A method and system for classifying blood samples in coagulation index detection based on MAML and spatial transformation

Abstract

The present application relates to the field of machine learning, in particular to a blood sample classification method and system for coagulation index detection based on MAML and spatial transformation, mainly using a spatial transformation module to classify coagulation indexes, the specific steps comprising: obtaining the curve images of prothrombin time (PT) and thrombin time (TT) samples, extracting features and creating a training set and a test set; inputting the training set images into a module combining spatial transformation and convolution network to generate processed images; then updating parameters using gradient descent of the training model, reconstructing a spatial transformation coagulation index classification model; setting hyperparameters for the abnormal coagulation index classification model; finding the optimal model training process through model training and iterative testing to realize the classification of coagulation indexes. The present application uses MAML algorithm and spatial transformation to gradually optimize a small number of samples to achieve good generalization ability, and combines neural network structure and learning algorithm, which has wide applicability and flexibility.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a blood sample classification method and system for coagulation index detection based on MAML and spatial transformation. BACKGROUND

[0002] With the development of medical science, the understanding of hemostasis and thrombosis and their development is becoming more and more in-depth. Hemostasis and thrombosis are closely related to the occurrence and treatment prognosis of many clinical diseases. Coagulation detection is the main application of thrombosis and hemostasis function in vitro diagnosis. Recent research and clinical practice shows that prothrombin time (PT) and activated partial thromboplastin time (APTT) have practical significance for the evaluation of coagulation function of preoperative patients.

[0003] PT refers to prothrombin time, which plays an important role in coagulation instruments. Prothrombin time is a commonly used coagulation function index, which is used to evaluate the extrinsic coagulation function of the coagulation system. By measuring the time of prothrombin converting to thrombin in blood samples, it can detect whether coagulation factors VII, V, X and prothrombin are normal. PT is often used to screen for bleeding diseases, evaluate liver function, monitor the effectiveness of anticoagulant therapy (such as warfarin therapy), etc.

[0004] APTT is a commonly used coagulation function index in coagulation instruments. It is used to evaluate the intrinsic coagulation function of the coagulation system. By measuring the time of prothrombin being activated to thrombin in blood samples, it can detect whether coagulation factors VIII, IX, XI, XII and the normal coagulation pathway are normal. APTT is often used to determine the cause of coagulopathy, screen for genetic coagulation factor deficiency or abnormality, evaluate the effect of anticoagulant therapy (such as heparin therapy), etc.

[0005] Thrombin time is a test method used to evaluate the function of the coagulation system. It refers to the time of observing the coagulation of plasma or blood samples after adding thrombin (an enzyme that promotes blood clotting) under specific experimental conditions. Thrombin time can be used to detect abnormal function of coagulation factors, abnormal liver function, effect of anticoagulants, and diagnosis and monitoring of some coagulation diseases.

[0006] Fibrin (FIB) is a plasma protein synthesized by the liver, also known as coagulation factor I. It is one of the key proteins in the process of blood clotting. When blood vessels are damaged, the coagulation system activates a series of enzyme reactions, in which fibrin is converted to fibrinogen. Through the action of enzymes, fibrinogen is polymerized into fibrin, forming a blood clot, and thus promoting the process of hemostasis. Fibrin plays an important role in blood clotting, hemostasis and repair of damaged tissues.

[0007] The purpose of MAML is to realize fast learning, and the key to realizing fast learning is that the gradient descent of the neural network model is accurate and fast. Let the neural network use the past task to learn to predict the gradient, so that when facing a new task, as long as the gradient prediction is accurate, the learning will be fast.

[0008] Therefore, in order to solve the above problems, a blood sample classification method and system for coagulation index detection based on MAML and spatial transformation are proposed to solve the above problems. SUMMARY

[0009] In view of the problems in the prior art, the present application provides a blood sample classification method and system for coagulation index detection based on MAML and spatial transformation. The method uses MAML to continuously update the parameters of the model, so that it can quickly adapt to the task, thereby improving the accuracy of the gradient optimization direction, and combining double-module joint training to improve the classification accuracy.

