Leukemia bone marrow cell microscopic image classification method based on deep learning

Abstract

The application discloses a leukemia bone marrow cell microscopic image classification method based on deep learning and belongs to the technical field of medical image processing. The method comprises the following steps: acquiring bone marrow cell microscopic image data, positioning a cell nucleus region and dividing a cytoplasm to obtain single cell image data; performing multi-scale feature extraction on the single cell image to obtain cell nucleus chromatin texture features and cytoplasm particle distribution features; inputting the two types of features into a pre-constructed graph convolution network to encode cell adjacency relationships and obtain cell spatial topological relationship features; performing cross-modal alignment and splicing on the topological features, the texture features and the particle features to obtain fused cell representation data; and based on the fused representation data, using a multi-level category discrimination network to perform leukemia cell type probability prediction and determining a classification result. The application fuses cell morphological features and spatial topological relationships, and improves the accuracy and interpretability of bone marrow cell classification.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically a deep learning-based method for classifying microscopic images of leukemia bone marrow cells. Background Technology

[0002] Classification of leukemia bone marrow cell microscopic images relies on accurate cell morphology identification. Traditional methods extract visual features such as texture, shape, and granule distribution of individual cells through image processing techniques, and then use classifiers such as support vector machines to determine cell type. This type of method treats each cell as an independent individual, focusing only on the microscopic image representation of the cell itself. However, bone marrow smears contain a large number of cell populations at different differentiation stages and of different lineages. The spatial arrangement and adjacency relationships between cells carry important information about the state of the hematopoietic microenvironment and pathological aggregation. Ignoring the spatial relationships between cells prevents the classification model from perceiving the abnormal distribution of cell populations caused by leukemia infiltration, limiting the improvement of discrimination accuracy. In addition, the texture of cell nuclear chromatin, the distribution of cytoplasmic granules, and the topological features abstracted from spatial relationships belong to different modalities, with significant differences in feature representation scale and distribution. Existing methods often use simple vector concatenation to combine multi-source features without explicit alignment of cross-modal features. Different modal features are prone to information conflicts or scale mismatches in joint representation, affecting the discriminative completeness of fused cell representations. The problem that needs to be solved is how to extract the spatial topological relationships of cell populations from bone marrow cell microscopic images and effectively integrate them with single-cell morphological features. Summary of the Invention

[0003] This paper presents a deep learning-based method for classifying leukemia bone marrow cell microscopic images. By performing graph structure modeling on the adjacency relationships between cells in bone marrow cell microscopic images, the method obtains the spatial topological relationship features of cells and achieves cross-modal alignment and fusion with the multi-scale morphological features of single cells, thereby improving the accuracy of leukemia cell type identification.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a deep learning-based method for classifying leukemia bone marrow cell microscopic images. This method includes: acquiring bone marrow cell microscopic image data, and performing nuclear region localization and cytoplasmic region segmentation on the bone marrow cell microscopic image data to obtain single-cell image data; performing multi-scale feature extraction on the single-cell image data to obtain nuclear chromatin texture features and cytoplasmic particle distribution features; inputting the nuclear chromatin texture features and cytoplasmic particle distribution features into a pre-constructed graph convolutional network for cell adjacency relationship encoding to obtain cell spatial topological relationship features; performing cross-modal feature alignment and splicing of the cell spatial topological relationship features with the nuclear chromatin texture features and cytoplasmic particle distribution features to obtain fused cell characterization data; based on the fused cell characterization data, using a multi-level class discriminant network to predict the probability of leukemia cell types, obtaining probability distribution data for each category, and determining the classification result of the bone marrow cell microscopic image based on the probability distribution data for each category.

[0005] As a preferred embodiment of the present invention, the process of acquiring single-cell image data specifically includes: performing grayscale and contrast normalization processing on bone marrow cell microscopic image data to obtain standardized grayscale image data; inputting the standardized grayscale image data into a pre-trained target detection network for nucleus center point localization to obtain nucleus center point coordinate data and nucleus bounding box data; based on the nucleus center point coordinate data and nucleus bounding box data, using a region growing algorithm to determine the boundary of the cytoplasmic region corresponding to each nucleus to obtain cytoplasmic region segmentation data; and based on the nucleus bounding box data and cytoplasmic region segmentation data, cropping an image region containing a complete single cell from the bone marrow cell microscopic image data to obtain single-cell image data. This method can accurately separate single-cell regions, effectively remove interference caused by the overlap of image background and cells, and provide clean cell units for subsequent analysis.

[0006] Before performing multi-scale feature extraction on single-cell image data, preferably, a step of color normalization is also included. This step specifically involves: calculating the channel mean and standard deviation for the red, green, and blue channels of the single-cell image data to obtain statistical parameter data for each channel; based on the statistical parameter data for each channel, subtracting the channel mean of the corresponding channel from the channel value of each pixel in the single-cell image data and then dividing by the channel standard deviation of the corresponding channel to obtain channel-normalized pixel data; multiplying the channel-normalized pixel data by the target red channel mean, target green channel mean, and target blue channel mean, and then adding the target red channel standard deviation, target green channel standard deviation, and target blue channel standard deviation respectively, to obtain the color-normalized single-cell image data. Color normalization effectively eliminates image color differences introduced under different slide preparation and staining conditions, improving the consistency and stability of feature extraction.

[0007] In the multi-scale feature extraction operation, preferably, the nucleus region in the single-cell image data is encoded using local binary patterns at different scales to obtain multi-scale texture encoded data; histogram statistics and concatenation are performed on the multi-scale texture encoded data to obtain the chromatin texture features of the nucleus; simultaneously, the oriented gradient histogram is calculated for the cytoplasm region in the single-cell image data to obtain the cytoplasmic gradient amplitude distribution data; block statistics and normalization are performed on the cytoplasmic gradient amplitude distribution data to obtain the cytoplasmic granule distribution features. Through multi-scale analysis and complementary feature design, the subtle texture changes of chromatin and the distribution morphology of cytoplasmic granules can be captured from different receptive fields, providing richer morphological evidence for distinguishing various types of leukemia cells.

