Bone marrow cell multi-scale identification method for blood disease classification
By constructing a bone marrow cell recognition model, the problems of insufficient sample size and insufficient feature extraction in bone marrow cell morphology test were solved, precise classification of cells and preliminary diagnosis of hematologic tumor diseases were achieved, and diagnostic efficiency was improved.
Patent Information
- Application Number
- CN202510619411.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-14
- Publication Date
- 2025-08-19
AI Technical Summary
The prior art has problems in bone marrow cell morphology tests, such as insufficient sample size, difficult to distinguish cell differences, insufficient feature extraction, and insufficient utilization of medical experts' knowledge, resulting in a high rate of misdiagnosis and missed diagnosis.
A bone marrow cell recognition model is constructed, including a layered feature extraction module, a feature enhancement module, a class-center learning framework module and a multi-layer perception machine. Through data augmentation, layered feature extraction, feature enhancement and class-center learning, combined with prior knowledge of medical experts, we can achieve accurate classification of cells.
It improves the accuracy and efficiency of cell classification, assists in the initial diagnosis of hematologic tumor diseases, and reduces the time of artificial microscopy.
Smart Images
Figure CN120510441A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a multi-scale bone marrow cell recognition method for blood disease classification. Background Art
[0002] Bone marrow cell morphology testing is an important examination method that assists in the diagnosis of blood system diseases and other related diseases by observing and analyzing the morphology of bone marrow cells. In the past, it mainly relied on inspectors to observe under a microscope, which was labor-intensive and time-consuming. In addition, different inspectors had large differences in the recognition of morphological characteristics such as cell color, size, and particle size, which made standardization difficult and prone to misdiagnosis and missed diagnosis. It is imperative to introduce artificial intelligence (AI) cell image recognition technology to assist in blood cell morphology testing. Artificial intelligence has the advantages of fast processing speed, high efficiency, easy storage and sharing of big data, and relative ease of standardization. Through the intervention of artificial intelligence cell image recognition technology, reducing labor costs and improving the efficiency and accuracy of manual microscopic examination of blood cells are important means to improve the quality of medical services.
[0003] Artificial intelligence cell image recognition technology is relatively mature in the application of peripheral blood cell morphology testing. In contrast, bone marrow cell morphology testing is often used to assist in the diagnosis of more complex blood system diseases and can provide more detailed and in-depth information. However, bone marrow aspirates are more complex cell specimens than peripheral blood smears. They contain a small number of areas suitable for cell detection, significant non-cellular debris, and more cell types. These cells often aggregate or overlap, making accurate AI-based bone marrow cell identification a challenging problem. The following problems exist in existing research on bone marrow cell morphology testing:
[0004] (1) Usually a small sample size is used for training, and the coverage of bone marrow cell types and disease categories is not comprehensive;
[0005] (2) In studies of fine-grained cell classification, the differences between cells are very small, making it difficult to effectively distinguish them;
[0006] (3) Single cell images can extract fewer effective features than other natural images, and research on effective feature extraction and feature enhancement is still insufficient;
[0007] (4) The method does not fully utilize the rich prior knowledge of medical experts on bone marrow cell morphology, which limits the performance of the method. Summary of the Invention
[0008] To address the above-mentioned issues, the present invention provides a multi-scale bone marrow cell recognition method for blood disease classification, comprising constructing and training a bone marrow cell recognition model, inputting an image to be recognized into the trained bone marrow cell recognition model to obtain a recognition result; the bone marrow cell recognition model comprises a hierarchical feature extraction module, a feature enhancement module, a class-centered learning framework module, and a multi-layer perceptron; and the training process of the bone marrow cell recognition model comprises the following steps:
[0009] S1. Obtain a set of bone marrow single cell images and preprocess them to obtain a qualified dataset. Annotate each image in the qualified dataset and divide it into a training set and a test set in proportion.
