Early screening method and system for blood tumors based on multi-channel blood scattergram

Abstract

The present invention provides a method and system for early screening of hematologic malignancies based on multi-channel blood scattergrams. The method comprises: obtaining a blood scattergram of a target subject; inputting each of the blood scattergrams into a pre-constructed feature extraction network to obtain deep feature maps output by the feature extraction network corresponding to each of the blood scattergrams; inputting each of the obtained deep feature maps into a pre-trained balanced optimization model to obtain an enhanced feature map output by the balanced optimization model; inputting each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier; fusing each of the enhanced feature maps to obtain a fused feature map, and obtaining a fused feature category prediction result based on the fused feature map, and obtaining a final category prediction result through weighted calculation. This method implements computer-assisted screening, reduces the empirical dependence of early screening for hematologic malignancies, and improves the accuracy and efficiency of screening results.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a method and system for early screening of blood tumors based on a multi-channel blood scattergram. Background Art

[0002] Lymphomas are a group of clonal malignancies that originate from B cells, T cells, or NK cells at various stages of differentiation. Currently, accurate diagnosis of lymphomas is based on morphology, histopathology, immunohistochemistry, flow cytometry immunophenotyping, in situ hybridization, cytogenetics, and molecular biology. However, in clinical practice, these tests are only performed in cases with a high suspicion of lymphoma, making it crucial to provide clinically relevant clues and guide timely testing. Lymphomas are more common in tissues outside the bone marrow and peripheral blood. However, when lymphomas develop a leukemic phase, morphologically abnormal cells may be seen in the peripheral blood, as seen in acute lymphoblastic leukemia, chronic lymphocytic leukemia, follicular lymphoma, mantle cell lymphoma, and plasma cell neoplasms. A complete blood count (CBC) is one of the most commonly performed routine tests in clinical practice, especially during the initial evaluation phase, with blood cell morphology analysis being a core method for diagnosing hematologic disorders. The term "abnormal lymphocytes" is used to describe lymphocytes suspected of malignant or clonal origin. The discovery of such abnormal cells in peripheral blood smears often provides key diagnostic evidence for lymphoproliferative disorders and plays an important role in the diagnosis and treatment of various lymphomas.

[0003] However, the traditional method of manual microscopic examination of peripheral blood smears has the limitations of being highly dependent on experience, inefficient, and susceptible to subjective bias.

[0004] To address the challenges of traditional methods, blood cell scatter plots are currently being used for supplementary screening. Fully automated hematology analyzers provide detailed information on multi-channel scatter plots, and abnormalities in cell distribution patterns in these scatter plots are valuable indicators for diagnosing various blood diseases, including leukemia. However, identifying abnormalities in scatter plot cell distribution patterns still has the limitation of being highly empirically dependent.

[0005] Therefore, how to realize computer-assisted screening, reduce the empirical dependence of early screening of blood tumors, and improve the accuracy and efficiency of screening results has become an urgent problem to be solved by technical personnel in this field. Summary of the Invention

[0006] The present invention aims to provide a method and system for early screening of blood tumors based on a multi-channel blood scattergram, so as to at least partially solve the above technical problems.

[0007] The present invention provides a method for early screening of blood tumors based on a multi-channel blood scattergram, the method comprising:

[0008] Obtaining a blood scatter plot of at least two blood cell detection dimensions of the target subject;

[0009] Inputting each of the blood scattergrams into a pre-built feature extraction network to obtain deep feature maps output by the feature extraction network that correspond to each of the blood scattergrams;

[0010] Inputting each of the obtained deep feature maps into a pre-trained balanced optimization model to obtain an enhanced feature map output by the balanced optimization model;

[0011] Inputting each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier;

[0012] Fusing the enhanced feature maps to obtain a fused feature map, and obtaining a fused feature category prediction result based on the fused feature map;

[0013] The category prediction results corresponding to each dimension and the fusion feature category prediction results are weighted to obtain the final category prediction result.

[0014] In some embodiments, obtaining a blood scattergram of at least two blood cell detection dimensions of a target subject specifically includes:

[0015] obtaining a blood sample from the target subject;

[0016] Inputting the blood sample into a blood cell analyzer to obtain a first channel scatter plot, a second channel scatter plot, and a third channel scatter plot generated by the blood cell analyzer;

[0017] The first channel scatter plot is a blood scatter plot of a white blood cell classification channel, the second channel scatter plot is a blood scatter plot of a basophil / nucleated red blood cell channel, and the third channel scatter plot is a blood scatter plot of a primitive cell / abnormal lymphocyte channel.