[0010] To achieve the above object, the present application provides the following technical scheme:

[0011] The present application provides a blood sample classification method for coagulation index detection based on MAML and spatial transformation, comprising the following steps:

[0012] (1) Obtain the curve image of prothrombin time PT and thrombin time TT sample, divide the curve image into single small task, each task contains different categories, different categories contain different samples, extract features to create training set and test set;

[0013] (2) Use the curve image generated in step (1) as input, and divide it into multiple tasks, so that it goes through two stages of meta-training and meta-testing. Meta-training abstracts general features and strategies from learning in multiple tasks, and the training set image is input into a module combining spatial transformation and convolutional network to generate processed images. Meta-testing evaluates the performance of the learning algorithm;

[0014] (3) After meta-training, the model updates the parameters theta by gradient descent to construct a coagulation index classification model based on MAML and spatial transformation;

[0015] (4) Set the hyperparameters of the abnormal coagulation index classification model, including adjusting the meta-learning rate, the number of traversals epoch, the data size batch size, the optimizer and the number of iterations;

[0016] (5) Fine-tune and iterate the trained model, apply the coagulation index classification model after iteration and fine-tuning to the test set, and classify based on the MAML and spatial transformation model.

[0017] Further, in step (1):

[0018] The curve image of the PT and TT sample is subjected to feature extraction, and then the curve image is cut or segmented, and after the curve is split into multiple paragraphs, the shape and bending degree of each paragraph are extracted, and the features of each part are fused after being extracted;

[0019] The region gradient and energy of the five points around the main line layout of each pixel point of the PT and TT two images are calculated, the gray value of the point with large energy is taken as the gray value of the point of the new image, and the test set is formed.

[0020] Further, in step (2):

[0021] The meta-training is divided into two modules, one part is spatial transformation, and the other part is a feature extraction module composed of four convolutional blocks;

[0022] The spatial module includes a positioning network, a grid generator and a sampler, and the feature extraction module includes a CONV convolution layer, a BN normalization layer, a RELU activation function and a POOLING pooling layer;

[0023] The positioning network is a parameter network θs that generates a transformation matrix from an input image with a height×width×channel size curve image feature map, the generator generates a sampling grid from the input parameters, and then the pixels of the input curve image are rearranged in the output image. The sampler maps the pixels of the input curve image to the pixels in the output image, and then the spatial network module is inserted into different convolutional layers for joint training, and finally the input data is globally associated and combined through the fully connected layer FC.

[0024] Further, in step (3):

[0025] The training model is trained using the following cross-entropy loss function:

[0026]

[0027] Cross-entropy loss function LT i (fu θ ) in which x (j) ,y (j) ~T i represents the jth sample randomly selected from task T i , x (j) is the input of the sample, fu θ (·) represents the model prediction probability distribution calculated according to the parameters θ, y (j) is the output of sample j, and θ represents the initial parameters and is updated iteratively in parallel;

[0028] During the training phase, single optimization and double optimization are used, the single optimization updates the data by a certain gradient, and the double optimization calculates the overall optimization direction of the model by weighted average of the loss;

[0029] The meta-objective is to minimize the loss function and obtain parameters with strong generalization ability, and the update formula is as follows:

[0030]

[0031] Wherein the θ in the meta-objective L q (θ) represents the parameters of the model, q represents the index of the loss function L q (·), n represents the loss weight of the query, used to calculate the weighted sum, N represents that N steps have been updated by the common gradient descent, I represents the total number of tasks, and i represents the start of the calculation of the weighted sum from the ith task, T represents the target and loss of the task training n times, i θ s and θ c represent the initialized parameters, and represent the parameterized function.