[0010] S2. Perform data augmentation on the training set to obtain an enhanced dataset, and combine the enhanced dataset with the training set to form an enhanced training dataset; the enhanced training dataset includes multiple input data, each input data including multiple single-cell images of a patient;
[0011] S3. The input data is passed through a hierarchical feature extraction module to obtain an initial feature sequence set; the initial feature sequence set includes multiple initial feature sequences;
[0012] S4. The initial feature sequence set is passed through the feature enhancement module to obtain an enhanced feature sequence set;
[0013] S5. Input the enhanced feature sequence set into the class center learning framework to obtain a class center distance result set, and calculate the cross entropy loss between each distance result in the class center distance result set and the true result;
[0014] S6. Perform statistical analysis based on the class center distance result set to obtain a prediction of the first patient's blood type;
[0015] S7. Aggregate all enhanced feature sequences in the enhanced feature sequence set using an attention mechanism to obtain bag-level aggregated features, input the aggregated features into a multi-layer perceptron to obtain a prediction of the second patient's blood type, and calculate a prediction loss based on the prediction of the second patient's blood type;
[0016] S8. Calculating a dual-branch prediction loss based on the first patient's blood type prediction and the second patient's blood type prediction;
[0017] S9. Optimize the model parameters according to the cross entropy loss, prediction loss, and dual-branch prediction loss until the model parameters converge.
[0018] Beneficial effects of the present invention:
[0019] The method proposed in the present invention can assist in the preliminary diagnosis of blood tumor diseases. In the present invention, a small number of cell type images are enhanced, and a hierarchical feature extraction module and a feature enhancement module are used to effectively retain spatial structural information and remove redundancy, and fully obtain more direct and effective features for classification; by introducing prior text knowledge and applying a class-centered learning mechanism, it can effectively distinguish between cell categories with slight differences and achieve accurate cell classification; by aggregating patient bone marrow cell features and statistically analyzing patient cell results, it can achieve classification prediction of blood diseases, and adjust model parameters through comparative learning strategies to optimize model performance. Compared with existing models, the present invention has improved the number of categories and accuracy of cell classification, and can assist in the preliminary diagnosis of blood tumor diseases and reduce manual microscopy time. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 Flow chart of the method of the present invention;
[0021] Figure 2 Schematic diagram of the model structure of the present invention;
[0022] Figure 3 This is a structural diagram of the hierarchical feature extraction module of the present invention;
[0023] Figure 4 This is the structural diagram of the feature enhancement module of the present invention, (a) space reconstruction unit, (b) channel reconstruction unit. DETAILED DESCRIPTION
[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0025] The present invention provides a multi-scale bone marrow cell recognition method for blood disease classification, including constructing and training a bone marrow cell recognition model, inputting the image to be recognized into the trained bone marrow cell recognition model to obtain a recognition result; the bone marrow cell recognition model includes a hierarchical feature extraction module, a feature enhancement module, a class center learning framework module and a multi-layer perceptron.
[0026] like Figure 1 、 Figure 2 As shown in Figure 2, the training process of the bone marrow cell recognition model includes the following steps:
[0027] S1. Obtain a set of bone marrow single cell images and preprocess them to obtain a qualified dataset. Annotate each image in the qualified dataset and divide it into a training set and a test set in proportion.
[0028] Specifically, in this embodiment of the present invention, a bone marrow fluid smear data set is collected for each patient, and multiple single-cell images are obtained from each bone marrow fluid smear data set. Each single-cell image is contrast-enhanced, resolution-enhanced, and normalized, and then unqualified single-cell images are filtered out to form a qualified data set. The final qualified data set includes cell image groups from different patients, each of which contains multiple single-cell images. Cell annotation is performed by experts with more than five years of experience in cell morphology testing, covering common blood cancer disease categories such as leukemia, lymphoma, and myeloma.
[0029] S2. Perform data augmentation on the training set to obtain an enhanced dataset, and combine the enhanced dataset with the training set to form an enhanced training dataset; the enhanced training dataset includes multiple input data, each input data including multiple single-cell images of a patient;
[0030] Specifically, data augmentation for the training set mainly involves enhancing a small number of single-cell images of patients. Specific operations include geometric transformation, color transformation, and mixup technology.
[0031] Geometric transformations include rotation, scaling, cropping, and flipping; these operations are simple and can quickly increase the amount of data.
[0032] Color transformation includes adjustments to brightness, contrast, and saturation, which can increase the diversity of images and improve the model's adaptability to different lighting conditions.
