Application method and system of self-supervised learning model in cell image classification process

Abstract

The present application discloses a method and system for applying a self-supervised learning model in the process of cell image classification. The embodiment of the present application uses a self-supervised learning method to learn an initialized cell classification model based on multiple unlabeled cell image data. The obtained initialized cell classification model is used as a pre-trained model for cell image classification, and the initialized cell classification model is adjusted using a transfer learning method after learning is completed. When applied, the unlabeled cell image data is input into the pre-trained model for processing, and the output result is input into the cell image classification network model, and the classification result of the cell image is output. In this way, since the initialized cell classification model is learned using a self-supervised learning method, the cell image classification process is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of computer image processing, and in particular to a method and system for applying a self-supervised learning model in a cell image classification process. Background Art

[0002] In recent years, supervised learning-based machine learning has made significant contributions to the field of artificial intelligence. Supervised learning involves inferring a function from labeled data using a trained machine learning network. However, training supervised learning-based machine learning networks requires a large amount of manually labeled training data. Due to the limited availability of manually labeled data and insufficient processing capabilities, supervised learning-based machine learning has only become a practical technology for commercialization in recent years. Supervised learning-based machine learning is only applicable when there is sufficient high-quality training data that can capture all possible scenarios. However, in the field of medical imaging, obtaining high-quality, accurately labeled training data is difficult, costly, and limited in quantity, making it impossible to obtain highly accurate machine learning networks for medical image classification.

[0003] There is growing interest in how to use massive amounts of unlabeled training data to train machine learning networks, thereby obtaining networks that can accurately classify cell images for subsequent applications. Consequently, self-supervised learning techniques have been proposed. Self-supervised learning can be considered an ideal state for machine learning network models: the model learns directly from unlabeled data, eliminating the need for labeled training data to produce a trained machine learning network. However, applying self-supervised learning techniques to the classification of cell images remains a pressing issue. Summary of the Invention

[0004] In view of this, an embodiment of the present application provides a method for applying a self-supervised learning model in the cell image classification process. This method can apply self-supervised learning technology to the learning process of the initialized cell classification model, thereby optimizing the cell image classification process in subsequent applications.

[0005] An embodiment of the present application also provides an application system of a self-supervised learning model in the process of cell image classification. The system can apply self-supervised learning technology to the learning process of the initialized cell classification model, thereby optimizing the cell image classification process in subsequent applications.

[0006] The embodiment of the present application is implemented as follows:

[0007] A method for applying a self-supervised learning model in a cell image classification process includes:

[0008] Based on multiple unlabeled cell image data, a self-supervised learning method is used to learn the initialized cell classification model, and the obtained initialized cell classification model is used as a pre-training model for cell image classification;

[0009] During application, unlabeled cell image data is input into the pre-training model for processing, and the output result is input into the cell image classification network model to output the classification result of the cell image.

[0010] The initialized cell classification model includes an image transformation layer, a convolutional neural network layer CNN, and a multi-layer perceptron MLP layer, wherein:

[0011] An image transformation layer, performing image transformation on the unlabeled cell image data;

[0012] CNN, performing convolution calculation on the unlabeled cell image data after image transformation;

[0013] The MLP performs perception calculation on the unlabeled cell image data after the image transformation by the convolution calculation, and obtains feature vectors corresponding to the unlabeled cell image data after each image transformation.

[0014] The learning of the cell classification model initialized based on multiple unlabeled cell image data includes:

[0015] After the unlabeled cell image data is transformed, multiple cell image enhancement data are obtained, and after CNN and MLP calculations are performed respectively, feature vectors are obtained;

[0016] During the learning process, in each iteration of the number of iterations set for the model, the internal parameters in the CNN and MLP are adjusted so that the difference values ​​of the feature vectors obtained in the feature space after the multiple cell image enhancement data from the same cell image data are processed by the CNN and MLP continuously decrease during the iteration process, and the difference values ​​of the feature vectors obtained in the feature space after the multiple cell image enhancement data from different cell image data are processed by the CNN and MLP continuously increase during the iteration process.

[0017] Before application, the method further comprises: adjusting the obtained initialized cell classification model by using a transfer learning method.

[0018] The step of adjusting the obtained initialized cell classification model by using a transfer learning method includes:

[0019] The obtained initialized cell classification model is used as a pre-training model and fine-tuned using labeled cell images.