[0008] In the preferred scheme for obtaining cell spatial topological relationship features, the Euclidean distance between any two cell nucleus center points is calculated based on the coordinates of the cell nucleus center point in the single-cell image data, resulting in inter-cell distance matrix data. Cell pairs with distances less than a preset distance threshold in the inter-cell distance matrix data are identified as adjacent cell pairs, and a cell adjacency graph is constructed based on these adjacent cell pairs, where nodes represent single cells and edges represent adjacency relationships. Cell nucleus chromatin texture features and cytoplasmic particle distribution features are used as initial node features, and multi-level neighborhood feature aggregation is performed through a graph convolutional network to obtain updated node feature data. Global pooling is then performed on the updated node feature data to obtain the cell spatial topological relationship features. This scheme enables the network to learn the spatial distribution patterns and organizational structures between cells, significantly aiding in identifying the pathological features of leukemia reflected by abnormal cell arrangement.

[0009] To fully integrate multi-source features, cross-modal feature alignment and stitching are performed. A first linear projection layer maps nuclear chromatin texture features to a common feature space, yielding first projected feature data. A second linear projection layer maps cytoplasmic particle distribution features to the common feature space, yielding second projected feature data. A third linear projection layer maps cell spatial topological relationship features to the common feature space, yielding third projected feature data. Element-wise addition and layer normalization are then performed on the first, second, and third projected feature data to obtain fused cell characterization data. Addition and fusion within the common feature space allows for mutual calibration and enhancement of feature information from different modalities, resulting in more discriminative overall cell expression.

[0010] When using a multi-level classifying network for classification, the network includes a coarse-grained discriminant subnetwork and a fine-grained discriminant subnetwork. Fusion cell characterization data is input into the coarse-grained discriminant subnetwork for class assignment prediction, yielding probability data for normal cells, reactive cells, and tumor cells. When the probability data for the tumor cell class exceeds a preset class threshold, the fusion cell characterization data is input into the fine-grained discriminant subnetwork for subtype assignment prediction, yielding probability data for each leukemia subtype. When the probability data for either the normal or reactive cell class exceeds a preset class threshold, the fusion cell characterization data is input into the fine-grained discriminant subnetwork for normal or reactive subtype assignment prediction, yielding probability data for each normal or reactive subtype. Based on the probability data for each leukemia subtype, normal subtype, or reactive subtype, the probability distribution data for each class is determined. This progressive discrimination method, from coarse to fine, ensures rapid screening of broad categories while also enabling precise subtype differentiation of key tumor cells and other categories, thus improving the accuracy and efficiency of classification.

[0011] Preferably, the multi-level class discriminant network performs feature dimensionality reduction based on the fused cell characterization data: the fused cell characterization data is sequentially input into the first and second fully connected layers of the multi-level class discriminant network for feature dimensionality reduction, resulting in dimensionality-reduced feature data; the dimensionality-reduced feature data is then input into the third fully connected layer for class score calculation, resulting in initial score data for each class; the initial score data for each class is then processed by a normalized exponential function mapping to obtain probability distribution data for each class; and the class with the highest probability value from the probability distribution data is selected as the classification result for the bone marrow cell microscopic image. Through hierarchical fully connected mapping and normalization, the compressed features can be effectively converted into confidence scores for each class, achieving stable probability output.

[0012] As a preferred embodiment of the present invention, the training process of the pre-constructed graph convolutional network is as follows: Training sample data of bone marrow cell microscopic images and corresponding cell type labeling data are acquired; the nucleus region is located and the cytoplasm region is divided in the bone marrow cell microscopic image training sample data to obtain single-cell training image data; multi-scale feature extraction is performed on the single-cell training image data to obtain training cell nucleus chromatin texture features and training cytoplasmic particle distribution features; based on the training cell nucleus chromatin texture features and training cytoplasmic particle distribution features, combined with the cell nucleus center point coordinate data in the single-cell training image data, training cell adjacency graph data is constructed; using the training cell adjacency graph data as input and the cell type labeling data as a supervision signal, the network parameters of the graph convolutional network are iteratively updated using the cross-entropy loss function to obtain the trained graph convolutional network. Furthermore, different weight coefficients are set for different cell type categories in the cross-entropy loss function, and these weight coefficients are inversely proportional to the frequency of occurrence of each category in the training samples. Using this weighted loss function can significantly alleviate the problem of imbalanced sample classes, prevent the model from favoring high-frequency classes, and improve the detection capability of rare leukemia cell types.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By acquiring single-cell image data from bone marrow cell microscopic images and extracting nuclear chromatin texture features and cytoplasmic granule distribution features, these features are used as initial node features. Cell distances are calculated based on the coordinates of the cell nucleus center point, and a cell adjacency graph is constructed. Then, a graph convolutional network is used to aggregate multi-level neighborhood features to obtain the spatial topological relationship features of the cells. This approach expands the analysis object of microscopic images from isolated single cells to cell populations. The graph convolutional network can perceive the aggregation patterns, density changes, and spatial proximity relationships of abnormal cells in specific regions within a bone marrow smear. The extracted spatial topological relationship features directly correspond to the spatial distribution abnormalities formed by the infiltration and replacement of normal hematopoietic cells by leukemia cells in the bone marrow microenvironment. This allows subsequent discrimination processes to obtain pathological clues at the population level beyond single-cell morphological information, overcoming the shortcomings of traditional methods that rely solely on single-cell visual features and miss spatial relationship information.