[0032] Further, in step (4):

[0033] The hyperparameters of the abnormal coagulation index classification model are set, including adjusting the meta-learning rate, the number of traversals epoch, the data size batch size, the optimizer and the number of iterations;

[0034] Epoch represents the number of training iterations through a complete training data set, epoch can be set to an integer between 1 and infinity, which is a parameter condition for stopping algorithm training. In addition to selecting a fixed epoch, other training stop conditions can also be used to stop algorithm training, such as the amount of change in model error over time, batch size represents the number of samples processed before updating the model, batch size must be greater than or equal to 1, and less than or equal to the number of samples in the training data set;

[0035] Among the hyperparameters, the learning rate of the training phase is 0.01, the double optimization learning rate is 0.001, the optimizer is the Adam optimizer, epoch is set to 30, and batch size is set to 4.

[0036] Further, in step (5):

[0037] The trained model is adjusted in learning rate, epoch, and batch size, the coagulation index classification model after iteration test and fine-tuning is applied to 150 test sets, the abnormal index and normal index are iteratively tested and fine-tuned again, the learning rate, epoch and batch size are fine-tuned according to the accuracy and the results of the receiver operating characteristic curve, the classification model with the highest accuracy is obtained by comparing the above data, and the final result is obtained.

[0038] A blood sample system for coagulation index detection based on MAML and spatial transformation:

[0039] A data acquisition unit is used for acquiring coagulation index data of a blood sample to be detected, thereby acquiring curve images of prothrombin time PT and thrombin time TT;

[0040] A feature extraction and training model unit is used for dividing the acquired curve images into a training set and a test set, extracting curve image features, and taking the acquired curve images as input, and passing through a spatial transformation module and a feature extraction module composed of four convolutional groups;

[0041] A detection result determination unit is used for global association and combination through a fully connected layer FC, fine-tuning and iteration test optimization performance, then applying the coagulation index classification model to the test set, and combining the model based on MAML and spatial transformation for classification, and finally outputting the classification result.

[0042] The present application has the following advantages:

[0043] MAML is used for rapid learning, so that the neural network can use the past task learning prediction gradient to improve the speed when facing new tasks, and the double module joint training is combined to improve the classification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0044] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application together with the embodiments of the application, and do not constitute a limitation on the application.

[0045] Figure 1 The working flow chart of the coagulation index detection of blood samples based on MAML and spatial transformation of the present application.

[0046] Figure 2 The model network diagram of the coagulation index detection of blood samples based on MAML and spatial transformation of the present application.

[0047] Figure 3A space transformation network structure schematic diagram of a blood sample classification method and system based on MAML and space transformation for coagulation indicator detection according to the present application.

[0048] Figure 4 An APTT curve image 1 of a sample coagulation function indicator according to the present application.

[0049] Figure 5 An APTT curve image 2 of a sample coagulation function indicator according to the present application.

[0050] Figure 6 An APTT curve image 3 of a sample coagulation function indicator according to the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0052] Embodiment 1

[0053] A blood sample classification method based on MAML and space transformation for coagulation indicator detection, comprising the following steps:

[0054] (1) Obtain the curve image of prothrombin time PT and thrombin time TT sample, divide the curve image into single small tasks, each task contains different categories, different categories contain different samples, create training set and test set after feature extraction;

[0055] (2) Use the curve image generated in step (1) as input, and divide it into multiple tasks, so that it goes through two stages of meta-training and meta-testing. Meta-training abstracts general features and strategies from learning in multiple tasks. The training set image is input into a module combining space transformation and convolutional network to generate processed images. Meta-testing evaluates the performance of the learning algorithm;

[0056] (3) After meta-training, the model performs gradient descent to update the parameters θ, and constructs a coagulation indicator classification model based on MAML and space transformation;

[0057] (4) Set the hyperparameters of the abnormal coagulation indicator classification model, including adjusting the meta-learning rate, the number of traversals epoch, the data size batch size, the optimizer, and the number of iterations;

[0058] (5) Fine-tune and iteratively test the trained model, apply the coagulation index classification model that has been iteratively tested and fine-tuned to the test set, and combine it with the model based on MAML and spatial transformation for classification.