[0020] The fine-tuning using the annotated cell image includes:

[0021] Adjust the last fully connected layer in the CNN of the initialized cell classification model, or adjust the last feature extraction layer and the last fully connected layer in the CNN, or adjust all feature extraction layers and the last fully connected layer in the CNN.

[0022] Prior to applying, the method further comprises:

[0023] The initialized cell classification model was experimentally verified to ensure that it met the verification requirements.

[0024] A self-supervised learning model application system in the process of cell image classification, the system includes: an acquisition module, a self-supervised learning module and an application module, wherein:

[0025] An acquisition module, used for acquiring multiple unlabeled cell image data;

[0026] A self-supervised learning module is used to learn an initialized cell classification model based on multiple unlabeled cell image data using a self-supervised learning method, and use the obtained initialized cell classification model as a pre-training model for cell image classification;

[0027] The application module is used to input the unlabeled cell image data into the pre-training model for processing, input the output result into the cell image classification network model, and output the classification result of the cell image.

[0028] The system further comprises: a supervised learning module, which is used to adjust the obtained initialized cell classification model by using a transfer learning method before the application module is applied.

[0029] As described above, the embodiment of the present application uses a self-supervised learning method to learn an initialized cell classification model based on multiple unlabeled cell image data, and uses the obtained initialized cell classification model as a pre-trained model for cell image classification. After the initialization of the cell classification model is learned, the model is adjusted by a transfer learning method. When applied, the unlabeled cell image data is input into the pre-trained model for processing, and the output result is input into the cell image classification network model, and the classification result of the cell image is output. In this way, since the initialized cell classification model is learned by a self-supervised learning method, and the initialized cell classification model is further adjusted by a transfer learning method, the cell image classification process is optimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1A flow chart of a method for applying a self-supervised learning model in a cell image classification process provided in an embodiment of the present application;

[0031] Figure 2 A flow chart of a method for learning and obtaining an initialized cell classification model provided in an embodiment of the present application;

[0032] Figure 3 Schematic diagram of the process of processing unlabeled cell image data by the initialized cell classification model provided in an embodiment of the present application;

[0033] Figure 4 A schematic diagram of the application system structure of a self-supervised learning model in the cell image classification process provided in an embodiment of the present application. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0035] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the application described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0036] The following specific embodiments are used to describe the technical solution of the present application in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0037] At present, the mainstream self-supervised learning technology adopts the idea of ​​contrastive learning, which performs self-supervised learning of the internal structure information in the machine learning network model based on unlabeled data. Usually, after performing various transformations on an image data, the optimization goal is to make different transformations of the same image data as close as possible in the feature space, and to make image data of different frames as far apart as possible in a specific space.

[0038] Therefore, the embodiment of the present application applies self-supervised learning technology. Based on multiple unlabeled cell image data, a self-supervised learning method is used to learn the initialized cell classification model, and the obtained initialized cell classification model is used as a pre-trained model for cell image classification. When applied, the unlabeled cell image data is input into the pre-trained model for processing, and the output result is input into the cell image classification network model to output the classification result of the cell image.

[0039] Before actually applying the initialized cell classification model, the embodiment of the present application also adjusts the initialized cell classification model, specifically using a transfer learning method for adjustment. The transfer learning method is a machine learning method that transfers the knowledge of the source field to the target field so that the target field can achieve good learning results. Usually, the amount of data in the source field is sufficient and the amount of data in the target field is small. This scenario is very suitable for the transfer learning method. Especially in the task of classifying cell images, the use of the transfer learning method can greatly improve the accuracy of the cell image classification results of the cell image classification network model learned when the labeled training sample data is insufficient.

[0040] In this way, the cell image classification process is optimized because the initialized cell classification model is learned in a self-supervised learning manner.

[0041] Figure 1 A flowchart of a method for applying a self-supervised learning model in a cell image classification process provided in an embodiment of the present application, wherein the specific steps include:

[0042] Step 101: Based on a plurality of unlabeled cell image data, an initialized cell classification model is learned using a self-supervised learning method, and the obtained initialized cell classification model is used as a pre-training model for cell image classification;

[0043] Step 102: When applied, the unlabeled cell image data is input into the pre-trained model for processing, and the output result is input into the cell image classification network model to output the classification result of the cell image.