[0014] Cellular nuclear chromatin texture features, cytoplasmic particle distribution features, and cell spatial topological relationship features are mapped to a common feature space through their respective corresponding linear projection layers, resulting in first, second, and third projection features with consistent dimensions. Element-wise addition and layer normalization are then performed on these three features to obtain fused cell representation data. This cross-modal feature alignment and stitching method unifies heterogeneous features from the nuclear region, cytoplasmic region, and intercellular topological relationships into a comparable representation space, eliminating differences in numerical magnitude and distribution characteristics between different modal features. This avoids the biased influence of a particular modal feature's scale dominance on the fusion result during direct stitching. After stabilizing the feature distribution through layer normalization, the fused cell representation data can fully retain complementary discriminative information from internal cell texture, particle structure, and external spatial positional relationships. This allows subsequent multi-level classification networks to accurately infer cell types based on richer fused features, especially for cell subtypes with similar morphological appearances but different spatial distribution patterns, demonstrating stronger identification capabilities. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of a deep learning-based method for classifying leukemia bone marrow cell microscopic images. Figure 2 This is a flowchart of bone marrow cell image feature extraction. Figure 3 This is a flowchart of the cell space topological relationship feature extraction process; Figure 4 This is a schematic diagram of multi-source cell feature projection fusion and layer normalization processing; Figure 5 This is a flowchart of a bone marrow cell classification method based on graph convolutional networks. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1This invention provides a deep learning-based method for classifying leukemia bone marrow cell microscopic images, comprising: acquiring bone marrow cell microscopic image data, and performing nucleus region localization and cytoplasmic region segmentation on the bone marrow cell microscopic image data to obtain single-cell image data; performing multi-scale feature extraction on the single-cell image data to obtain nucleus chromatin texture features and cytoplasmic particle distribution features; inputting the nucleus chromatin texture features and cytoplasmic particle distribution features into a pre-constructed graph convolutional network for cell adjacency relationship encoding to obtain cell spatial topological relationship features; performing cross-modal feature alignment and splicing of the cell spatial topological relationship features with the nucleus chromatin texture features and cytoplasmic particle distribution features to obtain fused cell characterization data; based on the fused cell characterization data, using a multi-level class discrimination network to predict the probability of leukemia cell types to obtain probability distribution data for each category, and determining the classification result of the bone marrow cell microscopic image based on the probability distribution data for each category.

[0019] Example 1

[0020] In specific implementation, please refer to Figure 2 The bone marrow cell microscopic image data were subjected to grayscale conversion and contrast normalization to obtain standardized grayscale image data. The grayscale conversion was performed using the following formula:

[0021] in, Represents the pixel after grayscale conversion grayscale value, Represents pixels in bone marrow cell microscopic image data Intensity values ​​in the red channel, Represents pixels in bone marrow cell microscopic image data In terms of the intensity value of the green channel, Represents pixels in bone marrow cell microscopic image data In the blue channel intensity values, coefficient 0.299 is the weighting coefficient for the red channel, coefficient 0.587 is the weighting coefficient for the green channel, and coefficient 0.114 is the weighting coefficient for the blue channel. These three weighting coefficients are determined according to the ITU-R BT.601 standard. The grayscale image after grayscale conversion undergoes contrast normalization. The mean and standard deviation of the grayscale values ​​of all pixels in the grayscale image are calculated. The grayscale value of each pixel is subtracted from the mean and then divided by the standard deviation to complete the contrast normalization process, resulting in standardized grayscale image data.

[0022] Standardized grayscale image data is input into a pre-trained object detection network for cell nucleus center point localization, yielding cell nucleus center point coordinates and cell nucleus bounding box data. The pre-trained object detection network employs an encoder-decoder architecture. The encoder contains four downsampling stages, each consisting of two convolutional layers and a max-pooling layer connected sequentially. The convolutional kernels of the convolutional layers are all 3×3 in size, and the max-pooling layer has a pooling window size of 2×2 and a stride of 2. The decoder contains four upsampling stages, each consisting of an upsampling layer, a concatenation layer, and two convolutional layers connected sequentially. The upsampling factor of the upsampling layer is 2. The concatenation layer concatenates the feature map output from the upsampling layer with the corresponding feature map from the encoder along the channel dimension. The convolutional kernels of the convolutional layers are all 3×3 in size. The output of the object detection network is divided into a center point heatmap branch and a bounding box regression branch. The center point heatmap branch consists of a convolutional layer with a sigmoid activation function, and has 1 output channel, representing the probability that each pixel is the center point of the cell nucleus. The bounding box regression branch consists of a convolutional layer with 2 output channels, representing the width and height of the bounding box, respectively. Standardized grayscale image data is used as input to the object detection network. After forward propagation, a center point heatmap and a bounding box regression map are obtained. Peak detection is performed on the center point heatmap, and the coordinates of pixels with a probability value greater than a preset center point threshold (set to 0.5) are extracted as the cell nucleus center point coordinates. Simultaneously, the width and height of the bounding box are extracted from the corresponding positions in the bounding box regression map. These are combined with the cell nucleus center point coordinates to generate the cell nucleus bounding box data, which is defined by the center point coordinates, width, and height.

[0023] Based on the cell nucleus center point coordinates and cell nucleus bounding box data, a region growing algorithm is used to determine the boundaries of the cytoplasmic region corresponding to each cell nucleus, resulting in cytoplasmic region segmentation data. For each set of cell nucleus center point coordinates, the cell nucleus center point coordinates are used as the initial seed point for region growing. The growth criterion is that the absolute difference between the gray value of the pixel to be grown and the gray value of the seed point is less than a preset gray-level similarity threshold, which is set to 0.15. Region growing is iterated in the eight-neighbor direction, marking pixels that meet the growth criterion as part of the cytoplasmic region, until no new pixels meet the growth criterion, thus obtaining the cytoplasmic region segmentation data.

[0024] Based on nuclear bounding box data and cytoplasmic region segmentation data, image regions containing complete single cells are cropped from bone marrow cell microscopic image data to obtain single-cell image data. During cropping, the nuclear bounding box data is used as the center, and the cropping extends outward to completely encompass the cytoplasmic region marked by the cytoplasmic region segmentation data, forming a rectangular region of interest. Color images within the rectangular region of interest are then extracted from the bone marrow cell microscopic image data to obtain single-cell image data.

[0025] After obtaining the single-cell image data, color normalization processing is performed on the single-cell image data. The color normalization processing steps include: calculating the channel mean and channel standard deviation for the red, green, and blue channels of the single-cell image data, respectively, to obtain the red channel mean, green channel mean, blue channel mean, red channel standard deviation, green channel standard deviation, and blue channel standard deviation; subtracting the red channel mean from the red channel value of each pixel in the single-cell image data and dividing by the red channel standard deviation, subtracting the green channel mean from the green channel value and dividing by the green channel standard deviation, and subtracting the blue channel mean from the blue channel value and dividing by the blue channel standard deviation, to obtain channel-normalized pixel data; for the channel-normalized pixel data, multiplying the normalized red channel value by the target red channel mean and adding the target red channel standard deviation, multiplying the normalized green channel value by the target green channel mean and adding the target green channel standard deviation, and multiplying the normalized blue channel value by the target blue channel mean and adding the target blue channel standard deviation, to obtain the color-normalized single-cell image data. The target red channel mean, target green channel mean, target blue channel mean, target red channel standard deviation, target green channel standard deviation, and target blue channel standard deviation are statistical values ​​pre-calculated from the standard reference bone marrow cell dataset. The target red channel mean is set to 0.5, the target green channel mean is set to 0.4, the target blue channel mean is set to 0.5, the target red channel standard deviation is set to 0.2, the target green channel standard deviation is set to 0.18, and the target blue channel standard deviation is set to 0.2.

