Biological cell image classification method, system and storage medium

Abstract

The application discloses a biological cell image classification method and system and a storage medium, relates to the field of biomedical analysis, and comprises the following steps: acquiring a biological cell image to be classified, using a pre-trained AlexNet network as a basic feature extractor to acquire a one-dimensional feature vector; acquiring key feature information through feature selection and dimension reduction, mapping the one-dimensional feature vector after dimension reduction to different feature spaces, and generating different feature representations through reconstruction; constructing multi-modal features, converting the multi-modal features into a feature map by adopting an attention mechanism, and generating a biological cell image; constructing a classification model based on a graph convolution network, pre-training by using transfer learning, taking the biological cell image to be identified as input, and acquiring the class probability of each biological cell image. The application combines deep learning and multi-modal fusion technology, improves the biological cell image classification precision through feature reconstruction and transfer learning, and provides reliable technical support for biomedical and clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of biomedical analysis technology, and more specifically, to a method, system, and storage medium for classifying biological cell images. Background Technology

[0002] Biological cell image classification is an important task in computer vision and biomedical analysis, and is widely used in disease diagnosis, drug screening, pathological research and other fields. With the rapid development of microscopy technology and imaging equipment, high-throughput cell image data can be generated, which has created an urgent need for automated and high-precision cell classification methods. Traditional cell classification methods mainly rely on manual observation and expert experience, but this method is inefficient, subjective, and difficult to handle large-scale data. Early biological cell image classification was mainly based on traditional image processing technology and machine learning methods. Common processes include: (1) image preprocessing; (2) feature extraction (such as morphological features, texture features, color features); (3) classifier design (such as support vector machine, random forest, K-nearest neighbor algorithm). Although these methods have achieved certain results, they rely on manually designed features, making it difficult to capture the complex structure and subtle differences of cell images, and their generalization ability is limited.

[0003] In recent years, breakthroughs in deep learning technology have significantly improved the performance of cell image classification. Convolutional neural networks, due to their powerful feature extraction capabilities, have become the mainstream method for cell classification. For example, classic network structures (such as AlexNet, VGG, and ResNet) can automatically learn multi-level features of cells through end-to-end training, significantly improving classification accuracy. However, biological cell image classification still faces many challenges: (1) cell morphology is diverse, and the same type of cells may exhibit significant differences under different conditions; (2) labeled data is scarce, and high-quality labeling requires the participation of domain experts, which is costly; (3) differences in imaging conditions (such as illumination and resolution) may affect the robustness of the model. Therefore, how to combine deep learning and multimodal fusion methods to improve the accuracy of biological cell image classification is an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a biological cell image classification method, system, and storage medium. By combining deep learning and multimodal fusion technology, the accuracy of biological cell image classification is improved through feature reconstruction and transfer learning.

[0005] The first aspect of this invention provides a method for classifying biological cell images, comprising the following steps:

[0006] The images of biological cells to be classified are acquired and preprocessed. A pre-trained AlexNet network is used as the basic feature extractor to obtain a one-dimensional feature vector.

[0007] The one-dimensional feature vector is subjected to feature selection and dimensionality reduction to obtain key feature information. The dimensionality-reduced one-dimensional feature vector is mapped to different feature spaces and different feature representations are generated by reconstruction.

[0008] Different feature representations are weighted and fused to construct multimodal features. An attention mechanism is then used to convert the multimodal features into feature maps to construct a biological cell profile.

[0009] A classification model is constructed based on graph convolutional networks, and a transfer learning strategy is used to pre-train the classification model. The pre-trained classification model is then adapted to biological cell image data, and the biological cell images to be classified are used as input to obtain the category probability of each biological cell image.

[0010] In this approach, images of the biological cells to be classified are acquired and preprocessed. A pre-trained AlexNet network is used as the basic feature extractor to obtain a one-dimensional feature vector, specifically:

[0011] Collect images of different types of biological cells under different imaging conditions, construct a biological cell image set, perform data augmentation and normalization processing on the biological cell image set, and generate a training set and a test set according to a preset ratio;

[0012] The AlexNet network was trained and validated using training and test sets. Dilated convolutions were introduced into the AlexNet network to replace standard convolutions, and dilated convolutions with stride of 1 were combined with global average pooling to replace the last max pooling layer. The fully connected classification layer in the AlexNet network was removed to construct a basic feature extractor.