[0044] In this method, the initialized cell classification model includes an image transformation layer, a convolutional neural network layer (CNN), and a multi-layer perceptron (MLP) layer, wherein,

[0045] An image transformation layer, performing image transformation on the unlabeled cell image data;

[0046] CNN, performing convolution calculation on the unlabeled cell image data after image transformation;

[0047] The MLP performs perception calculation on the unlabeled cell image data after the image transformation by the convolution calculation, and obtains feature vectors corresponding to the unlabeled cell image data after each image transformation.

[0048] In this method, the learning of the cell classification model initialized based on multiple unlabeled cell image data includes:

[0049] After the unlabeled cell image data is transformed, multiple cell image enhancement data are obtained, and after CNN and MLP calculations are performed respectively, feature vectors are obtained;

[0050] During the learning process, in each iteration of the number of iterations set for the model, the internal parameters in the CNN and MLP are adjusted so that the difference values ​​of the feature vectors obtained in the feature space after the multiple cell image enhancement data from the same cell image data are processed by the CNN and MLP continuously decrease during the iteration process, and the difference values ​​of the feature vectors obtained in the feature space after the multiple cell image enhancement data from different cell image data are processed by the CNN and MLP continuously increase during the iteration process.

[0051] In this method, the obtained initialized cell classification model is adjusted using transfer learning before application.

[0052] Specifically, the adjustment of the obtained initialized cell classification model using a transfer learning method includes:

[0053] The obtained initialized cell classification model is used as a pre-training model, and the cell classification model is fine-tuned using labeled cell images.

[0054] Here, the fine-tuning using annotated cell images includes:

[0055] Adjust the last fully connected layer in the CNN of the initialized cell classification model, or adjust the last feature extraction layer and the last fully connected layer in the CNN, or adjust all feature extraction layers and the last fully connected layer in the CNN.

[0056] In the above method, before applying, it also includes:

[0057] The initialized cell classification model was experimentally verified to ensure that it met the verification requirements.

[0058] As can be seen from the above method, the embodiment of the present application adopts the idea of ​​contrastive learning for self-supervised learning, and performs self-supervised learning of the internal structure data in the cell classification model initialized based on a large amount of unlabeled cell image data, and learns to obtain an initialized cell classification model that can process unlabeled cell image data, and its output results are used as a pre-trained model of the input features of the downstream cell image classification network model. The embodiment of the present application uses the above-mentioned learned pre-trained model and applies it to the classification task of cell image data. Prior to this, it uses transfer learning to make adjustments to further improve the classification accuracy of cell image data.

[0059] Figure 2 The flowchart of the method for learning and obtaining an initialized cell classification model provided in the embodiment of the present application includes the following specific steps:

[0060] Step 201: Obtain a training set of cell image data;

[0061] In this step, 2.6 million unlabeled cell image data can be obtained as a training set for learning the initialized cell classification model; 200,000 labeled cell image data can also be obtained as a subsequent adjustment of the learned initialized cell classification model using transfer learning;

[0062] Step 202: Based on the acquired training set of cell image data, learn the initialized cell classification model to obtain the initialized cell classification model;

[0063] In this step, contrastive learning is used to perform self-supervised learning. This involves performing various transformations on the unlabeled cell image data in the training set. The goal is to keep different transformations of the same cell image data as close as possible in the feature space, while keeping different cell image data as far apart as possible in the feature space. This helps learn the initialized cell classification model.

[0064] Step 203: Using the labeled cell image data, adjust the initialized cell classification model using a transfer learning approach;

[0065] In this step, the process of adjusting the initialized cell classification model is actually a supervised learning process;

[0066] Step 204: Experimentally verify the initialized cell classification model to ensure that it meets verification requirements.

[0067] exist Figure 2 In step 201, a training set of cell image data is obtained as follows.

[0068] Collect and organize 2.6 million unlabeled cell images for self-supervised learning. Collect and organize 200,000 labeled cell images for subsequent adjustments to the learned initialized cell classification model using transfer learning.