[0018] In some embodiments, the training process of the balance optimization model specifically includes:

[0019] Preprocessing the blood scatter plots of each channel of multiple groups of blood samples used for training;

[0020] Each preprocessed blood scatter plot and the corresponding category label are used as a pair of training data, and a training set is constructed;

[0021] Each blood scatter plot in the training set is input into the pre-built feature extraction network for training until the feature extraction network loss of each channel converges and the training is stopped to obtain the trained feature extraction network of the three channels;

[0022] Based on the training feature extraction network, the balanced optimization method is applied to train the classifier.

[0023] In some embodiments, during training:

[0024] Random sampling is used for the preprocessed blood scatter plot data of each channel in the training set, and the randomly sampled scatter plots are recorded as ;

[0025] Weighted sampling is performed based on the opposite number of the frequency of occurrence of the category to which the sample belongs in the training set. If the number of times the category to which the current sample belongs appears in the training set is , then the probability of the current sample being sampled is , where j is the total number of categories, weighted random sampling will be performed according to the inverse of the frequency of occurrence, and the scatter plot of weighted random sampling is recorded as ;

[0026] Sampling to and After that, load the trained feature extraction network and output the deep feature maps corresponding to the two scatter plots and , using a random mask with a value of 0 or 1 in the feature channel dimension ,Will Replace the feature map channels of a specific ratio in In the enhanced feature map ,in for The complement of The matrix is ​​multiplied element by element, and the replacement ratio is controlled by the parameter p;

[0027] The enhanced feature map is average pooled and then the prediction is optimized after the classifier output balance.

[0028] In some embodiments, the classifier training process specifically includes:

[0029] After obtaining the balanced optimization prediction, the image Tags and images Tags Perform linear fusion to obtain the fused label ;

[0030] Combine the balanced optimized category predictions with the fused labels Calculate the cross entropy loss and update the parameters of the classifier to obtain a trained three-channel classifier.

[0031] In some embodiments, during training:

[0032] In the first stage, the feature extraction network is trained using cross entropy loss. In the second stage, the parameters of the trained feature extraction network are frozen and the classifier is fine-tuned by fusion loss. The model predicts The loss is calculated separately for the labels of blood tumor class t and healthy class h, and the loss ratio is controlled by the fusion ratio p.

[0033] In some embodiments, the category prediction results corresponding to each dimension and the fusion feature category prediction results are weighted to obtain a final category prediction result, specifically including:

[0034] Set the feature output corresponding to the first channel scatter plot to The obtained category prediction result is , the feature output corresponding to the second channel scatter plot is The obtained category prediction result is , the feature output corresponding to the third channel scatter plot is The prediction results are ;

[0035] The three channels are spliced ​​and input into the fully connected layer and then activated by ReLU to map the fusion features. The fusion features are input into the fusion feature classifier to predict the category. ,right Calculate cross entropy loss with labels and update the fully connected layer and fusion classifier;

[0036] The category prediction results of each channel and the category prediction results of the fusion feature are weighted and calculated to obtain the final category prediction result.

[0037] The present invention also provides a blood tumor early screening system based on a multi-channel blood scattergram, the system comprising:

[0038] a scattergram acquisition unit, configured to acquire a blood scattergram of at least two blood cell detection dimensions of a target subject;

[0039] a feature extraction unit, configured to input each of the blood scattergrams into a pre-built feature extraction network to obtain deep feature maps output by the feature extraction network and corresponding to each of the blood scattergrams;

[0040] A feature enhancement unit, configured to input each of the obtained deep feature maps into a pre-trained balance optimization model to obtain an enhanced feature map output by the balance optimization model;

[0041] A category prediction unit, configured to input each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier;

[0042] A prediction result fusion unit is used to fuse the enhanced feature maps to obtain a fused feature map, and obtain a fused feature category prediction result based on the fused feature map;

[0043] The screening result generating unit is used to perform weighted calculation on the category prediction results corresponding to each dimension and the fusion feature category prediction results to obtain the final category prediction result.

[0044] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above-described methods when executing the program.

[0045] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which implements any of the above-described methods when executed by a processor.