[0013] Acquire images of biological cells to be classified under the current imaging conditions, perform standardized preprocessing on the images of biological cells to be classified, dynamically adjust the size of the preprocessed images of biological cells to be classified, input them into a basic feature extractor to extract multi-scale features, and fuse feature maps with different expansion rates through skip connections.

[0014] Global average pooling is used to average the fused feature maps to obtain a one-dimensional feature vector of the biological cell image to be classified.

[0015] In this scheme, feature selection and dimensionality reduction are performed on the one-dimensional feature vector to obtain key feature information, specifically as follows:

[0016] Using the one-dimensional feature vector extracted by the basic feature extractor as input, a LightGBM classifier is constructed. Hyperparameters are set, and the average feature importance of the one-dimensional feature vector is calculated through multi-fold cross-validation. The vectors are sorted according to the average feature importance, and one-dimensional feature vectors that meet the preset importance criteria are retained to generate a filtered subset of one-dimensional feature vectors.

[0017] The REFCV method is initialized using the one-dimensional feature vector subset, and the LightGBM classifier is trained iteratively. The one-dimensional feature vector with the lowest current importance is removed according to a dynamic ratio that changes with the number of iterations. The F1 score of each iteration is recorded, and the F1 score is used to characterize the classification accuracy.

[0018] After iterative training, a curve is plotted based on the number of features and classification accuracy. The number of features at the inflection point is selected as the optimal number of features. A subset of the one-dimensional feature vectors output by the REFCV method is obtained as the search space. The mouse swarm algorithm is initialized to optimize key features.

[0019] One-dimensional feature vector combinations are randomly generated based on the optimal number of features, and chaotic processing is introduced to optimize the initial population. The fitness function is constructed using the AUC value of cross-validation, and the mouse population is arranged in descending order of fitness.

[0020] In the position update of individual mice, the Levy flight strategy is introduced to optimize the movement step size, and random perturbation is introduced to avoid premature convergence. The top 5 solutions with fitness are retained in each generation to enter the next generation. Single feature perturbation and neighborhood search are performed on dominant individuals to enhance local search.

[0021] When the termination condition of the mouse swarm algorithm is met, the optimal one-dimensional feature vector combination is obtained based on the optimal solution and used as key feature information.

[0022] In this scheme, the dimensionality-reduced one-dimensional feature vector is mapped to different feature spaces, and different feature representations are generated through reconstruction, specifically as follows:

[0023] The one-dimensional feature vector after feature selection and dimensionality reduction is mapped to the frequency domain, wavelet domain, and graph structure domain. In the frequency domain mapping branch, the one-dimensional feature vector is reconstructed to generate a discrete signal sequence. Fast Fourier Transform is performed on the discrete signal sequence, and low-frequency components are extracted by variational mode decomposition. Spectral entropy and main frequency components are extracted from the low-frequency components as frequency domain features.

[0024] In the wavelet domain mapping branch, wavelet transform is performed on the one-dimensional feature components to extract the energy proportion of each layer in the wavelet decomposition and calculate the wavelet entropy as the wavelet domain feature.

[0025] In the graph structure domain, each one-dimensional feature component is used as a graph node. The Pearson correlation coefficient between the one-dimensional feature components is calculated to obtain the correlation and construct the edge structure between the nodes. A topological relationship graph is constructed based on the graph nodes and edge structure. The node centrality and average clustering coefficient in the topological relationship graph are calculated as graph structure features.

[0026] In this scheme, different feature representations are weighted and fused to construct multimodal features. An attention mechanism is then used to convert these multimodal features into feature maps to construct a biological cell profile. Specifically:

[0027] The frequency domain features, wavelet domain features, and graph structure features are aligned and normalized. The features after cross-modal unified representation are adaptively weighted, and the different feature representations are spliced ​​and reconstructed using weight information to construct multimodal features.

[0028] The multimodal features are formatted and reshaped into 2D feature maps. Modality labels are embedded to mark the source of the features. Spatial attention processing is applied to the preprocessed features. Spatial weights are calculated through convolution to generate attention maps. Spatial weighting is performed using the attention maps to achieve spatial feature enhancement.

[0029] Simultaneously, a channel attention mechanism is used to calculate channel weights, a modal correlation matrix is ​​introduced for cross-modal interaction, and channel weights are updated to achieve channel feature enhancement through channel weighting.

[0030] After upsampling, the original feature map, spatial attention feature map, and channel attention feature map are stitched together, and a biological cell image is generated after feature compression.