[0069] Because the number of negative samples in the cell image data far exceeds the number of positive samples, resulting in a significant long-tail distribution, applying this data to self-supervised learning tasks will reduce the model's ability to learn from the smaller number of positive samples. Therefore, the unlabeled image data should contain at least 10% positive samples. These cell image data do not require manual labeling and can be directly exported and used for self-supervised model learning. In the subsequent supervised learning phase, the cell image data requires image category labeling by professional pathologists. After processing by the initialized cell classification model, the final results are quality-controlled by three experienced pathologists. The cell classification labels are negative, positive, and impurity. The average number of labeled cell images per class in the training set is approximately 66,000, of which 20% is randomly selected from the training set as the test set.

[0070] Figure 3 A schematic diagram of the process of processing unlabeled cell image data by the initialized cell classification model provided in an embodiment of the present application is shown in the figure. After the unlabeled cell image data is transformed, multiple cell image enhancement data are obtained, which are processed by a convolutional neural network (CNN) and a multi-layer perceptron (MLP) respectively to obtain feature vectors. During the learning process, the internal parameters in the CNN and the MLP are adjusted so that the feature vectors obtained after the multiple cell image enhancement data from the same cell image data are processed by the CNN and the MLP are as close as possible in the feature space (attract), and the feature vectors obtained after the multiple cell image enhancement data from different cell image data are processed by the CNN and the MLP are as far away as possible in the feature space (repel).

[0071] During the self-supervised learning process of the model, the server used had a memory size of 256GB, a CPU model of 64-core Intel(R) Xeon(R) Gold 5218, a graphics card of GeForce RTX 2080Ti, and a video memory size of 40GB. The number of bytes supported for learning was 1024, and the initial learning rate was 0.1*byte / 256. The model optimization algorithm used was the stochastic gradient descent algorithm (SGD) combined with the momentum setting method (Momentum), the learning rate descent method was cosine annealing, and the weight decay size was set to 1e-4. The cell image data was randomly cropped to a size of 224*224, and the data enhancement method was random cropping and color transformation. The deep learning framework used was PyTorch. The number of training iterations (epochs) was 400.

[0072] The model uses MobileNetv3 as the backbone network during self-supervised learning, and the final output feature dimension after the backbone network is 256. The MLP in the model is a two-layer fully connected structure, and the activation function in the middle is ReLU. The model processing process is as follows Figure 3 As shown, the process is: first, the input image is subjected to random data enhancement, and after passing through the feature extraction network, a feature vector with a feature dimension of 128 dimensions is output. The idea of ​​contrastive learning is used for self-supervised learning. Usually, various transformations are performed on the image, and then the optimization goal is to make different transformations of the same image as close as possible in the feature space, and different images as far away as possible in the feature space. This is used as the target information for learning the model, and the learned model is used as a pre-training model. The specific implementation method of the model is to encode the feature vector output by the model into a contrast loss function, and through the back-propagation algorithm, continuously optimize the error of the model. Each iterative process updates the parameters in the model, and finally obtains the learned model. The specific calculation method of the loss function is:

[0073]

[0074] in, In the formula, z i and z j It represents the feature vector output after the same image has been enhanced twice and input into CNN and MLP structures, z k Represents the feature vectors output by other different images and their data-enhanced images after passing through the CNN and MLP structures. sim(z i ,z j ) is the similarity between two feature vectors, |z i | is the vector representation after feature regularization. t represents the hyperparameter, and the initial setting of t is 0.07.

[0075] After the model has undergone the above self-supervised learning, Figure 2 Step 203 in the method uses the labeled cell image data as a test set to perform supervised learning on the model and adjust the internal parameters of the model.

[0076] There are three approaches to adjusting the model:

[0077] 1) Freeze all CNN feature extraction layers of the model and only adjust the last fully connected layer in the model.

[0078] 2) Freeze the feature extraction layers of some CNNs of the model, and only adjust the feature extraction layer of the last block of CNN and the last fully connected layer in CNN.

[0079] 3) Without freezing, all the CNN feature extraction layers and the last layer of the model are adjusted.

[0080] The initial learning rates for adjusting the CNN structure of the model are set to 1e-1*Batc size / 256, 1e-2*Batc size / 256, and 1e-3*Batch size / 256 respectively. The initial learning rates for adjusting the fully connected layer (fc) structure of the last layer are set to 1e*Batc size / 256, 1e-1*Batc size / 256, and 1e-2*Batch size / 256 respectively. The algorithm optimizer used is Adam, the learning rate reduction method is cosine annealing, and the number of training iterations epoch is 200. The backbone network for adjusting the model is also MobileNetv3, and the dimension of the last classification layer is 1*3. The image is randomly cropped to a size of 224*224, and the data enhancement methods are random horizontal flipping, color conversion, random erasing, and random noise. After experimental verification, it is concluded that the average experimental results of the third method have the highest accuracy and the best effect in the three classifications of images. Therefore, the embodiments of the present application usually use the third method to adjust the model.