[0026] Multi-scale feature extraction was performed on the color-normalized single-cell image data to obtain nuclear chromatin texture features and cytoplasmic granule distribution features. When extracting nuclear chromatin texture features, local binary pattern encoding at different scales was performed on the nuclear region in the single-cell image data. The radius parameter of the local binary pattern encoding... The set of values ​​is {1, 2, 3}, which corresponds to the number of neighborhood points. The set of values ​​is {8, 16, 24}. For each set of radius parameters and the number of neighborhood points, a local binary pattern encoding value is calculated for each pixel within the cell nucleus region to obtain multi-scale texture encoding data. Histogram statistics are then performed on each set of multi-scale texture encoding data, and the number of intervals in the histogram is related to the number of neighborhood points. Related, when When the value is 8, the number of histogram intervals is 256. When the value is 16, the number of histogram intervals is 256. The number of histogram intervals is 256 when the histogram is 24. Histogram vectors at different scales are concatenated and spliced ​​to obtain the texture features of cell nuclear chromatin.

[0027] When extracting the distribution features of cytoplasmic particles, the directional gradient histogram is calculated for the cytoplasmic region in the single-cell image data. First, the horizontal gradient component and vertical gradient component of each pixel in the cytoplasmic region are calculated using the transpose of the horizontal gradient operator [-1,0,1] and the vertical gradient operator [-1,0,1], thus obtaining the gradient magnitude and gradient direction. The cytoplasmic region is divided into 8-pixel × 8-pixel cell units. Within each cell unit, the gradient direction is evenly divided into 9 directional intervals within the range of 0° to 180°. Weighted voting statistics are performed according to the gradient magnitude to obtain a 9-dimensional gradient histogram for each cell unit. Adjacent 2×2 cell units are combined into a block, and the blocks slide with the cell unit size as the step size. The gradient histograms of the 4 cell units in each block are concatenated and normalized using the L2 norm. A small constant ε=0.001 is added during normalization to avoid division by zero. The normalized histogram vectors of all blocks are concatenated to obtain the cytoplasmic gradient magnitude distribution data, i.e., the cytoplasmic particle distribution features.

[0028] Example 2

[0029] In specific implementation, please refer to Figure 3 Based on the coordinates of the nucleus center point of each cell in the single-cell image data, the Euclidean distance between any two nucleus center points is calculated, resulting in the inter-cell distance matrix data. The nucleus center point coordinates are derived from the localization results of the target detection network on the standardized grayscale image data. The nucleus center point coordinates of each cell are represented as a two-dimensional vector. It means that among them Number the cells. For the first The coordinates of the center point of the cell nucleus in each cell along the horizontal direction of the image. For the first The coordinates of the center point of the cell nucleus in each cell along the vertical direction of the image. Assume the single-cell image data contains... If there are multiple single cells, then the inter-cell distance matrix data is one. The matrix, the first in the matrix Line 1 Column elements Indicates the first The central point of the cell nucleus and the first The Euclidean distance between the centers of each cell nucleus is calculated using the following formula:

[0030] in, and The first The horizontal and vertical coordinates of the center point of each cell nucleus and The first The horizontal and vertical coordinates of the center point of each cell nucleus and The value range is 1 to integers, when hour .

[0031] Cell pairs in the intercellular distance matrix data that are less than a preset distance threshold are identified as adjacent cell pairs, and a cell adjacency graph is constructed based on these adjacent cell pairs. The preset distance threshold is set to 250 pixels, based on the following: the spatial resolution of bone marrow cell microscopic images after standardization is approximately 0.1 micrometers per pixel; the average diameter of normal bone marrow cells is approximately 12 to 20 micrometers; the intercellular matrix region between adjacent cells is approximately 10 to 30 micrometers; and the physical distance corresponding to 250 pixels is approximately 25 micrometers, which effectively connects adjacent cells and excludes intercellular associations that are too far apart. For each element in the intercellular distance matrix data... ,like Then the first The cell and the first Each cell is identified as a neighboring cell pair. Cell adjacency graph data is represented in a graph structure, denoted as . ,in A set of nodes, containing Each node corresponds to a complete single cell in the single-cell image data; Let be the set of edges, if the first edge is... The cell and the first If each cell is a neighboring cell pair, then at node and nodes Establish an undirected edge between them Adjacency matrix of cell adjacency graph data For one A symmetric matrix, when hour ,otherwise diagonal elements .

[0032] Cell nuclear chromatin texture features and cytoplasmic granule distribution features are used as the initial node features in the cell adjacency graph data. Multi-level neighborhood feature aggregation is then performed using a pre-constructed graph convolutional network to obtain updated node feature data. For each node... , to transform the nuclear chromatin texture feature vector cytoplasmic particle distribution feature vector By concatenating the nodes, we obtain the initial feature vectors of the nodes. ,symbol This represents a vector concatenation operation, where the initial feature vector of the node has a dimension of . ,in The dimension representing the texture features of nuclear chromatin. This represents the dimension of cytoplasmic particle distribution characteristics.

[0033] The pre-built graph convolutional network consists of two stacked graph convolutional layers. The input feature dimension of the first graph convolutional layer is... The first graph has an input feature dimension of 128 and an output feature dimension of 64, with the activation function being ReLU. The second graph's convolutional layer has an input feature dimension of 128 and an output feature dimension of 64, with the activation function being ReLU. The convolutional layer in the first graph uses a message-passing mechanism based on neighborhood aggregation. Layer graph convolutional layer node pair The feature update method is as follows: First, the neighborhood feature aggregation vector is calculated. Then, the aggregation vector is added to the transformed features of the node itself. Finally, feature transformation is performed using a learnable weight matrix and an activation function. Two graph convolutional layers implement first-order and second-order neighborhood aggregation, respectively. The first graph convolutional layer aggregates the feature information of direct neighbors, and the second graph convolutional layer further aggregates the feature information of indirect neighbors based on the output of the first graph convolutional layer, realizing multi-order neighborhood feature aggregation. After processing by two graph convolutional layers, the updated features of each node are obtained. The update features of all nodes constitute the node update feature data.