[0031] In this scheme, a classification model is built based on a graph convolutional network, and a transfer learning strategy is used to pre-train the classification model, specifically:

[0032] A classification model for biological cell images is constructed based on graph convolutional networks. Local feature aggregation is performed in the graph convolutional layers through a neighbor aggregation mechanism, an attention mechanism is added, node importance weights are calculated to enhance key regions, and residual connections are used to prevent information loss in deep graph convolutional layers.

[0033] A publicly available cell image dataset is selected as the source dataset for pre-training. Cross-entropy loss is used as the loss function for the cell type classification task. When the loss function converges, the pre-trained classification model is output.

[0034] In this scheme, a pre-trained classification model is adapted to biological cell image data. The biological cell images to be classified are used as input to obtain the category probability of each biological cell image. Specifically:

[0035] Obtain a biological cell portrait from the image of the biological cell to be classified, treat each pixel region as a node, and use the multimodal features, spatial coordinates and morphological features corresponding to the pixel region as node features. Establish an edge structure based on spatial proximity and node feature similarity to generate a graph structure corresponding to the biological cell portrait.

[0036] The pre-trained classification model is adapted to the biological cell image data, and the target domain is fine-tuned. The node feature matrix of the graph structure corresponding to the biological cell image is input into the classification model to obtain the category probability distribution of each node and generate a cell substructure segmentation map.

[0037] By integrating the prediction results of each node through a voting mechanism, the category prediction results of the biological cell image to be classified are obtained and output.

[0038] The second aspect of this invention provides a biological cell image classification system, which implements a biological cell image classification method. The system includes an image acquisition and preprocessing module, a basic feature extraction module, a feature selection and dimensionality reduction module, a multimodal feature mapping module, a biological cell image construction module, a classification module, and a result output module.

[0039] The image acquisition and preprocessing module acquires images of biological cells to be classified and performs image preprocessing.

[0040] The basic feature extraction module pre-trains an AlexNet network, removes its fully connected classification layer, retains the convolutional and pooling layers, and uses the pre-trained AlexNet network to obtain a one-dimensional feature vector.

[0041] The feature selection and dimensionality reduction module filters and reduces the dimensionality of the acquired one-dimensional feature vector to obtain key features with discriminative power.

[0042] The multimodal feature mapping module maps the dimensionality-reduced one-dimensional feature vector to different feature spaces, and performs weighted fusion of the representations in different feature spaces to generate more discriminative multimodal features.

[0043] The biological cell profiling module uses an attention mechanism to convert the multimodal features into feature maps to construct biological cell profiling.

[0044] The classification module constructs a classification model based on a graph convolutional network and uses a transfer learning strategy to pre-train the classification model. The pre-trained classification model is then adapted to the biological cell image data, and the class probability of each biological cell image is output.

[0045] The result output module outputs the classification results of the biological cell images and transmits and displays them according to a preset method.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention significantly improves the accuracy, robustness, and generalization ability of biological cell image classification by combining deep learning with multimodal feature fusion, integrating AlexNet feature extraction, feature reconstruction, biological cell profiling, and graph convolutional network (GCN) classification. Specifically, multimodal feature fusion enhances information utilization and improves feature discrimination ability, while transfer learning strategies reduce data dependence and improve few-shot learning capabilities.

[0048] By using feature mapping and reconstruction, biological cell profiles (similar to heatmaps or topological maps) are generated, making classification decisions visible. This allows for intuitive observation of which regions (such as the cell nucleus and membrane structures) contribute the most to classification, helping biomedical experts verify the reliability of the model and assisting in pathological analysis. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0050] Figure 1 A flowchart of a biological cell image classification method is shown;

[0051] Figure 2 A flowchart illustrating the acquisition of key feature information in an embodiment is shown;

[0052] Figure 3 A flowchart illustrating the reconstruction of different feature representations in the embodiments is shown;

[0053] Figure 4 A block diagram of a biological cell image classification system is shown. Detailed Implementation

[0054] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0056] Figure 1 A flowchart of a biological cell image classification method is shown.

[0057] like Figure 1As shown, this embodiment provides a biological cell image classification method, including:

[0058] S102, acquire images of biological cells to be classified and preprocess them, using a pre-trained AlexNet network as the basic feature extractor to obtain one-dimensional feature vectors;

[0059] S104, perform feature selection and dimensionality reduction on the one-dimensional feature vector to obtain key feature information, map the dimensionality-reduced one-dimensional feature vector to different feature spaces, and generate different feature representations through reconstruction.