[0081] After adjusting the model, the initialized cell classification model is experimentally verified to determine its classification effect as a pre-training model for subsequent cell image data classification.

[0082] Compare the classification effects of three transfer learning methods for classifying cell image data.

[0083] a) Use the model learned in this application as a pre-trained model.

[0084] b) Use a supervised model trained on ImageNet data as a pre-trained model. Specifically, a 1000-class classification model trained on the ILSVRC ImageNet 2012 data was used as the pre-trained model. This pre-trained model was applied to the downstream task of cell image data classification, and a three-class classification experiment was conducted on cell image data.

[0085] c) Use random initialization as the initialization model.

[0086] By comparing the classification results of cell image data obtained from experiments a, b, and c, the classification accuracy of experiment a is higher.

[0087] The following experimental conclusions were drawn from the above experimental results: A training set based on a large amount of unlabeled cell image data can effectively learn cell features. Compared with directly using a model pre-trained on Imagenet or random initialization, a better initialization model method is provided for processing downstream cell image classification tasks. This effectively improves the transfer learning effect of downstream cell image classification tasks. The initialized cell classification model is learned using a large amount of unlabeled cell image data. In subsequent use, the unlabeled cell image data is directly pre-processed before being processed by the downstream cell image classification model to obtain the final image classification result, reducing the downstream cell image classification model's demand for a large amount of labeled cell image data.

[0088] Figure 4 A schematic diagram of the application system structure of a self-supervised learning model in the process of cell image classification provided by an embodiment of the present application includes an acquisition module, a self-supervised learning module and an application module, wherein:

[0089] An acquisition module, used for acquiring multiple unlabeled cell image data;

[0090] A self-supervised learning module is used to learn an initialized cell classification model based on multiple unlabeled cell image data using a self-supervised learning method, and use the obtained initialized cell classification model as a pre-training model for cell image classification;

[0091] The application module is used to input the unlabeled cell image data into the pre-training model for processing, input the output result into the cell image classification network model, and output the classification result of the cell image.

[0092] The system also includes a supervised learning module for adjusting the obtained initialized cell classification model using a transfer learning method before the application module is applied.

[0093] It can be seen from the method and system provided in the embodiments of the present application that cell characteristics can be effectively learned based on a large amount of unlabeled cell image data, which effectively improves the transfer learning effect of the cell image classification task, utilizes a large amount of cell image data, and reduces the subsequent demand for a large amount of cell image data.

[0094] Those skilled in the art will appreciate that the features described in the various embodiments and / or claims of this disclosure may be combined and / or coupled in various ways, even if such combinations and / or couplings are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments and / or claims of this application may be combined and / or coupled in various ways, and all such combinations and / or couplings fall within the scope of this application.

[0095] Specific embodiments are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core ideas of the present application and is not intended to limit the present application. For those skilled in the art, changes can be made in the specific implementation methods and application scope based on the ideas, spirit and principles of the present application. Any modifications, equivalent replacements, improvements, etc. made by them should be included in the scope of protection of this application.