[0046] The present invention provides a method and system for early screening of blood tumors based on multi-channel blood scattergrams. The method obtains blood scattergrams of at least two blood cell detection dimensions of a target subject, inputs each of the blood scattergrams into a pre-built feature extraction network to obtain deep feature maps corresponding to each of the blood scattergrams output by the feature extraction network; inputs each of the obtained deep feature maps into a pre-trained balanced optimization model to obtain an enhanced feature map output by the balanced optimization model; inputs each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier; fuses each of the enhanced feature maps to obtain a fused feature map, and obtains a fused feature category prediction result based on the fused feature map; and performs a weighted calculation on the category prediction results corresponding to each dimension and the fused feature category prediction result to obtain a final category prediction result. The category prediction result is in the form of a predicted probability for each category, and the maximum value of the category prediction result is taken to output the prediction result for auxiliary screening. This achieves computer-assisted screening, reduces the empirical dependence of early screening of blood tumors, and improves the accuracy and efficiency of screening results.

[0047] In the screening of hematologic malignancies, the number of samples from healthy individuals is far greater than that from patients with the disease. This is especially true for rare tumor types, for which available samples are even scarcer. This sample imbalance (i.e., long-tail distribution) can cause machine learning models to favor the majority class (healthy samples) during training, resulting in weaker recognition of the minority class (tumor samples), leading to high missed diagnosis rates or insufficient specificity. In some embodiments, the methods and systems provided by the present invention incorporate H2T (Head-to-Tail Fusion, a feature fusion strategy in long-tail visual recognition) after the feature extraction network or during classifier training. This approach transfers the rich feature space information from the head class samples to the tail class samples, enhancing their diversity and semantic expression, thereby addressing the data bias problem caused by sample imbalance.

[0048] In practical applications, because different channels in a blood scatter plot (such as WDF, WNR, and WPC) reflect distinct physical and chemical properties of cells, existing methods fail to fully integrate and utilize this multi-channel information, reducing their classification capabilities. In some embodiments, the methods and systems provided by the present invention fuse multi-channel information through a cross-attention mechanism to achieve accurate classification of blood cancer diseases, resolving the issue of ineffective fusion of multi-channel information. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 A flow chart of the method for early screening of blood tumors based on a multi-channel blood scattergram provided by the present invention;

[0051] Figure 2 This is a structural block diagram of the blood tumor early screening system based on multi-channel blood scattergram provided by the present invention;

[0052] Figure 3 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0054] In a specific embodiment, Figure 1 As shown, the method for early screening of blood tumors based on a multi-channel blood scattergram provided by the present invention comprises the following steps:

[0055] S110: Obtain a blood scatter plot of at least two blood cell detection dimensions of the target subject; in this embodiment, the scatter plots obtained by the three channels of the blood cell analyzer are used as analysis targets. During operation, a blood sample of the target subject is obtained and input into the blood cell analyzer to obtain a first channel scatter plot, a second channel scatter plot, and a third channel scatter plot generated by the blood cell analyzer; wherein the first channel scatter plot is a blood scatter plot of the white blood cell classification channel, the second channel scatter plot is a blood scatter plot of the basophil / nucleated red blood cell channel, and the third channel scatter plot is a blood scatter plot of the primitive cell / abnormal lymphocyte channel.

[0056] In a specific usage scenario, a blood sample is obtained from the subject, and a fully automatic blood cell analyzer is used to generate blood scatter plot images of three channels: WDF (white blood cell classification channel), WNR (basophil / nucleated red blood cell channel), and WPC (original cell / abnormal lymphocyte channel). Assume that the input scatter plot images of each channel are , , The obtained scatter plot image is preprocessed as necessary. The coordinate axis part of the image is set to the background color according to the pixel position. Then the image size is uniformly adjusted to H×W×C (where H is the image height, W is the image width, and C is the number of channels). The pixel values ​​are normalized (scaling the pixel values ​​to [0, 1]) to make it suitable as input data for the subsequent feature extraction network.

[0057] S120: Inputting each of the blood scattergrams into a pre-built feature extraction network to obtain deep feature maps output by the feature extraction network and corresponding to each of the blood scattergrams;

[0058] S130: Inputting each of the obtained deep feature maps into a pre-trained balance optimization model to obtain an enhanced feature map output by the balance optimization model;

[0059] S140: Inputting each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier;

[0060] S150: fusing the enhanced feature maps to obtain a fused feature map, and obtaining a fused feature category prediction result based on the fused feature map;

[0061] S160: performing weighted calculation on the category prediction results corresponding to each dimension and the fusion feature category prediction results to obtain a final category prediction result.