[0034] Global pooling is performed on the node update feature data to obtain the cell spatial topological relationship features. The global pooling operation uses average pooling, and the calculation formula is as follows:

[0035] in, A feature vector representing the topological relationships in cell space. This represents the total number of nodes in the cell adjacency graph data. Indicates the first The updated feature vectors of each node are processed through a two-layer graph convolutional network. The cellular spatial topological relationship feature vectors have a dimension of 64, serving as an encoding representation of the spatial organization of cell populations in the entire bone marrow cell microscopic image.

[0036] Example 3

[0037] In specific implementation, please refer to Figure 4 The first linear projection layer maps the nuclear chromatin texture features to a common feature space, resulting in the first projected feature data. The nuclear chromatin texture feature vector is denoted as... ,in The dimension representing the texture features of nuclear chromatin is determined by the sum of the number of histogram intervals at different scales in multi-scale local binary pattern encoding. The first linear projection layer consists of a fully connected layer without an activation function; its learnable parameters include the weight matrix. and bias vector ,in The feature dimension represents the common feature space. The dimension was set to 256, based on the following criteria: balancing feature representation capability and computational efficiency, 256 dimensions can fully accommodate alignment information for texture features, particle distribution features, and spatial topological relationships, while maintaining compatibility with the input dimensions of subsequent multi-level class discriminant networks. The calculation method for the first projection feature data is as follows:

[0038] in, This represents the first projection feature data. This is the weight matrix of the first linear projection layer. This is the bias vector of the first linear projection layer. The feature vector of chromatin texture in the cell nucleus, symbolized as " " indicates matrix multiplication.

[0039] The cytoplasmic particle distribution features are mapped to a common feature space through a second linear projection layer, resulting in second projected feature data. The cytoplasmic particle distribution feature vector is denoted as... ,in The dimension representing the distribution characteristics of cytoplasmic granules is determined by the total length of the normalized histogram vectors of all blocks in the oriented gradient histogram calculation. The second linear projection layer consists of a fully connected layer without an activation function; its learnable parameters include the weight matrix. and bias vector The calculation method for the second projection feature data is as follows:

[0040] in, This represents the second projection feature data. This is the weight matrix of the second linear projection layer. This is the bias vector for the second linear projection layer. This represents the characteristic vector of cytoplasmic particle distribution.

[0041] The cell space topological relationship features are mapped to the common feature space through a third linear projection layer, resulting in third-projection feature data. The cell space topological relationship feature vector is denoted as... ,in Dimensions representing the topological characteristics of cell space. The value is 64, consistent with the output feature dimension of the second graph convolutional layer in the graph convolutional network. The third linear projection layer consists of a fully connected layer without an activation function; its learnable parameters include the weight matrix. and bias vector The calculation method for the third projection feature data is as follows:

[0042] in, This represents the third projection feature data. This is the weight matrix of the third linear projection layer. This is the bias vector of the third linear projection layer. This represents the feature vector of the cell space topological relationship.

[0043] The first, second, and third projected feature data are subjected to element-wise summation and layer normalization to obtain fused cell characterization data. The element-wise summation operation sums the three projected feature vectors along corresponding dimensions in the common feature space. The calculation method is as follows:

[0044] in, This represents the eigenvector after element-wise addition.

[0045] The element-wise addition result is subjected to layer normalization, calculated as follows:

[0046] in, This indicates the fusion of cell characterization data. The scaling parameter vector representing the layer normalization. The translation parameter vector represents the normalized layer. This represents element-wise multiplication. express The mean of all dimension values, express The variance of all dimension values Represents the numerical stability constant. The value is set to 0.00001, based on the following principles: effectively preventing the denominator from being zero under floating-point precision, while having a negligible impact on the normalization result. (Scaling parameter vector) Translation parameter vector These are learnable parameters, updated via backpropagation during model training, with initial values ​​set to vectors of all 1s and all 0s, respectively. The fused cell representation data has a dimension of 256 and serves as input to the subsequent multi-level classification network.

[0047] Example 4

[0048] In practice, the fused cell characterization data is input into a multi-level classification network. The fused cell characterization data is denoted as... ,in For the dimensions of the public feature space, The value is 256. The multi-level class discriminant network consists of two branches: a coarse-grained discriminant subnetwork and a fine-grained discriminant subnetwork. The coarse-grained discriminant subnetwork is responsible for predicting the major category, while the fine-grained discriminant subnetwork is responsible for predicting the subtype.

[0049] The coarse-grained discriminant subnetwork consists of three fully connected layers connected sequentially: a first fully connected layer, a second fully connected layer, and a third fully connected layer. The first fully connected layer has an input feature dimension of 256 and an output feature dimension of 128, with ReLU activation. The second fully connected layer has an input feature dimension of 128 and an output feature dimension of 64, with ReLU activation. The third fully connected layer has an input feature dimension of 64 and an output feature dimension of 3, corresponding to the three categories: normal cells, reactive cells, and tumor cells. No activation function is used in the third fully connected layer. The fused cell characterization data is sequentially passed through the first and second fully connected layers for dimensionality reduction, resulting in 64-dimensional dimensionality-reduced feature data. This dimensionality-reduced feature data is then input into the third fully connected layer for class score calculation, yielding a class score vector. , The three components represent the raw scores of the normal cell category, the reactive cell category, and the tumor cell category, respectively. The category score vector is processed using a normalized exponential function mapping, calculated as follows:

[0050] in, Indicates the first The probability values ​​of each major category The value range is an integer from 1 to 3, where 1 corresponds to the normal cell category, 2 corresponds to the reactive cell category, and 3 corresponds to the tumor cell category. Indicates the first The raw scores for each major category; Indicates the first The raw scores for each major category The summation index is an integer ranging from 1 to 3; Represented by natural constant An exponential function with base 1. Through the above calculations, probability data for normal cell categories, reactive cell categories, and tumor cell categories are obtained.