[0060] S106, weighted fusion of different feature representations to construct multimodal features, and an attention mechanism to convert the multimodal features into feature maps to construct a biological cell profile;

[0061] S108: A classification model is constructed based on a graph convolutional network, and a transfer learning strategy is used to pre-train the classification model. The pre-trained classification model is then adapted to biological cell image data. The biological cell images to be classified are used as input to obtain the category probability of each biological cell image.

[0062] It should be noted that images of different types of biological cells under different imaging conditions are collected, such as blood cells, cancer cells, and stem cells; the imaging conditions include fluorescence microscopy, confocal microscopy, etc.; a biological cell image set is constructed, and the biological cell image set is augmented using methods such as rotation, flipping, brightness adjustment, and random cropping to improve the model's generalization ability; the images are standardized or normalized to reduce the influence of factors such as illumination and contrast; and training and test sets are generated according to a preset ratio.

[0063] The AlexNet network was trained and validated using training and test sets. Since the receptive field of standard convolution is fixed, it is difficult to capture multi-scale structures in cell images, such as organelles of different sizes. Therefore, dilated convolution is introduced into the AlexNet network to replace standard convolution, which expands the receptive field without increasing the number of parameters, maintains the feature map resolution, and avoids information loss caused by early downsampling. It is suitable for cell image analysis that needs to take into account both global context and local details. The first two max-pooling layers are retained to ensure basic downsampling. The last max-pooling layer is replaced with a dilated convolution with a stride of 1 combined with global average pooling. The fully connected classification layer in the AlexNet network is removed to construct a basic feature extractor. Images of the biological cells to be classified, acquired under the current imaging conditions, are acquired and standardized. The size of the preprocessed images is dynamically adjusted and input into the basic feature extractor to extract multi-scale features. Low-level features include local features such as cell boundaries and nuclear membranes; mid-level features include subcellular structures such as nucleoli and mitochondria; and high-level features include global spatial relationships, such as cell arrangement patterns. Feature maps with different expansion rates are fused using skip connections. Global average pooling is used to average the fused feature maps to obtain a one-dimensional feature vector of the biological cell image to be classified. The basic feature extractor improves the detection rate of overlapping cells and small organelles, preserves the spatial information of cells, and meets the requirements of real-time analysis scenarios.

[0064] Figure 2 A flowchart illustrating the process of obtaining key feature information in an embodiment is shown.

[0065] According to an embodiment of the present invention, feature selection and dimensionality reduction are performed on the one-dimensional feature vector to obtain key feature information, specifically as follows:

[0066] S202, using the one-dimensional feature vector extracted by the basic feature extractor as input, construct a LightGBM classifier, set hyperparameters, calculate the average feature importance of the one-dimensional feature vector through multi-fold cross-validation, sort according to the average feature importance, retain one-dimensional feature vectors that meet the preset importance standard, and generate a filtered subset of one-dimensional feature vectors.

[0067] S204. Initialize the REFCV method using the one-dimensional feature vector subset, iteratively train the LightGBM classifier, remove the one-dimensional feature vector with the lowest current importance according to the dynamic ratio that changes with the number of iterations, record the F1 score of each iteration, and use the F1 score to characterize the classification accuracy.

[0068] S206. After iterative training, a curve is plotted based on the number of features and classification accuracy. The number of features at the inflection point is selected as the optimal number of features. A subset of the one-dimensional feature vectors output by the REFCV method is obtained as the search space. The mouse swarm algorithm is initialized to optimize key features.

[0069] S208, randomly generate a combination of one-dimensional feature vectors according to the optimal number of features, introduce chaotic processing to optimize the initial population, construct a fitness function using the AUC value of cross-validation, and arrange the mouse population in descending order of fitness.

[0070] S210 introduces the Levy flight strategy to optimize the movement step size in the position update of individual mice, and introduces random perturbation to avoid premature convergence. The top 5 solutions with fitness are retained in each generation to enter the next generation. Single feature perturbation and neighborhood search are performed on dominant individuals to enhance local search.

[0071] S212, when the termination condition of the mouse swarm algorithm is met, the best one-dimensional feature vector combination is obtained based on the optimal solution as the key feature information.