[0062] In summary, this paper proposes a method and system for early screening of hematologic malignancies based on multi-channel blood scatter plot information fusion and long-tail distribution optimization. The core concept of this method is to extract features from blood scatter plots in three channels: WDF, WPC, and WNR, using a deep learning network. Long-tail classification is then used to enhance the recognition of rare classes (hematologic malignancies). Furthermore, a cross-attention mechanism is used to fuse multi-channel information, ultimately achieving accurate classification of hematologic malignancies.

[0063] In step S130, the training process of the balance optimization model specifically includes:

[0064] Preprocessing the blood scatter plots of each channel of multiple groups of blood samples used for training;

[0065] Each preprocessed blood scatter plot and the corresponding category label are used as a pair of training data, and a training set is constructed;

[0066] Each blood scatter plot in the training set is input into the pre-built feature extraction network for training until the feature extraction network loss of each channel converges and the training is stopped to obtain the trained feature extraction network of the three channels;

[0067] Based on the training feature extraction network, the balanced optimization method is applied to train the classifier.

[0068] Specifically, in the preprocessing stage, the coordinate axes and the coordinate axis letters in the blood scatter plots of each channel corresponding to the blood sample are converted to background colors, and the processed blood scatter plots are saved as bitmaps (PNG format) to form blood scatter plots of the blood sample after preprocessing in each channel. 、 、 , according to the diagnostic category of the blood sample Record labels for the scatter plots of the three channels. Resize the images to a uniform size of H × W × C (where H is the image height, W is the image width, and C is the number of channels), and normalize the pixel values ​​(scaling them to [0, 1]). The preprocessed blood scatter plots and their corresponding labels form the training set.

[0069] For each channel The preprocessed image , independently perform the following operations: Use EfficientNet-B1 as the feature extraction network. For the input image After passing through the EfficientNet-B1 network, the deep feature map of the image is obtained before the global average pooling layer. , whose size is a×b×dim (where a and b are the width and height of the deep feature map, and dim is the number of feature map channels). After global average pooling, the auxiliary classifier is input, and the auxiliary classifier outputs the initial prediction. The cross entropy loss is calculated by combining the initial prediction with the label of the current image and the parameters of the feature extraction network are updated. In the first stage, the training is stopped after the loss of the feature extraction network of each channel converges, and the feature extraction networks of the three channels are trained. , , .

[0070] Based on the training feature extraction network, the balanced optimization method is applied to train the classifier.

[0071] During the above training process:

[0072] Random sampling is used for the preprocessed blood scatter plot data of each channel in the training set, and the randomly sampled scatter plots are recorded as ; Then, weight sampling is performed according to the opposite number of the frequency of occurrence of the category to which the sample belongs in the training set. If the number of times the category to which the current sample belongs appears in the training set is , then the probability of the current sample being sampled is , where j is the total number of categories, and weighted random sampling (Weighted Random Sampling) will be performed according to the inverse of the frequency of occurrence. The scatter plot of weighted random sampling is recorded as .

[0073] Sampling to and Finally, load the feature extraction network trained in the first stage and output the deep feature maps corresponding to the two scatter plots and . Use a random mask with a value of 0 or 1 in the feature channel dimension (The corresponding dimension is 0, which means it is retained The feature of the corresponding dimension is 1, which means it is replaced by Corresponding dimension features), Replace the feature map channels of a specific ratio in In the enhanced feature map ,Right now ,in for The complement of ,the replacement ratio is controlled by the parameter p. The enhanced feature map is average pooled and then predicted by the classifier output balance optimization.

[0074] In step S140, the classifier training process specifically includes:

[0075] After obtaining the balanced optimization prediction, the image Tags and images Tags Perform linear fusion to obtain the fused label ;

[0076] Combine the balanced optimized category predictions with the fused labels Calculate the cross entropy loss and update the parameters of the classifier to obtain a trained three-channel classifier.

[0077] Specifically, during training:

[0078] In the first stage, the feature extraction network is trained using cross entropy loss. In the second stage, the parameters of the trained feature extraction network are frozen and the classifier is fine-tuned by fusion loss. The model predicts The loss is calculated with the fusion label, and the loss ratio is controlled by the fusion ratio p.