[0051] When the probability data of tumor cell major categories exceeds a preset major category threshold, the fused cell characterization data is input into the fine-grained discriminant subnetwork for subtype classification prediction. The preset major category threshold is set to 0.6, based on the principle that 0.6 as a major category determination threshold strikes a balance between clinical sensitivity and specificity. When the tumor major category probability exceeds 0.6, it indicates the presence of strong tumor cell characteristic signals, providing sufficient confidence for fine-grained subtype discrimination. The fine-grained discriminant subnetwork consists of a fourth fully connected layer, a fifth fully connected layer, and a sixth fully connected layer connected sequentially. The fourth fully connected layer has an input feature dimension of 256, an output feature dimension of 128, and uses the ReLU activation function. The fifth fully connected layer has an input feature dimension of 128, an output feature dimension of 64, and uses the ReLU activation function. The sixth fully connected layer has an input feature dimension of 64, and an output feature dimension of... , This represents the total number of leukemia subtypes. According to the World Health Organization's classification of leukemia, it includes eight subtypes: acute lymphoblastic leukemia (ALL) L1, acute lymphoblastic leukemia (ALL) L2, acute myeloid leukemia (AML) undifferentiated, acute myeloid leukemia (AML) partially differentiated, acute promyelocytic leukemia (APML), acute myelomonocytic leukemia (AML), acute monocytic leukemia (Monocytic leukemia), and acute erythroleukemia (ALL). The value is set to 8. The sixth fully connected layer does not use an activation function. The fused cell characterization data is sequentially passed through the fourth and fifth fully connected layers for feature dimensionality reduction, then through the sixth fully connected layer for class score calculation. The output of the sixth fully connected layer is processed using a normalized exponential function mapping. The formula structure for this mapping is consistent with that in the coarse-grained discriminant subnetwork, with a summation range of 1 to... The probability data for each leukemia subtype were obtained, and the probability data for each leukemia subtype contained a total of 8 probability values.

[0052] When the probability data for the normal cell category exceeds a preset category threshold, the fused cell characterization data is input into the fine-grained discriminant subnetwork for normal subtype classification prediction. At this point, the fine-grained discriminant subnetwork uses the fourth and fifth fully connected layers shared with the leukemia subtype branch, and the output feature dimension of the sixth fully connected layer is set to... , This represents the total number of categories with normal subtypes. The value is set to 5, corresponding to primitive granulocytes, promyelocytes, metamyelocytes, metamyelocytes, and band neutrophils, respectively. No activation function is used in the sixth fully connected layer. A normalized exponential function is applied to the output of the sixth fully connected layer, with a summation range of 1 to... The probability data of each normal subtype were obtained.

[0053] When the probability data for reactive cell categories exceeds a preset category threshold, the fused cell characterization data is input into the fine-grained discriminant network for reactive subtype classification prediction. In this case, the fine-grained discriminant network uses the fourth and fifth fully connected layers shared with the leukemia subtype branch, and the output feature dimension of the sixth fully connected layer is set to... , This represents the total number of categories of reactive subtypes. The value is set to 4, corresponding to reactive lymphocytes, reactive monocytes, reactive neutrophils, and reactive eosinophils, respectively. No activation function is used in the sixth fully connected layer. The output of the sixth fully connected layer is processed using a normalized exponential function mapping, with a summation range of 1 to... The probability data of each reactive subtype were obtained.

[0054] Based on the probability data of each leukemia subtype, normal subtype, or reactive subtype, probability distribution data for each category is determined. Each category's probability distribution data is a probability vector composed of all subtype probability values ​​output by the fine-grained discriminant subnetwork; its dimension depends on the activated discriminant branch, and can be 8-dimensional, 5-dimensional, or 4-dimensional. The category with the highest probability value is selected from the probability distribution data as the classification result for the bone marrow cell microscopic images.

[0055] Example 5

[0056] In specific implementation, please refer to Figure 5The study acquired training sample data of bone marrow cell microscopic images and corresponding cell type annotation data. The nucleus region was located and the cytoplasm region was divided into single-cell training image data. The bone marrow cell microscopic image training sample data originated from desensitized digital images of clinical bone marrow smears. Each image was annotated by hematology experts according to the World Health Organization's classification of leukemia, with each identifiable cell type labeled. The cell type annotation data included 17 categories: normal blast cells, normal promyelocytes, normal metamyelocytes, normal metamyelocytes, normal band neutrophils, reactive lymphocytes, reactive monocytes, reactive neutrophils, reactive eosinophils, acute lymphoblastic leukemia L1 subtype, acute lymphoblastic leukemia L2 subtype, acute myeloid leukemia undifferentiated, acute myeloid leukemia partially differentiated, acute promyelocytic leukemia, acute myelomonocytic leukemia, acute monocytic leukemia, and acute erythroleukemia. The bone marrow cell microscopic image training sample data were subjected to grayscale and contrast normalization to obtain standardized grayscale image data. The standardized grayscale image data were input into a pre-trained target detection network to obtain the cell nucleus center point coordinate data and cell nucleus bounding box data. The boundary of the cytoplasm region corresponding to each cell nucleus was determined by the region growing algorithm to obtain cytoplasm region segmentation data. Based on the cell nucleus bounding box data and cytoplasm region segmentation data, the image region containing a complete single cell was cropped to obtain single-cell training image data.

[0057] Color normalization is performed on the single-cell training image data: The mean, mean, and mean values ​​of the red, green, and blue channels, as well as the standard deviations of the red, green, and blue channels, are calculated for each pixel. Then, the red channel value of each pixel is subtracted from the mean and divided by the standard deviation; the green channel value is subtracted from the mean and divided by the standard deviation; and the blue channel value is subtracted from the mean and divided by the standard deviation. Finally, the normalized red channel value is multiplied by the target red channel mean and added to the target red channel standard deviation; the normalized green channel value is multiplied by the target green channel mean and added to the target green channel standard deviation; and the normalized blue channel value is multiplied by the target blue channel mean and added to the target blue channel standard deviation. This yields the color-normalized single-cell training image data. The target red channel mean is set to 0.5, the target green channel mean to 0.4, the target blue channel mean to 0.5, the target red channel standard deviation to 0.2, the target green channel standard deviation to 0.18, and the target blue channel standard deviation to 0.2.