[0033] Mixup technology is a simple and effective data augmentation method that expands the data set by linear interpolation, which helps improve the generalization ability and robustness of the model. The specific implementation process of the Mixup operation is as follows: two single-cell images A and B, and their corresponding labels La and Lb, are randomly selected from the training set; an interpolation parameter λ is generated for these two single-cell images. λ is a random number between 0 and 1, which determines the contribution ratio of the two single-cell images A and B when generating a new single-cell image. If λ is close to 1, the new single-cell image will be closer to the single-cell image A. If λ is close to 0, the new single-cell image will be closer to the single-cell image B. According to the interpolation parameter λ, the pixel values of the two single-cell images A and B are linearly mixed to generate a new single-cell image. Each pixel value of the new single-cell image is λ times the corresponding pixel value of the single-cell image A plus (1-λ) times the corresponding pixel value of the single-cell image B. The label is one-hot encoded. Similarly, the labels La and Lb are linearly mixed according to the interpolation parameter λ. The new label Lnew = λLa + (1-λ)Lb. The one-hot encoding of the new label is generated by selecting the maximum index value in Lnew.
[0034] S3. The input data is passed through a hierarchical feature extraction module to obtain an initial feature sequence set; the initial feature sequence set includes multiple initial feature sequences.
[0035] Specifically, if Figure 3 As shown in the figure, the hierarchical feature module is constructed using CrossFormer++, which consists of 4 stages, each of which includes 1 cross-scale embedding layer (CEL), 3 CrossFormer blocks and 2 amplitude cooling layers (ACL); each CrossFormer block consists of a long-short distance attention (LSDA) module and a dynamic position bias (DPB) module.
[0036] The cross-scale embedding layer is used to generate input tokens for each stage. In the first stage, the cross-scale embedding layer receives the image as input and samples patches using four kernels of different sizes: 4×4, 8×8, 16×16, and 32×32. These four kernels have the same center but different sizes, with a consistent stride length to generate the same number of tokens, which are then concatenated into a single embedding vector. In stages 2-4, two kernel sizes are used to further extract features: 2×2 and 4×4, respectively. Setting the stride length to 2×2 reduces the number of generated embedding tokens to one-quarter that of the previous stage.
[0037] The long-short distance attention module reduces computational burden by splitting the self-attention module into short-distance attention and long-distance attention, while preserving both small-scale and large-scale features. The amplitude cooling layer, a layer that is even lighter than the cross-scale embedding layer, mainly consists of a depthwise separable convolution and a normalization layer (without residual connections). This layer effectively reduces the activation amplitude while reducing the introduced computational cost and avoiding the vanishing gradient problem.
[0038] Progressive Group Size (PGS) gradually increases the group size from shallow to deep layers to accommodate feature extraction requirements at different levels. In CrossFormer's self-attention map, shallow layer labels focus primarily on local regions, while deep layer labels tend to focus more globally. Different group sizes are set in the four stages of CrossFormer++, enabling shallow layers to better capture local features and deep layers to better capture global features. In this embodiment of the present invention, the group sizes for the four stages are set to 4, 8, 16, and 32, respectively.
[0039] Specifically, step S3 includes:
[0040] For each single-cell image in the input data, it is input into the hierarchical feature module, and the initial feature sequence including different levels of features is obtained through the hierarchical information bottleneck strategy;
[0041] The initial feature sequence sets are obtained by combining the initial feature sequences of all single-cell images in the input data.
[0042] The hierarchical information bottleneck strategy is a method based on information theory. It aims to optimize the hierarchical feature extraction process through information theory constraints. It is mainly divided into three parts: retaining task-related features by maximizing the mutual information between hierarchical features and corresponding labels; filtering redundant information by minimizing the mutual information between hierarchical features and input data; and disentangling inter-hierarchical features by minimizing the mutual information between features at different levels.
[0043] S4. The initial feature sequence set is passed through the feature enhancement module to obtain an enhanced feature sequence set.
[0044] Specifically, the feature enhancement module is constructed using SCConv, such as Figure 4 As shown in Figure 1, it consists of a spatial reconstruction unit (SRU) and a channel reconstruction unit (CRU). The SRU uses a separation-reconstruction approach to suppress spatial redundancy, while the CRU uses a split-transform-fusion strategy to reduce channel redundancy. The SRU and CRU units are arranged sequentially. The input feature X first passes through the SRU to obtain the spatially refined feature Xw; then passes through the CRU to obtain the channel-refined feature as the output.