[0079] In step S150, the category prediction results corresponding to each dimension and the fusion feature category prediction results are weighted to obtain a final category prediction result, which specifically includes:

[0080] Set the feature output corresponding to the first channel scatter plot to The obtained category prediction result is , the feature output corresponding to the second channel scatter plot is The obtained category prediction result is , the feature output corresponding to the third channel scatter plot is The prediction results are ;

[0081] The three channels are spliced ​​and input into the fully connected layer and then activated by ReLU to map the fusion features. The fusion features are input into the fusion feature classifier to predict the category. ,right Calculate cross entropy loss with labels and update the fully connected layer and fusion classifier;

[0082] The category prediction results of each channel and the category prediction results of the fusion feature are weighted and calculated to obtain the final category prediction result.

[0083] For ease of understanding, the process of feature extraction and balance optimization for each single channel is described as a whole below. It should be understood that the process of feature extraction and balance optimization for other channels is the same.

[0084] For each channel The preprocessed image , independently perform the following operations:

[0085] 1. Basic feature extraction network:

[0086] EfficientNet-B1 is used as the basic feature extraction network. EfficientNet uses a unified composite coefficient To jointly expand the depth d, width w and input resolution r of the network, the basic scaling rule can be expressed as: , , The constraints are ,in , , EfficientNet-B1 is a specific configuration in this series that uses the MBConv (Mobile Inverted Bottleneck Convolution) module as its core building block. The MBConv module contains a point-by-point convolution (1×1 convolution) for channel expansion, a depthwise separable convolution for spatial filtering, and another point-by-point convolution for channel compression. It also typically includes a Squeeze-and-Excitation (SE) optimization module to dynamically adjust feature responses between channels.

[0087] For the input image , after the EfficientNet-B1 network After processing, the deep feature map of the channel is obtained before the global average pooling layer .

[0088] 2. Balance optimization

[0089] Because blood cancer samples are relatively scarce compared to healthy samples, they exhibit a typical long-tail distribution problem. To mitigate the model's bias toward head-class samples and improve the recognition of tail-class (tumor) samples, the H2T (Head-to-Tail Fusion) method is introduced after the feature extraction network or during classifier training. This method aims to enhance the diversity and semantic expression of tail-class samples by transferring the rich feature space information from head-class samples to tail-class samples.

[0090] Fusion at the feature map level: Random sampling is used for the blood scatter plot data after preprocessing of each channel in the training set, and the randomly sampled scatter plots are recorded as ; Weight sampling is performed based on the opposite number of the frequency of occurrence of the category to which the sample belongs in the training set. If the number of times the category to which the current sample belongs appears in the training set is , then the probability of the current sample being sampled is , where j is the total number of categories, weighted random sampling will be performed according to the inverse of the frequency of occurrence, and the scatter plot of weighted random sampling is recorded as ; Sampling to and After that, load the trained feature extraction network and output the deep feature maps corresponding to the two scatter plots and , using a random mask with a value of 0 or 1 in the feature channel dimension ,Will Replace the feature map channels of a specific ratio in In the enhanced feature map ,in for The complement of It is an element-wise matrix multiplication operation, and the replacement ratio is controlled by the parameter p.

[0091] Two-stage training strategy: H2T is applied in the second stage of two-stage training (classifier adjustment stage). The first stage uses cross entropy loss to train the feature extractor. The second stage freezes the parameters of the feature extractor and fine-tunes the classifier through the fusion loss. The model predicts The loss is calculated separately for the labels of blood tumor class t and healthy class h and the loss ratio is controlled by the fusion ratio p.

[0092] 3. Multi-channel decision fusion:

[0093] In order to comprehensively utilize the information of the three channels and capture the intrinsic correlation between them, the feature vectors output by each channel are fused. The feature outputs of the three channels are , , , the predictions are , , The three channels are concatenated and the fusion features are mapped through the fully connected layer. , predicting categories by fusing features , and constrained by cross entropy loss. The predictions of the three channels and the fusion features are normalized and weighted to obtain the final result, which is ,in is the fusion weight for each channel.

[0094] In the above specific embodiments, the method and system for early screening of blood tumors provided by the present invention have at least the following technical effects:

[0095] Specific fusion of multi-channel information: A fusion strategy of three specific blood scatter plot channel information, WDF, WNR, and WPC, was proposed and applied for early screening of blood tumor diseases.

[0096] Targeted basic network application: EfficientNet-B1 is selected as the basic network for feature extraction for each specific channel mentioned above to balance computational efficiency and feature extraction capabilities.

[0097] Introduction of advanced long-tail classification strategy: To address the classification bias problem caused by the scarcity of blood tumor samples, the H2T long-tail classification method is innovatively applied to the basic network of each channel for model enhancement, significantly improving the ability to identify rare tumor samples.