Claims

1. A method for applying a self-supervised learning model in the process of cell image classification, characterized in that: include: Based on multiple unlabeled cell image data, a self-supervised learning method is used to learn an initialized cell classification model, and the obtained initialized cell classification model is adjusted using a transfer learning method. The adjusted initialized cell classification model is used as a pre-training model for cell image classification; During application, unlabeled cell image data is input into the pre-trained model for processing to obtain an output result of a feature vector for characterizing the input cell image data, and the output result is input into a downstream cell image classification network model of the cell classification model to obtain a classification result of the cell image; in, The initialized cell classification model includes an image transformation layer, a convolutional neural network layer CNN, and a multi-layer perceptron MLP layer, wherein: An image transformation layer, performing image transformation on the plurality of unlabeled cell image data respectively; CNN, performing convolution calculation on the multiple unlabeled cell image data after image transformation; MLP, performing perception calculation on the plurality of unlabeled cell image data after the image transformation by convolution calculation, to obtain a feature vector corresponding to each of the unlabeled cell image data after the image transformation; The learning of the cell classification model initialized by self-supervised learning based on multiple unlabeled cell image data includes: After the plurality of unlabeled cell image data are transformed, a plurality of cell image enhancement data of each unlabeled cell image data are obtained, and after CNN and MLP calculations are performed respectively, a feature vector of each cell image enhancement data is obtained; During the learning process, in each iteration, the internal parameters of CNN and MLP are adjusted so that the difference in the feature space between the feature vectors obtained after the enhanced cell image data from the same cell image data are processed by CNN and MLP is continuously reduced during the iteration, and the difference in the feature space between the feature vectors obtained after the enhanced cell image data from different cell image data are processed by CNN and MLP is continuously increased during the iteration; The unlabeled cell image data includes positive samples and negative samples.

2. The method according to claim 1, wherein In the learning process, in each iteration, the internal parameters of the CNN and the MLP are adjusted so that the difference value of the feature vectors of each cell image enhancement data from the same cell image data in the feature space is continuously reduced during the iteration process, and the difference value of the feature vectors of each cell image enhancement data from different cell image data in the feature space is continuously increased during the iteration process, including: Each feature vector is encoded into a contrast loss function, and the internal parameters are updated through back propagation. Among them, the specific calculation method of the loss function is: in, z i and z j Represents the feature vector of the same image after two data enhancements, z k Represents the feature vectors of other different images after two data enhancements. sim(z i ,z j ) is the similarity between two feature vectors, |z i |、|z j | is the vector representation after feature regularization, and t represents the hyperparameter.

3. The method according to claim 1, wherein The step of adjusting the obtained initialized cell classification model by using a transfer learning method includes: The obtained initialized cell classification model is used as a pre-training model and fine-tuned using labeled cell images.

4. The method according to claim 3, wherein The fine-tuning using the annotated cell image includes: Adjust the last fully connected layer in the CNN of the initialized cell classification model, or adjust the last feature extraction layer and the last fully connected layer in the CNN, or adjust all feature extraction layers and the last fully connected layer in the CNN.

5. The method according to claim 1, wherein Prior to applying, the method further comprises: The initialized cell classification model was experimentally verified to ensure that it met the verification requirements.

6. A self-supervised learning model application system in the process of cell image classification, characterized in that: The system includes: an acquisition module, a self-supervised learning module and an application module, wherein: An acquisition module, used for acquiring multiple unlabeled cell image data; The self-supervised learning module is used to learn the cell classification model initialized by self-supervised learning based on multiple unlabeled cell image data. A supervised learning module is used to adjust the obtained initialized cell classification model using a transfer learning method, and use the adjusted initialized cell classification model as a pre-training model for cell image classification; An application module is used to input the unlabeled cell image data into the pre-trained model for processing, obtain an output result of a feature vector used to characterize the input cell image data, input the output result into a downstream cell image classification network model of the cell classification model, and output a classification result of the cell image; in, The initialized cell classification model includes an image transformation layer, a convolutional neural network layer CNN, and a multi-layer perceptron MLP layer, wherein: An image transformation layer, performing image transformation on the plurality of unlabeled cell image data respectively; CNN, performing convolution calculation on the multiple unlabeled cell image data after image transformation; The MLP performs perceptual calculations on the plurality of unlabeled cell image data after image transformation by convolution calculations to obtain feature vectors corresponding to the unlabeled cell image data after each image transformation; the learning of the cell classification model initialized based on the plurality of unlabeled cell image data includes: After the plurality of unlabeled cell image data are transformed, a plurality of cell image enhancement data of each unlabeled cell image data are obtained, and after CNN and MLP calculations are performed respectively, a feature vector of each cell image enhancement data is obtained; During the learning process, in each iteration, the internal parameters of CNN and MLP are adjusted so that the difference in the feature space between the feature vectors obtained after the enhanced cell image data from the same cell image data are processed by CNN and MLP is continuously reduced during the iteration, and the difference in the feature space between the feature vectors obtained after the enhanced cell image data from different cell image data are processed by CNN and MLP is continuously increased during the iteration; The plurality of unlabeled cell image data include positive samples and negative samples.

Images