[0059] According to Example 1, the specific steps are as follows:

[0060] (1) This experiment used APTT and FIB curve images of 500 samples to create the training and test sets. Figures 4 to 6 The image shown is the APTT curve image of a portion of the samples used in this experiment. Figures 4 to 6 To illustrate, three different APTT curve images are shown.

[0061] The training set creation process is as follows: A coagulation analyzer is a device used in laboratories to detect thrombosis and hemostasis. This device uses a percentage detection method to measure the clotting time of blood. In this method, the intensity of transmitted light is defined as 0% when the reagent is first added to the sample and before coagulation has occurred. When the coagulation reaction is completely finished, the intensity of transmitted light is defined as 100%. Therefore, 50% of the transmitted light intensity corresponds to half of the coagulation time. Coagulation indicators can be obtained by analyzing the time-response curve of transmitted light intensity. The time required to reach a predetermined light intensity level is defined as the coagulation time. Using this method, six commonly used coagulation indicators can be obtained, including prothrombin time (PT), activated partial thromboplastin time (APTT), fibrinogen (FIB), thrombin time (TT), D-dimer, and fibrin degradation products (FDP). The above curve image feature extraction methods can be combined to obtain more accurate and comprehensive results, thereby improving the accuracy of the training set data.

[0062] The test set is created as follows: The gradient and energy of the regions surrounding each pixel in the image (upper, lower, left, right, and itself) are calculated. The gray values ​​of the pixels with higher energy are selected as the gray values ​​of the corresponding pixels in the new image, thus forming the test set. The test set is a dataset used to evaluate the final performance of the model and does not overlap with the training set; therefore, the gray values ​​of the pixels with higher energy are selected as the data for the test set.

[0063] (2) The generated image is input, divided into multiple tasks, and then subjected to meta-training and meta-testing two stages. Meta-training abstracts general features and strategies from learning of multiple tasks to reduce training time. Meta-testing evaluates the performance of the learning algorithm. In detail, the spatial transformation network includes a positioning network, a grid generator, and a sampler. The spatial module includes the positioning network, the grid generator, and the sampler. The feature extraction module includes a convolution layer CONV, a normalization layer BN, an activation function RELU, and a pooling layer POOLING. Meta-training and meta-testing belong to two processes in meta-learning. Meta-learning is a technology that learns how to learn. The training process and the testing process of meta-learning each require two types of data. Meta-learning can help the algorithm automatically infer the optimal algorithm or optimal hyperparameters suitable for different tasks, thereby improving the generalization ability and adaptability of the model. The spatial transformation network can enhance the robustness of the network to image deformation, rotation, and other geometric deformations by learning spatial transformation of input data. In the end-to-end training process, the transformation parameters are adaptively learned without manually setting the transformation mode and parameters. The positioning network is usually composed of a multilayer perceptron and some auxiliary layers. In the feature extraction module, RELU and other layers output feature maps. The curve smoothing differential of the operation result is more obvious in the intensity transformation of light, a nonlinear curve relationship is introduced to depict complex changes in input and train complex models. The bias term of the connection weight and the neuron of the fully connected layer FC is optimized through the training process. The pooling layer can reduce overfitting, reduce the input size, and improve performance.

[0064] The positioning network generates a transformation matrix parameter network θs from the input image feature map of height x width x channel size. The generator generates a sampling grid from the input parameter. Then, the pixels of the input curve image are rearranged in the output image. The sampler maps the pixels of the input curve image to the pixels in the output image. Finally, the fully connected layer FC globally associates and combines the input data, which facilitates easy embedding of the entire module into a convolutional neural network. Then, the spatial network module is inserted into different convolutional layers for joint training to improve classification accuracy.

[0065] (3) The output image is input into the meta-training model to update the parameter θ by gradient descent, and a coagulation index classification model based on MAML and spatial conversion is constructed.