[0098] Efficient feature and prediction fusion mechanism: A cross-attention mechanism is used to deeply fuse feature information and / or prediction information from three enhanced channels, effectively capturing complementary information between multiple channels and further improving classification accuracy.

[0099] Holistic screening method and system: The constructed complete method workflow, from multi-channel data acquisition, single-channel enhancement processing, to multi-channel fusion decision-making, forms a novel and efficient technical solution for early screening of hematological malignancies with potential clinical application value. This method can be integrated into a computer-assisted screening system for automated processing.

[0100] Through the above technical solution, the present invention can more effectively deal with the category imbalance problem in blood scatter plot data, make full use of multi-channel information, thereby significantly improving the accuracy and recall rate of early screening of blood tumor diseases, especially rare types of tumors, and providing more powerful auxiliary diagnostic tools for clinicians.

[0101] In addition to the above method, the present invention also provides a blood tumor early screening system based on multi-channel blood scattergram, such as Figure 2 As shown, the system includes:

[0102] a scattergram acquisition unit 310 for acquiring a blood scattergram of at least two blood cell detection dimensions of a target subject;

[0103] A feature extraction unit 320 is configured to input each of the blood scattergrams into a pre-built feature extraction network to obtain deep feature maps output by the feature extraction network that correspond to each of the blood scattergrams;

[0104] A feature enhancement unit 330 is configured to input each of the obtained deep feature maps into a pre-trained balance optimization model to obtain an enhanced feature map output by the balance optimization model;

[0105] A category prediction unit 340 is configured to input each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier;

[0106] A prediction result fusion unit 350 is configured to fuse the enhanced feature maps to obtain a fused feature map, and obtain a fused feature category prediction result based on the fused feature map;

[0107] The screening result generating unit 360 is configured to perform weighted calculation on the category prediction results corresponding to each dimension and the fusion feature category prediction results to obtain a final category prediction result.

[0108] In some embodiments, obtaining a blood scattergram of at least two blood cell detection dimensions of a target subject specifically includes:

[0109] obtaining a blood sample from the target subject;

[0110] Inputting the blood sample into a blood cell analyzer to obtain a first channel scatter plot, a second channel scatter plot, and a third channel scatter plot generated by the blood cell analyzer;

[0111] The first channel scatter plot is a blood scatter plot of a white blood cell classification channel, the second channel scatter plot is a blood scatter plot of a basophil / nucleated red blood cell channel, and the third channel scatter plot is a blood scatter plot of a primitive cell / abnormal lymphocyte channel.

[0112] In some embodiments, the training process of the balance optimization model specifically includes:

[0113] Preprocessing the blood scatter plots of each channel of multiple groups of blood samples used for training;

[0114] Each preprocessed blood scatter plot and the corresponding category label are used as a pair of training data, and a training set is constructed;

[0115] Each blood scatter plot in the training set is input into the pre-built feature extraction network for training until the feature extraction network loss of each channel converges and the training is stopped to obtain the trained feature extraction network of the three channels;

[0116] Based on the training feature extraction network, the balanced optimization method is applied to train the classifier.

[0117] In some embodiments, during training:

[0118] Random sampling is used for the preprocessed blood scatter plot data of each channel in the training set, and the randomly sampled scatter plots are recorded as ;

[0119] Weighted sampling is performed based on the opposite number of the frequency of occurrence of the category to which the sample belongs in the training set. If the number of times the category to which the current sample belongs appears in the training set is , then the probability of the current sample being sampled is , where j is the total number of categories, weighted random sampling will be performed according to the inverse of the frequency of occurrence, and the scatter plot of weighted random sampling is recorded as ;

[0120] Sampling to and After that, load the trained feature extraction network and output the deep feature maps corresponding to the two scatter plots and , using a random mask with a value of 0 or 1 in the feature channel dimension ,Will Replace the feature map channels of a specific ratio in , forming an enhanced feature map ,in for The complement of , the replacement ratio is controlled by the parameter p;

[0121] The enhanced feature map is average pooled and then the prediction is optimized after the classifier output balance.

[0122] In some embodiments, the classifier training process specifically includes:

[0123] After obtaining the balanced optimization prediction, the image Tags and images Tags Perform linear fusion to obtain the fused label ;

[0124] Combine the balanced optimized category predictions with the fused labels Calculate the cross entropy loss and update the parameters of the classifier to obtain a trained three-channel classifier.