[0045] In the present invention, each initial feature sequence in the initial feature sequence set is input into the feature enhancement module to obtain the corresponding enhanced feature sequence to form an enhanced feature sequence set.
[0046] S5. Input the enhanced feature sequence set into the class center learning framework to obtain the class center distance result set, and calculate the cross entropy loss between each distance result in the class center distance result set and the true result.
[0047] Specifically, the initialization operation of the class center learning framework module can directly reflect the prior distribution differences between categories, reducing the blind exploration of the class center learning framework module in the early stage of training; if it is randomly initialized, it may cause the class center to fall into a suboptimal distribution, making the model unable to effectively separate categories. The introduction of prior knowledge can be converted into semantic constraints in the feature space; for a small amount of category sample data, adding text data can reduce the interference of noise; in addition, the introduction of text data can avoid the class center from over-fitting the training samples and improve the generalization of the model. However, if you only rely on text data for initialization, the text data may not be able to fully reflect the cell morphological characteristics, resulting in deviations in the initialized class center. Secondly, there are differences between the text features and image features extracted by the NLP model. If no prior alignment is performed, the initialization may be invalid. Therefore, the present invention chooses to add image features in the process of class center initialization.
[0048] The present invention initializes the class-centered learning framework module through prior knowledge text, and the process includes:
[0049] S51. Collect prior knowledge text, including WHO blood disease classification guidelines, cell morphology descriptions, key attributes of cell types, and public knowledge bases;
[0050] S52. Input the prior knowledge text into the pre-trained BioBERT to extract the key attributes in the prior knowledge text, convert the prior knowledge text into a structured table, and then generate a text feature vector for the text description; use PCA to reduce the dimension of the text feature vector to the same space as the image feature, specifically expressed as
[0051]
[0052] Among them, Proj(·) is a linear projection layer that matches the text feature dimension to the image feature space. c represents the image features (in this invention, it refers to the features obtained by passing the single cell image through the hierarchical feature module), T c represents the text feature vector, and α is the control factor that controls the text prior strength. It represents the feature space anchor point that combines text prior values and image statistical laws, and directly affects the position of the classification boundary.
[0053] S53. Perform K-means clustering on the reduced text feature vector to obtain clustering results, where the number of clusters equals the number of cell classes. Use the silhouette coefficient to assess clustering quality to ensure that the text features are separable.
[0054] Specifically, the enhanced feature sequence set is input into the class center learning framework, and for each enhanced feature sequence in the enhanced feature sequence set, the distance result between it and the class center parameter of each cell category is calculated, and finally the class center distance result set is obtained; each distance result in the class center distance result set is expressed as
[0055]
[0056] Among them, D n,i represents the n=1,2,…,N enhanced feature sequence x n The distance between the class center parameters of the i=1, 2, ..., C cell categories, C represents the number of cell categories, N represents the number of enhanced feature sequences; M represents the length of the class center parameter, W i represents the weight matrix corresponding to the i-th cell category; σ 2 Represents the variance parameter of the Gaussian kernel, that is, the degree of influence of the distance between the sample point and the category center parameter on the distance metric; T represents the transposition operation, Represents the square of Euclidean arithmetic.
[0057] Calculate the cross entropy loss L1 between each distance result and the true result, expressed as
[0058]
[0059] Among them, y n,i represents the true probability that the nth enhanced feature sequence belongs to the i-th cell category.
[0060] Specifically, the weight matrix and center parameters can be updated during training, and the class center parameters and weight matrix are updated once for each training batch. The weight matrix is updated by gradient descent of the cross entropy loss L1, and then the updated weight matrix is used to update the class center parameters of each cell class. The calculation formula is as follows:
[0061]
[0062] in, represents the class center parameter of the i-th cell class in the t-th iteration training; γ represents the learning rate, which is used to update the step size during training; β is a hyperparameter used to control the intensity of inter-class competition; represents the updated weight matrix, that is, the weight matrix of the i-th cell category in the t-th iteration training; F i represents the average of all enhanced feature sequences belonging to the i-th cell class in a batch. The second term in the formula represents the intra-class aggregation term, which pushes the class center closer to the feature distribution of cell class i in the current batch. The third term represents the inter-class separation term, which is independent of the samples in the current batch and only calculates the repulsive force based on the class centers of other cell classes, forcing the class center of cell class i to move away from the centers of other classes, thereby enhancing inter-class separability.