[0066] The training model is trained using the following cross-entropy loss function:

[0067]

[0068] Cross-entropy loss function LT i (fu θ ) where x (j) ,y (j) ~Ti Indicates from task T i The j-th sample is randomly selected from the sample, x (j) For the input of the sample, fu θ (·) represents the model prediction probability distribution calculated based on parameter θ, y (j) For sample j, θ represents the initial parameters that are iteratively updated in parallel.

[0069] During the training phase, single optimization and double optimization are used. Single optimization performs gradient updates on the data, while double optimization calculates the overall optimization direction of the model by weighted averaging of the loss.

[0070] The meta-objective is to minimize the loss function and obtain parameters with strong generalization ability. The update formula is shown below:

[0071]

[0072] Among them, the metatarget L q In (θ), θ represents the model parameters, and q represents the loss function L. q The index of (·), The loss weights for the nth query are used to calculate the weighted sum, where N represents the N steps of co-gradient descent updates, I represents the total number of tasks, and i indicates that the weighted sum is calculated starting from the i-th task. T represents the objective and loss of the task training n times. i The distribution that belongs to the meta-training task, θ s and θ c Indicates initialization parameters, and This represents a parameterized function.

[0073] The detailed derivation of the formula is as follows:

[0074] First, initialize the parameters including θ. s and θ c The parameterized function is and T i The distribution belongs to the meta-training task, and the parameters θ are randomly initialized. s and θ c The parameter is first updated to θ i In each iteration, according to the loss function L Ti For parameter θ i gradient The update is performed, and the magnitude of the update is determined by the learning rate α. and Make adjustments. Then gradually change the parameter θ. iThe value of the loss function is minimized, so that the overall loss function is minimized as much as possible. In this way, by multiple iterations and updates, the parameters θ can be optimized to gradually learn better representation of samples, and the update formula is

[0075]

[0076] The n-step sum of the n-step descending update is

[0077]

[0078] Wherein The loss weight of the n-time query is used to calculate the weighted sum, α represents the learning rate, N represents the N-step gradient descent update, I represents the total number of tasks, i represents the calculation of the weighted sum from the i-th task, The target and loss of the n-time training on the task are represented, The update number of the parameterized function when the task is reached is represented.

[0079] (4) The hyperparameters of the abnormal coagulation index classification model are set, including adjusting the meta-learning rate, the traversal number of times epoch, the data size batch size, the optimizer and the iteration number; by marking the normal sample and the abnormal sample as binary numbers 1 and 0 respectively, the mean of the coagulation items between the normal and abnormal data sets is verified, it is assumed that the mean values of the two populations are equal, then the Pearson correlation coefficient between the two is observed, the correlation item change of the normal index and the abnormal index is observed. The Pearson correlation coefficient is a correlation analysis method, which can be used to measure the linear correlation between two variables, that is, the relationship between the normal sample and the abnormal sample is observed.

[0080] In the hyperparameters, the training stage learning rate is 0.01, the double optimization learning rate is 0.001, the optimizer is the Adam optimizer, the epoch is set to 30, and the batch size is set to 4.

[0081] (5) The trained model is fine-tuned and iteratively tested, the coagulation index classification model after iterative testing and fine-tuning is applied to the test set, and the model based on MAML and spatial transformation is classified.

[0082] The error acceptance ratio, the error rejection rate and the value of the generalized autoregressive of the coagulation index classification system based on MAML and spatial variation of the application are as follows:

[0083] FAR false accepts 0.18％ FRR false rejects 0.18％ GAR generalized autoregressive 99.82％

[0084] Wherein FAR refers to the proportion of false acceptance of abnormal indicators in the system, the lower the proportion, the better the classification effect of the system model on abnormal indicators, FRR refers to the proportion of false rejection of normal indicators, the lower the proportion, the better the classification effect of the system, GAR refers to a statistical model, which can better explain and predict the behavior of variables and sequences, and provide more complex and flexible pattern analysis and prediction ability. The higher the index, the stronger the analysis and prediction ability of the system, which means the better the classification effect.