[0125] In some embodiments, during training:

[0126] In the first stage, the feature extraction network is trained using cross entropy loss. In the second stage, the parameters of the trained feature extraction network are frozen and the classifier is fine-tuned by fusion loss. The model predicts The loss is calculated with the fusion label, and the loss ratio is controlled by the fusion ratio p.

[0127] In some embodiments, the category prediction results corresponding to each dimension and the fusion feature category prediction results are weighted to obtain a final category prediction result, specifically including:

[0128] Set the feature output corresponding to the first channel scatter plot to The obtained category prediction result is , the feature output corresponding to the second channel scatter plot is The obtained category prediction result is , the feature output corresponding to the third channel scatter plot is The prediction results are ;

[0129] The three channels are spliced ​​and input into the fully connected layer and then activated by ReLU to map the fusion features. The fusion features are input into the fusion feature classifier to predict the category. ,right Calculate cross entropy loss with labels and update the fully connected layer and fusion classifier;

[0130] The category prediction results of each channel and the category prediction results of the fusion feature are weighted and calculated to obtain the final category prediction result.

[0131] In the above-mentioned specific embodiment, the method and system for early screening of blood tumors based on multi-channel blood scattergrams provided by the present invention obtains blood scattergrams of at least two blood cell detection dimensions of the target subject, inputs each of the blood scattergrams into a pre-constructed feature extraction network to obtain deep feature maps corresponding to each of the blood scattergrams output by the feature extraction network; inputs each of the obtained deep feature maps into a pre-trained balanced optimization model to obtain an enhanced feature map output by the balanced optimization model; inputs each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier; fuses each of the enhanced feature maps to obtain a fused feature map, and obtains a fused feature category prediction result based on the fused feature map; performs weighted calculation on the category prediction results corresponding to each dimension and the fused feature category prediction result to obtain a final category prediction result, in the form of a predicted probability for each category, and takes the maximum value of the category prediction result to output the prediction result for auxiliary screening. This achieves computer-assisted screening, reduces the empirical dependence of early screening of blood tumors, and improves the accuracy and efficiency of screening results.

[0132] In the screening of hematologic malignancies, the number of samples from healthy individuals is far greater than that from patients with the disease. This is especially true for rare tumor types, for which available samples are even scarcer. This sample imbalance (i.e., long-tail distribution) can cause machine learning models to favor the majority class (healthy samples) during training, resulting in weaker recognition of the minority class (tumor samples), leading to high missed diagnosis rates or insufficient specificity. In some embodiments, the methods and systems provided by the present invention incorporate H2T (Head-to-Tail Fusion, a feature fusion strategy in long-tail visual recognition) after the feature extraction network or during classifier training. This approach transfers the rich feature space information from the head class samples to the tail class samples, enhancing their diversity and semantic expression, thereby addressing the data bias problem caused by sample imbalance.

[0133] In practical applications, because different channels in a blood scatter plot (such as WDF, WNR, and WPC) reflect distinct physical and chemical properties of cells, existing methods fail to fully integrate and utilize this multi-channel information, reducing their classification capabilities. In some embodiments, the methods and systems provided by the present invention fuse multi-channel information through a cross-attention mechanism to achieve accurate classification of blood cancer diseases, resolving the issue of ineffective fusion of multi-channel information.

[0134] Figure 3 An example of a physical structure diagram of an electronic device is shown below. Figure 3 As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 may call the logic instructions in the memory 430 to execute the above method.

[0135] Furthermore, the logic instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0136] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the above method.

[0137] In yet another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is configured to execute the above method when executed by a processor.

[0138] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0139] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for early screening of blood tumors based on multi-channel blood scattergram, characterized in that: The method comprises: Obtaining a blood scatter plot of at least two blood cell detection dimensions of the target subject; Inputting each of the blood scattergrams into a pre-built feature extraction network to obtain deep feature maps output by the feature extraction network that correspond to each of the blood scattergrams; Inputting each of the obtained deep feature maps into a pre-trained balanced optimization model to obtain an enhanced feature map output by the balanced optimization model; Inputting each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier; Fusing the enhanced feature maps to obtain a fused feature map, and obtaining a fused feature category prediction result based on the fused feature map; Perform weighted calculation on the category prediction results corresponding to each dimension and the fusion feature category prediction results to obtain the final category prediction result; The step of obtaining a blood scatter plot of at least two blood cell detection dimensions of the target subject specifically includes: obtaining a blood sample from the target subject; Inputting the blood sample into a blood cell analyzer to obtain a first channel scatter plot, a second channel scatter plot, and a third channel scatter plot generated by the blood cell analyzer; Wherein, the first channel scatter plot is a blood scatter plot of a white blood cell classification channel, the second channel scatter plot is a blood scatter plot of a basophil / nucleated red blood cell channel, and the third channel scatter plot is a blood scatter plot of a primitive cell / abnormal lymphocyte channel; The training process of the balance optimization model specifically includes: Preprocessing the blood scatter plots of each channel of multiple groups of blood samples used for training; Each preprocessed blood scatter plot and the corresponding category label are used as a pair of training data, and a training set is constructed; Each blood scatter plot in the training set is input into the pre-built feature extraction network for training until the feature extraction network loss of each channel converges and the training is stopped to obtain the trained feature extraction network of the three channels; Based on the training feature extraction network, the balanced optimization method is applied to train the classifier.

