correlation degree value and the preset angle fusion weight allocation interval; Perform label matching on the third line label and the fourth line label, and determine the pre-selected angle fusion weight corresponding to the fourth line label that meets the label matching as the angle fusion weight; Step 6: Visualize the analysis results according to the analysis results and analysis purpose; Among them, step 6: according to the analysis results and analysis purposes, the analysis results are visualized, including: Analyze the purpose and obtain the display type, which includes static display and dynamic display; If the display type is static display, the analysis results are displayed statically; If the display type is dynamic display, the correlation analysis results of the associated images are obtained; Based on the correlation analysis results and analysis results, predict the future development trend of ATP; Generate trend reports based on future ATP development trends; Parse trend reports and obtain report semantics; According to the analysis purpose, the analysis conversation record is obtained, and the analysis model is trained based on the analysis conversation record; Input the report semantics into the analysis model to obtain the output analysis semantics; Determine the experimental guidance strategy based on the analysis semantics and the preset guidance strategy library; The experimental guidance strategy is displayed in a preset blank display window at the trend end display point of the trend report.
2. The organoid ATP analysis method based on multi-instance learning according to claim 1, characterized in that: Step 2: Determine the learning unit based on the organoid ATP annotation dataset, including: Obtaining example construction strategies in the learning unit; example construction strategies include: constructing examples based on rules and automatically constructing examples based on models; Construct strategies based on examples and identify learning units.
3. The organoid ATP analysis method based on multi-instance learning as claimed in claim 2, characterized in that: Also includes: Before determining the learning units, resampling of the organoid ATP annotation dataset was performed; The learning unit is determined according to the resampled data obtained by resampling.
4. The organoid ATP analysis method based on multi-instance learning according to claim 1, characterized in that: Step 3: Based on the deep network model of multi-instance learning, determine the ATP analysis model according to the learning unit, including: A feature extractor based on a convolutional neural network determines the target features according to the learning units; Determine example weights based on the attention mechanism; Determine the aggregated features based on the example weights and the target features; Based on the regression network of the fully connected network, the ATP analysis model is trained according to the aggregated features.
5. The organoid ATP analysis method based on multi-instance learning according to claim 4, characterized in that: Based on the regression network of the fully connected network, the ATP analysis model is trained according to the aggregated features, including: Get the network output of the regression network; Re-encode the network output and obtain the re-encoded target vector; According to the re-encoding process, the ATP encoding loss is determined; the ATP encoding loss includes: the binary cross entropy loss of the integer part and the first mean square error loss of the fractional part; According to the predicted contributions of different vector dimensions of the target vector, the second mean square error loss after ATP normalization is obtained; Determine the loss function based on the ATP coding loss and the second mean square error loss; The loss function is optimized based on the Adam optimizer to obtain the ATP analysis model after optimization.
6. An organoid ATP analysis system based on multi-instance learning, characterized in that: include: The dataset acquisition subsystem is used to obtain the organoid ATP annotation dataset; The learning unit determination subsystem is used to determine the learning unit based on the organoid ATP annotation dataset; The multi-instance learning subsystem is used to determine the ATP analysis model based on the deep network model of multi-instance learning according to the learning unit; A target analysis image acquisition subsystem is used to acquire target analysis images; The analysis subsystem is used to input the target analysis image into the ATP analysis model to obtain the analysis result; The dataset acquisition subsystem acquires the organoid ATP annotation dataset and performs the following operations: Collect tumor organoid images; The ATP activity of tumor organoid images was detected and labeled based on the bioluminescence method to obtain an organoid ATP annotation dataset. Based on the bioluminescence method, ATP activity of tumor organoid images was detected and labeled to obtain an organoid ATP annotation dataset, including: Obtain bioluminescence assay specifications; Determine a standard behavior sequence based on bioluminescence detection specifications and tumor organoid images; the standard behavior sequence includes: multiple standard behavior processes; Obtain the testing process data of the staff based on bioluminescence method; According to the detection process data, obtain the detection behavior process; Determine a matching feature vector pair according to a first behavior label of the detection behavior process and a second behavior label of the standard behavior process; Determine the first vector angle of the matching feature vector pair; Determine the first behavior label or the second behavior label corresponding to the matching feature vector pair, and use it as the third behavior label; Get the angle fusion weight corresponding to the label in the third line, and determine the angle of the second vector according to the angle fusion weight and the angle of the first vector; If the angle sum of the second vector angles is less than or equal to a preset target threshold, then based on the detection results of the corresponding detection process data, an organoid ATP annotation data set is obtained; Get the angle fusion weight corresponding to the third line label, including: Get the fourth line as the historical detection behavior of the label; Determine the detection records of historical detection behaviors, parse the detection records, and obtain the detection accuracy; According to the correlation between historical detection behaviors and detection accuracy, the correlation degree of the correlation relationship is quantified to obtain a correlation degree value; Determine the preselected angle fusion weight according to the correlation degree value and the preset angle fusion weight allocation interval; Perform label matching on the third line label and the fourth line label, and determine the pre-selected angle fusion weight corresponding to the fourth line label that meets the label matching as the angle fusion weight; An organoid ATP analysis system based on multi-instance learning also performs the following operations; Visualize the analysis results according to the analysis results and analysis purpose; in, According to the analysis results and analysis purpose, the analysis results are visualized, including: Analyze the purpose and obtain the display type, which includes static display and dynamic display; If the display type is static display, the analysis results are displayed statically; If the display type is dynamic display, the correlation analysis results of the associated images are obtained; Based on the correlation analysis results and analysis results, predict the future development trend of ATP; Generate trend reports based on future ATP development trends; Parse trend reports and obtain report semantics; According to the analysis purpose, the analysis conversation record is obtained, and the analysis model is trained based on the analysis conversation record; Input the report semantics into the analysis model to obtain the output analysis semantics; Determine the experimental guidance strategy based on the analysis semantics and the preset guidance strategy library; The experimental guidance strategy is displayed in a preset blank display window at the trend end display point of the trend report.