[0072] It should be noted that the LightGBM classifier is constructed by setting hyperparameters such as learning rate, tree depth, and feature score type. The average feature importance is calculated using 10-fold cross-validation, and features are sorted in descending order of cumulative contribution. Features with a cumulative contribution of 85% or higher are retained. Redundant features in the one-dimensional feature vector are eliminated using the LightGBM classifier. The REFCV method is initialized using the feature subset filtered by LightGBM. In each iteration, the LightGBM classifier is trained using 5-fold cross-validation, removing the 10% of features with the lowest current importance. The F1 score is recorded for each iteration. Iteration stops when the performance on the validation set decreases by more than 1% for three consecutive iterations. A curve of feature quantity versus classification accuracy is plotted, and the number of features at the inflection point is selected. The proportion of features removed in each iteration is dynamically adjusted using the REFCV method. The feature subset output by the REFCV method is used to initialize the mouse swarm algorithm. One-dimensional feature vector combinations are randomly generated based on the optimal number of features. Each feature combination is treated as an individual mouse position, and the optimal feature combination is found through swarm intelligence search. During iterative optimization, LightGBM is used to train the classification model, and the AUC value of the validation set is used as the fitness value. A penalty term is preferentially set based on the Pearson correlation coefficient between features; when the Pearson correlation coefficient between features exceeds a preset threshold, a deduction is made to avoid excessive redundancy in feature combinations. A Levy flight random step size is introduced in the position update of individual mice to provide long-distance jump capability, escaping local optima, and random perturbation is introduced to avoid premature convergence. In each iteration, an elite retention strategy is used to retain the top 5 solutions with the best fitness for the next generation, preventing the loss of high-quality feature combinations. For dominant individuals, 1-2 feature bits are randomly flipped for single-feature perturbation, and the Hamming distance between individual mice is evaluated for neighborhood search, achieving enhanced local search. One-dimensional feature vectors are processed into feature vectors, and through biomimetic intelligent optimization, the most diagnostically valuable feature combinations are efficiently extracted from massive image features while ensuring biomedical interpretability.

[0073] Figure 3 A flowchart illustrating the reconstruction of different feature representations in the embodiments is shown.

[0074] According to an embodiment of the present invention, the dimensionality-reduced one-dimensional feature vector is mapped to different feature spaces, and different feature representations are generated through reconstruction, specifically as follows:

[0075] S302, the one-dimensional feature vector after feature selection and dimensionality reduction is mapped to the frequency domain, wavelet domain and graph structure domain. In the frequency domain mapping branch, the one-dimensional feature vector is reconstructed into a signal to generate a discrete signal sequence. Fast Fourier transform is performed on the discrete signal sequence, and low-frequency components are extracted by variational mode decomposition. Spectral entropy and main frequency components are extracted from the low-frequency components as frequency domain features.

[0076] S304, In the wavelet domain mapping branch, perform wavelet transform on the one-dimensional feature components, extract the energy proportion of each layer in the wavelet decomposition, and calculate the wavelet entropy as the wavelet domain feature.

[0077] S306, in the graph structure domain, each one-dimensional feature component is used as a graph node, the Pearson correlation coefficient between the one-dimensional feature components is calculated to obtain the correlation and the edge structure between the nodes is constructed, and a topological relationship graph is constructed based on the graph nodes and edge structure. The node centrality and average clustering coefficient in the topological relationship graph are calculated as graph structure features.

[0078] It should be noted that in the frequency domain mapping branch, a Fast Fourier Transform is performed to obtain the complex spectrum. The analytical signals of each mode are calculated through Variational Mode Decomposition (VMD). Intrinsic Mode Functions (EMFs) are obtained through frequency center updates and mode reconstruction. Low-frequency components representing the aggregation of biological features are extracted. Spectral entropy and dominant frequency components are extracted from these low-frequency components as frequency domain features. The spectral entropy primarily measures feature complexity, and the dominant frequency component is obtained by collecting the maximum five peak frequencies. In the wavelet domain mapping branch, the Daubechies4 basis function is used for decomposition to obtain approximation coefficients and detail coefficients. The energy proportion of each layer is extracted based on these approximation coefficients and detail coefficients, and wavelet entropy is calculated as wavelet domain features. In the graph structure domain, a topological relationship graph is constructed based on the graph nodes and edge structure. The number of directly connected edges of a node represents its centrality. High-centrality nodes correspond to key morphological features in the cell image, and centrality reflects the topological importance of this feature in the cell structure. The clustering coefficient defines the degree of tightness between a node's neighbors. It finds all the neighboring nodes of a node, counts the number of edges in the neighboring node set, calculates the maximum possible number of edges, calculates the local clustering coefficient based on the node's neighbor data and the actual number of edges between the node's neighbors, integrates the local clustering coefficients and takes the average to obtain the average clustering coefficient, and quantifies the topological relationship between features.