2. The method for early screening of blood tumors based on multi-channel blood scattergram according to claim 1, characterized in that: During training: Random sampling is used for the preprocessed blood scatter plot data of each channel in the training set, and the randomly sampled scatter plots are recorded as ; Weighted sampling is performed based on the opposite number of the frequency of occurrence of the category to which the sample belongs in the training set. If the number of times the category to which the current sample belongs appears in the training set is , then the probability of the current sample being sampled is , where j is the total number of categories, weighted random sampling will be performed according to the inverse of the frequency of occurrence, and the scatter plot of weighted random sampling is recorded as ; Sampling to and After that, load the trained feature extraction network and output the deep feature maps corresponding to the two scatter plots and , using a random mask with a value of 0 or 1 in the feature channel dimension ,Will Replace the preset ratio of feature map channels in In the enhanced feature map ,in for The complement of It is an element-wise multiplication operation of the matrix, and the replacement ratio is controlled by the parameter p; The enhanced feature map is average pooled and then the prediction is optimized after the classifier output balance.

3. The method for early screening of blood tumors based on multi-channel blood scattergram according to claim 2, characterized in that: The training process of the classifier specifically includes: After obtaining the balanced optimization prediction, the image Tags and images Tags Perform linear fusion to obtain the fused label ; Combine the balanced optimized category predictions with the fused labels Calculate the cross entropy loss and update the parameters of the classifier to obtain a trained three-channel classifier.

4. The method for early screening of blood tumors based on multi-channel blood scattergram according to claim 3, characterized in that: During training: In the first stage, the feature extraction network is trained using cross entropy loss. In the second stage, the parameters of the trained feature extraction network are frozen and the classifier is fine-tuned by fusion loss. The model predicts The loss is calculated separately for the labels of blood tumor class t and healthy class h, and the loss ratio is controlled by the fusion ratio p.

5. The method for early screening of blood tumors based on multi-channel blood scattergram according to claim 4, characterized in that: The category prediction results corresponding to each dimension and the fusion feature category prediction results are weighted to obtain the final category prediction result, specifically including: Set the feature output corresponding to the first channel scatter plot to The obtained category prediction result is , the feature output corresponding to the second channel scatter plot is The obtained category prediction result is , the feature output corresponding to the third channel scatter plot is The prediction results are ; The three channels are spliced ​​and input into the fully connected layer and then activated by ReLU to map the fusion features. The fusion features are input into the fusion feature classifier to predict the category. ,right Calculate cross entropy loss with labels and update the fully connected layer and fusion classifier; The category prediction results of each channel and the category prediction results of the fusion feature are weighted and calculated to obtain the final category prediction result.

6. A multi-channel blood scattergram-based early screening system for blood tumors, used to implement the method according to any one of claims 1 to 5, characterized in that: The system comprises: a scattergram acquisition unit, configured to acquire a blood scattergram of at least two blood cell detection dimensions of a target subject; a feature extraction unit, configured to input each of the blood scattergrams into a pre-built feature extraction network to obtain deep feature maps output by the feature extraction network and corresponding to each of the blood scattergrams; A feature enhancement unit, configured to input each of the obtained deep feature maps into a pre-trained balance optimization model to obtain an enhanced feature map output by the balance optimization model; A category prediction unit, configured to input each of the enhanced feature maps into a pre-trained classifier to obtain a category prediction result output by the classifier; A prediction result fusion unit is used to fuse the enhanced feature maps to obtain a fused feature map, and obtain a fused feature category prediction result based on the fused feature map; The screening result generating unit is used to perform weighted calculation on the category prediction results corresponding to each dimension and the fusion feature category prediction results to obtain the final category prediction result.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images