[0085] Embodiment 2

[0086] A system for detecting blood samples of coagulation indicators based on MAML and spatial transformation

[0087] A data acquisition unit is used to acquire the coagulation indicator data of the blood sample to be detected, thereby acquiring the curve images of prothrombin time PT and thrombin time TT. The curve images of PT and TT are divided into single small tasks, each task contains different categories, and different categories contain different samples;

[0088] A feature extraction and training model unit is used to divide the acquired curve images into a training set and a test set. After extracting the curve image features, the acquired curve images are input into the spatial transformation module and the feature extraction module composed of four convolutional groups. Meta-training abstracts general features and strategies from learning of multiple tasks, the training set image is input into the module combining spatial transformation and convolutional network to generate processed images, thereby reducing the training time;

[0089] A detection result determination unit is used to perform global association and combination through a fully connected layer FC, and to apply the coagulation indicator classification model to the test set after fine-tuning and iterative test optimization. The classification model is combined with the model obtained based on MAML and spatial transformation for classification, and finally the classification result is output.

[0090] By adopting the technical scheme, the application provides a blood sample classification method and system for coagulation index detection based on MAML and spatial transformation, aiming to improve the speed and accuracy of coagulation index classification, and by using MAML and spatial transformation modules to classify coagulation indexes and realize fast learning, the neural network uses the past task learning prediction gradient to improve the speed when facing new tasks, and combines double-module joint training to improve the classification accuracy. In detail, first, the curve images of prothrombin time PT and thrombin time TT samples are acquired, features are extracted, and a training set and a test set are created, then the training set images are input into a module combining spatial transformation and convolution network to generate processed images, then the training model is used for gradient descent to update parameters, a spatial transformation coagulation index classification model is reconstructed, finally, the abnormal coagulation index classification model is set with hyperparameters, and the optimal model training process is found through model training and iterative testing, finally, the classification of coagulation indexes is realized, and the model can quickly adapt to the task, thereby improving the accuracy of the gradient optimization direction.

[0091] Although the present application has been disclosed in the preferred embodiments as above, it is not intended to limit the present application, and any person skilled in the art can make various modifications and modifications without departing from the spirit and scope of the present application, therefore, the protection scope of the present application should be defined by the claims.

Claims

1. A classification method for blood samples in coagulation index detection based on MAML and spatial transformation, characterized in that, Includes the following steps: (1) Obtain curve images of prothrombin time (PT) and thrombin time (TT) samples, divide the curve images into individual small tasks, each task contains different categories, and different samples are contained under different categories. After extracting features, create training and test sets. (2) Use the curve image generated in step (1) as input and divide it into multiple tasks, so that it goes through two stages: meta-training and meta-testing. Meta-training abstracts general features and strategies from the learning of multiple tasks, and generates processed images by combining the training set image input space transformation with the convolutional network module. Meta-testing evaluates the performance of the learning algorithm. (3) After meta-training, the model is subjected to gradient descent to update the parameters θ, and a coagulation index classification model based on MAML and spatial transformation is constructed. Step (3) includes the following steps: The model was trained using the following cross-entropy loss function: Cross-entropy loss function middle Indicates from the task Randomly select the j-th sample from the sample. For the input of the sample, Indicates based on parameters The calculated model predicts the probability distribution. For the output of sample j, This is represented as initial parameters being iteratively updated in parallel. During the training phase, single optimization and double optimization are used. Single optimization performs gradient updates on the data, while double optimization calculates the overall optimization direction of the model by weighted averaging of the loss. The meta-objective is to minimize the loss function and obtain parameters with strong generalization ability. The update formula is shown below: Meta-objective In Represents the parameters of the model. Represents the loss function The index, The loss weights for the query at n iterations are used to calculate the weighted sum, where N represents the number of steps updated by common gradient descent, and I represents the total number of tasks. Indicates from the first The weighted average is calculated for each task. This represents the objective and loss after n training iterations of the task. Distributions belonging to the meta-training task. and Indicates initialization parameters, and Indicates a parameterized function; (4) Set hyperparameters for the abnormal coagulation index classification model, including adjusting the meta-learning rate, number of epochs, batch size, optimizer, and number of iterations; Step (4) includes the following steps: Hyperparameter settings for the abnormal coagulation index classification model include adjusting the meta-learning rate, number of epochs, batch size, optimizer, and number of iterations. (5) Fine-tune and iteratively test the trained model, apply the coagulation index classification model that has been iteratively tested and fine-tuned to the test set, and combine it with the model based on MAML algorithm and spatial transformation for classification.