[0079] It should be noted that the frequency domain features, wavelet domain features, and graph structure features are dimensionally aligned and normalized. The features after cross-modal unified representation are adaptively weighted. Initial weights are preset based on modal reliability. Adaptive weights are obtained through iterative updates and then normalized. Different feature representations are spliced ​​and reconstructed using the weight information to construct multimodal features. The multimodal features are formatted and reshaped into 2D feature maps. Modal labeling is then performed to mark the source of the features. For example, channels 1-3 correspond to frequency domain features, channels 4-6 correspond to wavelet features, and channel 7 corresponds to graph structure features. Spatial attention processing is applied to the preprocessed features. Spatial weights are calculated through convolution to generate an attention map. The attention map is then used for spatial weighting to enhance spatial features, with high-weight regions corresponding to key regions. Simultaneously, a channel attention mechanism is introduced. The SE-Net structure is used to calculate channel weights, and a learnable modal relevance matrix is ​​introduced for cross-modal interaction. Channel weights are updated to enhance channel features. After upsampling through transposed convolution, the original feature map, spatial attention feature map, and channel attention feature map are concatenated. After feature compression, dimensionality reduction is performed to generate a biological cell image.

[0080] It should be noted that the biological cell image classification model is constructed based on graph convolutional networks. In the graph convolutional layers, local feature aggregation is performed through a neighbor aggregation mechanism, an attention mechanism is added, node importance weights are calculated to enhance key regions, and residual connections are used to prevent information loss in deep graph convolutional layers. A publicly available cell image dataset is selected as the source dataset for pre-training, and cross-entropy loss is used as the loss function for the cell type classification task. When the loss function converges, the pre-trained classification model is output.

[0081] To obtain a biological cell profile from an image of a biological cell to be classified, each pixel region is treated as a node, and the multimodal features, spatial coordinates, and morphological features corresponding to the pixel region are used as node features. An edge structure is established based on spatial proximity and node feature similarity. For example, nodes within a radius of 10 pixels are fully connected, and nodes with a cosine similarity > 0.7 are used to establish edges, generating a graph structure corresponding to the biological cell profile. A pre-trained classification model is adapted to the biological cell profile data, and fine-tuning is performed for the target domain. The edge weights are adjusted according to the target data. The node feature matrix of the graph structure corresponding to the biological cell profile is input into the classification model to obtain the class probability distribution of each node, generating a cell substructure segmentation map. The prediction results of each node are integrated through a voting mechanism, with key nodes (attention weight > 0.9) having veto power. The class prediction results of the biological cell image to be classified are obtained and output.

[0082] It should be noted that the process involves collecting historical biological cell image recognition and classification examples, extracting the optimal one-dimensional feature vector combination selected from the historical biological cell images in the examples, obtaining the imaging conditions of the historical biological cell images, matching the imaging conditions with the optimal one-dimensional feature vector combination, and storing the optimal one-dimensional feature vector combination under different imaging conditions. The imaging conditions of the biological cell images to be classified are then obtained, and L1 norm is used for category matching based on the imaging conditions of the biological cell images to be classified. Based on the matching results, the corresponding optimal one-dimensional feature vector combination is extracted from the database as candidate key features. Each candidate key feature is used to train a LightGBM classifier, and the candidate key feature with the highest classification accuracy is extracted as the key feature of the biological cell image to be classified.

[0083] Figure 4 A block diagram of a biological cell image classification system is shown.

[0084] The second embodiment of the present invention provides a biological cell image classification system 4, which realizes a biological cell image classification method. The system includes an image acquisition and preprocessing module 401, a basic feature extraction module 402, a feature selection and dimensionality reduction module 403, a multimodal feature mapping module 404, a biological cell portrait construction module 405, a classification module 406, and a result output module 407.

[0085] The image acquisition and preprocessing module acquires images of biological cells to be classified and performs image preprocessing.

[0086] The basic feature extraction module pre-trains an AlexNet network, removes its fully connected classification layer, retains the convolutional and pooling layers, and uses the pre-trained AlexNet network to obtain a one-dimensional feature vector.

[0087] The feature selection and dimensionality reduction module filters and reduces the dimensionality of the acquired one-dimensional feature vector to obtain key features with discriminative power.

[0088] The multimodal feature mapping module maps the dimensionality-reduced one-dimensional feature vector to different feature spaces, and performs weighted fusion of the representations in different feature spaces to generate more discriminative multimodal features.

[0089] The biological cell profiling module uses an attention mechanism to convert the multimodal features into feature maps to construct biological cell profiling.