[0063] S6. Perform statistical analysis based on the class center distance result set to obtain a prediction of the first patient's blood type.
[0064] Specifically, based on the class center distance result set obtained by the class center learning framework, that is, the cell prediction results of each single-cell image of the patient, the proportion of each cell type and the total number of bone marrow nucleated cells are counted; the cell morphological characteristics are obtained according to the feature extraction module, and the mean and variance of various cell morphological parameters are counted; based on the patient's cell number and cell morphological parameter statistical results, combined with the WHO blood disease diagnostic standards, blood disease classification is realized to obtain the patient's first patient blood type prediction.
[0065] S7. Aggregate each enhanced feature sequence in the enhanced feature sequence set through the attention mechanism to obtain bag-level aggregated features, input the aggregated features into the multi-layer perceptron to obtain the second patient's blood type prediction, and calculate the prediction loss based on the second patient's blood type prediction.
[0066] Specifically, all enhanced feature sequences are aggregated through the attention mechanism to obtain a bag-level aggregate feature, which is expressed as
[0067]
[0068] Where Z represents the aggregated feature, Y represents the enhanced feature sequence set, A represents the attention matrix, and Transpose() represents transposition. The present invention uses a class-wise attention network to calculate the attention matrix A. The class-wise attention network assigns an attention weight to each cell sample instance, indicating the importance of the instance to each category, specifically expressed as
[0069] A=softmax(W(tanh(U(Y))⊙sigmoid(V(Y))))
[0070] Among them, U(), V(), W() represent fully connected layers, sigmoid() represents activation function, tanh() represents activation function, softmax() represents activation function, ⊙ represents element-by-element multiplication, Represents matrix multiplication.
[0071] Specifically, all aggregated features are input into the multilayer perceptron to obtain the corresponding second patient blood type prediction. The prediction loss is calculated based on each second patient blood type prediction, which is expressed as
[0072]
[0073] Where B represents the number of patients (number of samples) in each training batch. represents the blood type prediction of the second patient corresponding to the kth sample, Y k Indicates the true type corresponding to the kth sample.
[0074] S8. Calculate the dual-branch prediction loss based on the first patient blood type prediction and the second patient blood type prediction.
[0075] Specifically, the dual-branch prediction loss is expressed as (calculating this loss is to constrain the prediction difference between the two branches)
[0076]
[0077] Among them, B represents the number of samples in each training batch, represents the first blood prediction type of the k-th sample, and CE() represents the cross entropy loss.
[0078] S9. Optimize the model parameters according to the cross entropy loss, prediction loss, and dual-branch prediction loss until the model parameters converge.
[0079] The gradients of cross entropy loss, prediction loss and dual-branch prediction loss are calculated independently, and parameter updates are achieved through back propagation and gradient accumulation.
[0080] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "rotation" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.
[0081] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A multi-scale identification method for bone marrow cells for blood disease classification, characterized by: Constructing and training a bone marrow cell recognition model, inputting the image to be recognized into the trained bone marrow cell recognition model to obtain a recognition result; the bone marrow cell recognition model includes a hierarchical feature extraction module, a feature enhancement module, a class center learning framework module and a multi-layer perceptron; The training process of the bone marrow cell recognition model includes the following steps: S1. Obtain a set of bone marrow single cell images and preprocess them to obtain a qualified dataset. Annotate each image in the qualified dataset and divide it into a training set and a test set in proportion. S2. Perform data augmentation on the training set to obtain an enhanced dataset, and combine the enhanced dataset with the training set to form an enhanced training dataset; the enhanced training dataset includes multiple input data, each input data including multiple single-cell images of a patient; S3. The input data is passed through a hierarchical feature extraction module to obtain an initial feature sequence set; the initial feature sequence set includes multiple initial feature sequences; S4. The initial feature sequence set is passed through the feature enhancement module to obtain an enhanced feature sequence set; S5. Input the enhanced feature sequence set into the class center learning framework to obtain a class center distance result set, and calculate the cross entropy loss between each distance result in the class center distance result set and the true result; S6. Perform statistical analysis based on the class center distance result set to obtain a prediction of the first patient's blood type; S7. Aggregate all enhanced feature sequences in the enhanced feature sequence set through an attention mechanism to obtain bag-level aggregated features, input the aggregated features into a multi-layer perceptron to obtain a prediction of the second patient's blood type, and calculate a prediction loss based on the prediction of the second patient's blood type; S8. Calculating a dual-branch prediction loss based on the first patient's blood type prediction and the second patient's blood type prediction; S9. Optimize the model parameters according to the cross entropy loss, prediction loss, and dual-branch prediction loss until the model parameters converge.
2. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 1, characterized in that: The hierarchical feature module is built using CrossFormer++, which consists of 4 stages, each of which includes 1 cross-scale embedding layer, 3 CrossFormer blocks and 2 amplitude cooling layers; each CrossFormer block consists of a long-short distance attention module and a dynamic position deviation module.
3. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 1, characterized in that: Step S3 specifically includes: For each single-cell image in the input data, it is input into the hierarchical feature module, and the initial feature sequence including different levels of features is obtained through the hierarchical information bottleneck strategy; The initial feature sequence set is obtained by combining the initial feature sequences of all single-cell images in the input data; The hierarchical information bottleneck strategy includes: maximizing the mutual information between hierarchical features and corresponding labels, minimizing the mutual information between hierarchical features and input data, and minimizing the mutual information between features at different levels.
4. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 1, characterized in that: The feature enhancement module is constructed using SCConv, which includes a spatial reconstruction unit and a channel reconstruction unit.
5. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 1, characterized in that: The initialization process of the class-centered learning framework module includes: S51. Collect prior knowledge text, including WHO blood disease classification guidelines, cell morphology descriptions, key attributes of cell types, and public knowledge bases; S52. Input the prior knowledge text into the pre-trained BioBERT to obtain the text feature vector; use PCA to reduce the dimension of the text feature vector; reduce the dimension to the same space as the image feature, specifically expressed as Among them, Proj(·) is the linear projection layer, μ c represents the image features, T c represents the text feature vector, α is the control factor; Represents the feature space anchor point that combines text prior values and image statistical laws; S53. Perform K-means clustering on the text feature vector after dimensionality reduction to obtain the clustering result.
6. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 1, characterized in that: The enhanced feature sequence set is input into the class center learning framework. For each enhanced feature sequence in the enhanced feature sequence set, the distance result between it and the class center parameter of each cell category is calculated, and finally the class center distance result set is obtained. Each distance result in the class center distance result set is expressed as Among them, D n,i represents the n=1,2,…,N enhanced feature sequence x n The distance between the class center parameters of the i=1, 2, ..., C cell categories, C represents the number of cell categories, N represents the number of enhanced feature sequences; M represents the length of the class center parameter, W i represents the weight matrix corresponding to the i-th cell category, σ 2 represents the variance parameter of the Gaussian kernel; T represents the transpose operation, represents the square of Euclidean arithmetic; Calculate the cross entropy loss L1 between each distance result and the true result, expressed as Among them, y n,i represents the true probability that the nth enhanced feature sequence belongs to the i-th cell category.
7. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 6, characterized in that: During the training process, the class center parameters and weight matrix are updated once for each training batch. The weight matrix is updated by gradient descent of the cross entropy loss, and then the class center parameters are updated using the updated weight matrix. The calculation formula is as follows: in, represents the class center parameter of the i-th cell category in the t-th iteration training; γ represents the learning rate; β is a hyperparameter, represents the updated weight matrix; F i represents the average value of all enhanced feature sequences belonging to the i-th cell category in a batch, and C represents the number of cell categories.
8. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 1, characterized in that: Aggregate any enhanced feature sequence through the attention mechanism to obtain the bag-level aggregated features, expressed as A=softmax(W(tanh(U(Y))⊙sigmoid(V(Y)))) Among them, Z represents the aggregated feature, Y represents the enhanced feature sequence, A represents the attention matrix, U(), V(), W() represent the fully connected layer, sigmoid() represents the activation function, tanh() represents the activation function, softmax() represents the activation function, Transpose() represents the transposition, ⊙ represents the element-by-element multiplication, Represents matrix multiplication.
9. The multi-scale bone marrow cell recognition method for blood disease classification according to claim 1, characterized in that: The prediction loss is calculated based on the second patient's blood type prediction, which is expressed as Among them, B represents the number of samples in each training batch, Denotes the blood type prediction of the second patient of the kth sample, Y k Indicates the true type corresponding to the kth sample.