2. The blood sample classification method for coagulation index detection based on MAML and spatial transformation according to claim 1, characterized in that, Step (1) includes the following steps: Feature extraction is performed on the curve images of PT and TT samples. Then, the curve images are cut or segmented. After splitting the curve into multiple segments, the shape and curvature features of each segment are extracted. The features of each part are then fused together. Calculate the gradient and energy of the region around the main line layout of each pixel in the two images PT and TT, including the top, bottom, left, right, and itself. Take the gray value of the pixel with the higher energy as the gray value of the pixel in the new image to form a test set.

3. The blood sample classification method for coagulation index detection based on MAML and spatial transformation according to claim 1, characterized in that, Step (2) includes the following steps: Meta-training is divided into two modules: one is spatial transformation, and the other is a feature extraction module consisting of four convolutional blocks. The spatial module includes a localization network, a mesh generator, and a sampler; the feature extraction module includes a CONV convolutional layer, a BN normalization layer, a ReLU activation function, and a POOLING pooling layer. The localization network is a parametric network that generates a transformation matrix from the curve image feature map (height × width × channel size) of the input image. The generator generates a sampling grid for the input parameters, then rearranges the pixels of the input curve image in the output image. The sampler maps the pixels of the input curve image to the pixels in the output image. The spatial network module is then inserted into different convolutional layers for joint training. Finally, the input data is globally correlated and combined through a fully connected layer (FC).

4. The blood sample classification method for coagulation index detection based on MAML and spatial transformation according to claim 1, characterized in that: Among the hyperparameters, the learning rate during the training phase is 0.01, the double-optimization learning rate is 0.001, the optimizer is the Adam optimizer, the epoch is set to 30, and the batch size is set to 4.

5. The blood sample classification method for coagulation index detection based on MAML and spatial transformation according to claim 4, characterized in that: The meta-learning rate is adjusted as follows: if the gradient of a certain component often maintains the same sign in two adjacent steps, the cumulative result of that component may be positive, and the learning rate is increased in this case; if the gradients in two adjacent steps often have opposite signs, the corresponding cumulative result may be negative, and the learning rate is decreased in this case.

6. The blood sample classification method for coagulation index detection based on MAML and spatial transformation according to claim 1, characterized in that, Step (5) includes the following steps: The training model was adjusted in terms of learning rate, number of epochs, and batch size. The coagulation index classification model, which had been iteratively tested and fine-tuned, was applied to a test set of 150 samples. Abnormal and normal indices were iteratively tested and fine-tuned again. The learning rate, epochs, and batch size were fine-tuned based on the obtained accuracy and the results of the receiver operating characteristic curve. The above data were compared, and the classification model with the highest accuracy was the final result.

7. A system for detecting coagulation indicators in blood samples based on MAML and spatial transformation, performing the classification method for blood samples in detecting coagulation indicators based on MAML and spatial transformation as described in claim 1, characterized in that, include: Data acquisition unit: used to acquire coagulation index data of the blood sample to be tested, thereby obtaining curve images of prothrombin time (PT) and thrombin time (TT); Feature extraction and training model unit: The acquired curve image is divided into training set and test set. After extracting the curve image features, the acquired curve image is used as input and passed through the spatial transformation module and the feature extraction module consisting of four convolutions. The detection result determination unit: After global association and combination through a fully connected layer (FC), fine-tuning and iterative testing to optimize performance, the coagulation index classification model is applied to the test set, and the model obtained based on MAML and spatial transformation is combined for classification. Finally, the classification results are output.

Images