[0090] The classification module constructs a classification model based on a graph convolutional network and uses a transfer learning strategy to pre-train the classification model. The pre-trained classification model is then adapted to the biological cell image data, and the class probability of each biological cell image is output.

[0091] The result output module outputs the classification results of the biological cell images and transmits and displays them according to a preset method.

[0092] The third embodiment of the present invention provides a computer-readable storage medium including a biological cell image classification method program. When the biological cell image classification method program is executed by a processor, it implements the steps of a biological cell image classification method.

[0093] In the several embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection of modules may be electrical, mechanical, or other forms. Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes: mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, and other media capable of storing program code.

[0094] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A biological cell image classification method, characterized by, The method comprises the following steps: acquiring a biological cell image to be classified and preprocessing, using a pre-trained AlexNet network as a basic feature extractor to obtain a one-dimensional feature vector; performing feature selection and dimension reduction on the one-dimensional feature vector to obtain key feature information, mapping the one-dimensional feature vector after dimension reduction to different feature spaces, and generating different feature representations through reconstruction; performing weighted fusion on different feature representations to construct a multi-modal feature, converting the multi-modal feature into a feature map by using an attention mechanism to construct a biological cell portrait; constructing a classification model based on a graph convolution network, pre-training the classification model by using a transfer learning strategy, adapting the pre-trained classification model to biological cell portrait data, taking the biological cell portrait to be classified as input, and obtaining the class probability of each biological cell image; performing feature selection and dimension reduction on the one-dimensional feature vector to obtain key feature information, specifically: using the one-dimensional feature vector extracted by the basic feature extractor as input, constructing a LightGBM classifier, setting hyperparameters, calculating the average feature importance of the one-dimensional feature vector through multi-fold cross-validation, sorting according to the average feature importance, retaining one-dimensional feature vectors meeting a preset importance standard to generate a subset of one-dimensional feature vectors after screening; initializing the REFCV method using the subset of one-dimensional feature vectors, iteratively training the LightGBM classifier, removing the one-dimensional feature vector with the lowest current importance according to a dynamic proportion varying with the number of iterations, recording the F1 score of each iteration, and using the F1 score to represent the classification accuracy; after iterative training, drawing a curve according to the feature quantity and the classification accuracy, selecting the feature quantity at the inflection point as the best feature quantity, obtaining the subset of one-dimensional feature vectors output by the REFCV method as a search space, and initializing a swarm algorithm to optimize key features; randomly generating a one-dimensional feature vector combination according to the best feature quantity, introducing chaos processing to optimize the initial population, using the AUC value of cross-validation to construct a fitness function, and arranging the mouse population in descending order of fitness; introducing the Levy flight strategy to optimize the moving step in the position update of the mouse individual, introducing random disturbance to avoid premature convergence, retaining the Top 5 solutions with the highest fitness in each generation to enter the next generation, and performing single feature disturbance and neighborhood search on the superior individuals to enhance local search; when the termination condition of the swarm algorithm is reached, the best one-dimensional feature vector combination is obtained according to the optimal solution, which is used as the key feature information; mapping the one-dimensional feature vector after dimension reduction to different feature spaces to generate different feature representations, specifically: mapping the one-dimensional feature vector after feature selection and dimension reduction to the frequency domain, wavelet domain and graph structure domain, reconstructing the one-dimensional feature vector in the frequency domain mapping branch to generate a discrete signal sequence, performing fast Fourier transform on the discrete signal sequence, extracting low-frequency components through variational mode decomposition, and extracting frequency spectrum entropy and main frequency components from the low-frequency components as frequency domain features; In the wavelet domain mapping branch, the one-dimensional feature components are subjected to wavelet transform, the energy proportion of each layer in the wavelet decomposition is extracted, and the wavelet entropy is calculated as the wavelet domain feature; In the graph structure domain, each one-dimensional feature component is taken as a graph node, the Pearson correlation coefficient between the one-dimensional feature components is calculated to obtain the correlation, the edge structure between the nodes is constructed, the topological relationship graph is constructed according to the graph nodes and the edge structure, and the node centrality and the average clustering coefficient in the topological relationship graph are calculated as the graph structure feature; The different feature representations are weighted and fused to construct a multi-modal feature, and the multi-modal feature is converted into a feature map using an attention mechanism to construct a biological cell image, specifically: The frequency domain feature, the wavelet domain feature and the graph structure feature are aligned and normalized, the adaptive weighting is performed on the features after the cross-modal unified representation, the different feature representations are spliced and reconstructed using the weight information, and the multi-modal feature is constructed; The multi-modal feature is formatted and reshaped into a 2D feature map, and a modal label embedding is performed to mark the feature source, the spatial attention processing is performed on the preprocessed feature, the spatial weight is calculated through convolution, the attention map is generated, the spatial weighting is performed using the attention map to realize the spatial feature enhancement; At the same time, the channel attention mechanism is used to calculate the channel weight, the modal correlation matrix is introduced for cross-modal interaction, the channel weight is updated for channel weighting to realize the channel feature enhancement; After the upsampling operation, the original feature map, the spatial attention feature map and the channel attention feature map are spliced, and the biological cell image is generated after feature compression.