[0058] Multi-scale feature extraction was performed on the color-normalized single-cell training image data to obtain the nuclear chromatin texture features and cytoplasmic granule distribution features of the training cells. When extracting the nuclear chromatin texture features, multi-scale local binary pattern encoding was performed on the nuclear region in the single-cell training image data, with a radius parameter... The set of values ​​is {1, 2, 3}, which corresponds to the number of neighborhood points. The value set is {8, 16, 24}. Histogram statistics are performed on the local binary pattern encoded values ​​at each scale. The number of histogram intervals is within... The histogram vectors at different scales (8, 16, and 24) were all set to 256. The histogram vectors at different scales were concatenated to obtain the training cell nuclear chromatin texture features. When extracting the training cytoplasmic particle distribution features, the oriented gradient histogram was calculated for the cytoplasmic region in the single-cell training image data. The cytoplasmic region was divided into 8-pixel × 8-pixel cell units, and the gradient direction was divided into 9 directional intervals within the range of 0° to 180°. Two × 2 cell units were combined into a block with a block sliding step of 8 pixels. The gradient histogram within each block was normalized using the L2 norm, with a small constant of 0.001 added during normalization. The normalized histogram vectors of all blocks were concatenated to obtain the training cytoplasmic particle distribution features.

[0059] Based on the chromatin texture features of the training cell nuclei and the distribution features of the training cytoplasm granules, and combined with the coordinate data of the cell nucleus center points in the single-cell training image data, a training cell adjacency graph is constructed. Assume that the single-cell training image data contains... The first single cell, the... The coordinates of the center point of the cell nucleus of each cell are: , No. The coordinates of the center point of the cell nucleus of each cell are: Calculate the Euclidean distance between any two cell nucleus centers. ,in and The value range is 1 to The integer. Cell pairs with an Euclidean distance less than a preset distance threshold are identified as adjacent cell pairs. The preset distance threshold is set to 250 pixels. This is based on the fact that the spatial resolution of bone marrow cell microscopic images after standardization is approximately 0.1 micrometers per pixel, and the physical distance corresponding to 250 pixels is approximately 25 micrometers, which can effectively connect adjacent cells and exclude cell associations that are too far apart. The training cell adjacency graph data is represented in a graph structure, and the node set contains There are 12 nodes, each corresponding to a single cell. The edge set consists of undirected edges defined by all neighboring cell pairs. The initial features of each node are formed by concatenating the training cell nuclear chromatin texture feature vector and the training cytoplasmic particle distribution feature vector.

[0060] Using training cell adjacency graph data as input and cell type labeling data as supervision signal, the cross-entropy loss function is used to iteratively update the network parameters of the graph convolutional network (GCNN) to obtain the trained GCNN. The GCNN consists of two stacked graph convolutional layers. The input feature dimension of the first graph convolutional layer is the sum of the dimensions of the training cell nuclear chromatin texture features and the training cytoplasmic particle distribution features, with an output feature dimension of 128 and a ReLU activation function. The input feature dimension of the second graph convolutional layer is 128, the output feature dimension is 64, and the activation function is ReLU. The two graph convolutional layers implement first-order and second-order neighborhood aggregation, respectively. The Adam optimizer is used during training, with an initial learning rate of 0.001, a weight decay coefficient of 0.0001, a batch size of 32, and 100 training generations. During each generation of training, the training cell adjacency graph data is input into the graph convolutional network for forward propagation to obtain node update features. After global average pooling, the node update features are input into a classification fully connected layer. The input dimension of the classification fully connected layer is 64, and the output dimension is 17, corresponding to 17 cell type categories. The output of the classification fully connected layer is mapped by a normalized exponential function to obtain the predicted probability of each category.

[0061] The cross-entropy loss function is calculated as follows:

[0062] in, This represents the cross-entropy loss value. This is the cell type category index, with values ​​ranging from 1 to 17. Indicates the first The weight coefficients of the category, and the weight coefficients are related to the category. The frequency of each category in the bone marrow cell microscopic image training sample data is inversely proportional to its occurrence. The calculation method is as follows: Statistically analyze the training sample data of bone marrow cell microscopic images... Number of cells in each category Calculate the first Frequency of each category ,in Let be the total number of cells of all categories in the training sample data of bone marrow cell micrographs. ; Indicates the first The true labels of each category are encoded using one-hot encoding. If a training sample belongs to the first category... Each category The value is 1 if it is set to 1, otherwise the value is 0. This indicates that the graph convolutional network is applied to the first... The predicted probability values ​​for each category are obtained by mapping the scores output by the fully connected classification layer through a normalized exponential function; Represented by natural constant The algorithm uses a logarithmic function with base 0.05. Gradients are calculated by backpropagation of the network parameters of the graph convolutional network and the classification fully connected layer based on the cross-entropy loss value. The Adam optimizer is then used to update all learnable parameters in both the graph convolutional network and the classification fully connected layer, completing one generation of training. This process is repeated until the preset number of training generations is reached, resulting in a trained graph convolutional network. After training, the classification fully connected layer is removed, and the network parameters of the two graph convolutional layers are retained for encoding the cell spatial topological features of actual bone marrow cell microscopic image data.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application 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 this application should be included within the scope of protection of this application.

Claims

1. A deep learning-based method for classifying microscopic images of leukemia bone marrow cells, characterized in that, The method includes the following steps: Obtain bone marrow cell microscopic image data, and perform nucleus region localization and cytoplasm region division on the bone marrow cell microscopic image data to obtain single cell image data; Multi-scale feature extraction was performed on the single-cell image data to obtain the nuclear chromatin texture features and cytoplasmic particle distribution features; The cell nuclear chromatin texture features and the cytoplasmic particle distribution features are input into a pre-constructed graph convolutional network to encode cell adjacency relationships, thereby obtaining cell spatial topological relationship features; The cell spatial topological relationship features, the cell nuclear chromatin texture features, and the cytoplasmic particle distribution features are aligned and stitched together across modalities to obtain fused cell characterization data; Based on the fused cell characterization data, a multi-level classification network is used to predict the probability of leukemia cell types, obtain probability distribution data for each category, and determine the classification result of bone marrow cell microscopic images based on the probability distribution data for each category.

2. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 1, characterized in that, The process of acquiring bone marrow cell microscopic image data and performing nuclear region localization and cytoplasmic region segmentation on the bone marrow cell microscopic image data to obtain single-cell image data includes the following steps: The bone marrow cell microscopic image data were subjected to grayscale conversion and contrast normalization to obtain standardized grayscale image data. The standardized grayscale image data is input into a pre-trained target detection network to locate the cell nucleus center point, thereby obtaining cell nucleus center point coordinate data and cell nucleus bounding box data. Based on the cell nucleus center point coordinate data and the cell nucleus bounding box data, a region growing algorithm is used to determine the boundary of the cytoplasmic region corresponding to each cell nucleus, thereby obtaining cytoplasmic region segmentation data. Based on the cell nuclear bounding box data and the cytoplasmic region segmentation data, an image region containing a complete single cell is cropped from the bone marrow cell microscopic image data to obtain the single-cell image data.

3. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 2, characterized in that, The process of performing multi-scale feature extraction on the single-cell image data to obtain nuclear chromatin texture features and cytoplasmic granule distribution features includes the following steps: Local binary pattern encoding at different scales is performed on the cell nucleus region in the single-cell image data to obtain multi-scale texture encoded data; Histogram statistics and concatenation are performed on the multi-scale texture coding data to obtain the cell nuclear chromatin texture features; The orientation gradient histogram of the cytoplasm region in the single-cell image data is calculated to obtain the cytoplasmic gradient amplitude distribution data. The cytoplasmic gradient amplitude distribution data is subjected to block statistical and normalization processing to obtain the cytoplasmic particle distribution characteristics.

4. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 3, characterized in that, The step of inputting the nuclear chromatin texture features and the cytoplasmic granule distribution features into a pre-constructed graph convolutional network to encode cell adjacency relationships and obtain cell spatial topological relationship features includes the following steps: Based on the coordinates of the cell nucleus center point of each cell in the single-cell image data, the Euclidean distance between any two cell nucleus center points is calculated to obtain the inter-cell distance matrix data. Cell pairs in the inter-cell distance matrix data that are less than a preset distance threshold are identified as adjacent cell pairs, and cell adjacency graph data is constructed based on the adjacent cell pairs. The nodes of the cell adjacency graph data are single cells, and the edges of the cell adjacency graph data are adjacency relationships. The cell nuclear chromatin texture features and the cytoplasmic particle distribution features are used as the initial node features of the cell adjacency graph data. Multi-order neighborhood feature aggregation is performed through the graph convolutional network to obtain node update feature data. The node update feature data is subjected to global pooling to obtain the cell spatial topological relationship features.

5. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 4, characterized in that, The process of aligning and stitching the cell spatial topological features with the cell nuclear chromatin texture features and the cytoplasmic particle distribution features across modalities to obtain fused cell characterization data includes the following steps: The cell nuclear chromatin texture features are mapped to a common feature space through a first linear projection layer to obtain first projection feature data; The cytoplasmic particle distribution features are mapped to the common feature space through a second linear projection layer to obtain second projection feature data. The cell space topological relationship features are mapped to the common feature space through the third linear projection layer to obtain the third projection feature data; The first projection feature data, the second projection feature data, and the third projection feature data are subjected to element-wise addition and layer normalization to obtain the fused cell characterization data.

6. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 5, characterized in that, The process of predicting the probability of leukemia cell types using a multi-level classification network based on the fused cell characterization data, obtaining probability distribution data for each category, and determining the classification result of the bone marrow cell microscopic image based on the probability distribution data for each category includes the following steps: The fused cell characterization data is sequentially input into the first fully connected layer and the second fully connected layer of the multi-level category discrimination network for feature dimensionality reduction to obtain dimensionality-reduced feature data. The dimensionality-reduced feature data is input into the third fully connected layer of the multi-level class discrimination network to calculate the class scores and obtain the initial score data for each class. The initial score data for each category is processed by a normalized exponential function mapping to obtain the probability distribution data for each category; The category with the highest probability value is selected from the probability distribution data of each category as the classification result of the bone marrow cell micrograph.

7. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 1, characterized in that, The pre-built graph convolutional network is trained in the following way: Obtain bone marrow cell microscopic image training sample data and corresponding cell type labeling data, and perform cell nucleus region localization and cytoplasm region division on the bone marrow cell microscopic image training sample data to obtain single cell training image data; Multi-scale feature extraction is performed on the single-cell training image data to obtain the nuclear chromatin texture features and cytoplasmic particle distribution features of the training cells; Based on the chromatin texture features of the training cell nuclei and the distribution features of the training cell granules, and combined with the coordinate data of the cell nucleus center point in the single-cell training image data, a training cell adjacency graph is constructed. Using the training cell adjacency graph data as input and the cell type labeling data as supervision signal, the cross-entropy loss function is used to iteratively update the network parameters of the graph convolutional network to obtain the trained graph convolutional network.

8. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 7, characterized in that, The cross-entropy loss function sets different weight coefficients for different cell type categories, and the weight coefficients are inversely proportional to the frequency of each category in the training samples.

9. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 1, characterized in that, The method of using a multi-level classification discriminant network to predict the probability of leukemia cell types includes the following steps: The multi-level classification network includes a coarse-grained classification subnetwork and a fine-grained classification subnetwork; The fused cell characterization data is input into the coarse-grained discriminant network for major category prediction to obtain the probability data of normal cell major category, the probability data of reactive cell major category, and the probability data of tumor cell major category. When the probability data of the tumor cell category is greater than the preset category threshold, the fusion cell characterization data is input into the fine-grained discriminant network for subtype classification prediction to obtain the probability data of each leukemia subtype. When the probability data of the normal cell category or the probability data of the reactive cell category is greater than the preset category threshold, the fused cell characterization data is input into the fine-grained discriminant subnetwork for normal or reactive subtype classification prediction, and the probability data of each normal subtype or each reactive subtype is obtained. Based on the probability data of each leukemia subtype, the probability data of each normal subtype, or the probability data of each reactive subtype, the probability distribution data of each category are determined.

10. The deep learning-based microscopic image classification method for leukemia bone marrow cells according to claim 1, characterized in that, Before performing multi-scale feature extraction on the single-cell image data, a step of color normalization of the single-cell image data is also included, which specifically includes: The channel mean and channel standard deviation of the red, green and blue channels in the single-cell image data are calculated respectively to obtain the statistical parameter data of each channel; Based on the statistical parameter data of each channel, the channel mean of the corresponding channel is subtracted from the channel value of each pixel in the single-cell image data, and then divided by the channel standard deviation of the corresponding channel to obtain the channel-standardized pixel data. The channel-normalized pixel data is multiplied by the mean of the target red channel, the mean of the target green channel, and the mean of the target blue channel, respectively, and then the standard deviations of the target red channel, the target green channel, and the target blue channel are added to obtain the color-normalized single-cell image data.

Images