2. The biological cell image classification method of claim 1, wherein, An image of a biological cell to be classified is obtained and preprocessed, and a pre-trained AlexNet network is used as a basic feature extractor to obtain a one-dimensional feature vector, specifically: Different types of biological cell images under different imaging conditions are collected to construct a biological cell image set, the biological cell image set is subjected to data enhancement and normalization processing, and a training set and a test set are generated according to a preset proportion; The training set and the test set are used to train and verify the AlexNet network, a dilated convolution is introduced into the AlexNet network to replace a standard convolution, a dilated convolution with a step size of 1 is used in combination with a global average pooling to replace the last maximum pooling layer, and a fully connected classification layer in the AlexNet network is removed to construct a basic feature extractor; An image of a biological cell to be classified is obtained and preprocessed, and a pre-trained AlexNet network is used as a basic feature extractor to obtain a one-dimensional feature vector, specifically: A classification model is constructed based on a graph convolution network, and a transfer learning strategy is used to pretrain the classification model, specifically:

3. The biological cell image classification method of claim 1, wherein, ​ The classification model of biological cell images is constructed based on a graph convolution network, local feature aggregation is performed through a neighbor aggregation mechanism in a graph convolution layer, an attention mechanism is added, a node importance weight is calculated to enhance a key region, and a residual connection is used to prevent information loss of a deep graph convolution layer; A public cell image dataset is selected as a source dataset for pre-training, a cross-entropy loss is used as a loss function for a cell type classification task, and when the loss function converges, a pre-trained classification model is output.

4. The biological cell image classification method of claim 1, wherein, The pre-trained classification model is adapted to biological cell image data, biological cell images to be classified are input, and a class probability of each biological cell image is obtained, specifically as follows: The biological cell images to be classified are obtained, each pixel region is taken as a node, and the multi-modal features, spatial coordinates and morphological features corresponding to the pixel region are taken as node features, an edge structure is established according to spatial proximity and node feature similarity, and a graph structure corresponding to the biological cell image is generated; The pre-trained classification model is adapted to biological cell image data, target domain fine-tuning adaptation is performed, the node feature matrix of the graph structure corresponding to the biological cell image is input into the classification model, a class probability distribution of each node is obtained, and a cell substructure segmentation map is generated; The prediction results of each node are integrated through a voting mechanism, and a class prediction result of the biological cell image to be classified is obtained and output.

5. A biological cell image classification system, characterized by, The biological cell image classification method of any one of claims 1-4 is implemented, and the system comprises an image acquisition and preprocessing module, a basic feature extraction module, a feature selection and dimension reduction module, a multi-modal feature mapping module, a biological cell image construction module, a classification module and a result output module; The image acquisition and preprocessing module acquires biological cell images to be classified and performs image preprocessing; The basic feature extraction module pre-trains an AlexNet network, removes the full connection classification layer thereof, retains the convolution layer and the pooling layer, and uses the pre-trained AlexNet network to obtain a one-dimensional feature vector; The feature selection and dimension reduction module filters and reduces the dimension of the obtained one-dimensional feature vector to obtain key features with discriminability; The multi-modal feature mapping module maps the one-dimensional feature vector after dimension reduction to different feature spaces, weights and fuses the representations of different feature spaces, and generates more discriminative multi-modal features; The biological cell image construction module converts the multi-modal features into a feature map using an attention mechanism to construct a biological cell image; The classification module constructs a classification model based on a graph convolution network, pre-trains the classification model using a transfer learning strategy, adapts the pre-trained classification model to biological cell image data, and outputs a class probability of each biological cell image; The result output module outputs the classification result of the biological cell image and transmits and displays the classification result in a preset manner.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium comprises a biological cell image classification method program, and the biological cell image classification method program is executed by a processor to implement the steps of the biological cell image classification method of any one of claims 1-4.

Images