Identification method of bone marrow hematopoietic cell dysplasia based on neural network

Through the neural network-based method, the VGG classification neural network model is constructed to train the bone marrow cells and nucleus sample sets, which solves the problem of subjective factors in the diagnosis of bone marrow hematopoietic cell development abnormality, and realizes the accurate identification of bone marrow hematopoietic cell development abnormality and Diagnosis improves the objectivity and efficiency of diagnosis.

CN118968509APending Publication Date: 2024-11-15INST OF HEMATOLOGY & BLOOD DISEASES HOSPITAL CHINESE ACADEMY OF MEDICAL SCI & PEKING UNION MEDICAL COLLEGE
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411014780.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art has subjective factors in the diagnosis of abnormal development of bone marrow hematopoietic cells, resulting in low accuracy of interpretation results. The artificial intelligence-assisted reading function has not yet been applied in detail, making it difficult to effectively evaluate the proportion of abnormal development cells.

Method used

Using a neural network-based method, a VGG classification neural network model is constructed by obtaining bone marrow cell images, training the cell and nucleus sample sets, using ROC curves to obtain thresholds, and establishing a relationship function between the number of abnormal developmental cells and integral values. Achieve objective recognition of abnormal development of bone marrow cells.

Benefits of technology

The diagnostic accuracy and work efficiency of bone marrow hematopoietic cell development abnormalities have been improved, and the objectification and quantification of the interpretation standards of three-line development abnormalities have been achieved, reducing the influence of subjective factors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968509A_ABST
    Figure CN118968509A_ABST
Patent Text Reader

Abstract

The invention provides a bone marrow hematopoietic cell dysplasia identification method based on a neural network, and the method comprises the following steps: dividing bone marrow cell images according to cell types to obtain a plurality of cell sample sets, extracting cell nucleus images in the bone marrow cell images in the cell sample sets to obtain a plurality of cell nucleus sample sets, separately training the neural network by using a cell sample set and a cell nucleus sample set of different cell types, obtaining a threshold value corresponding to a maximum dentate index according to an ROC curve of a neural network model prediction result, and setting an integral value used for quantifying the prediction result, and establishing a relation function between the number of the dysplastic cells and the integral value based on the obtained threshold value, and obtaining an anomaly recognition result of the current bone marrow cell image. The three-line dysplasia interpretation method has the advantages that three-line dysplasia interpretation is objective and quantitative, the working efficiency of bone marrow morphology inspection personnel can be improved, and the interpretation accuracy can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of neural networks, and in particular relates to a method for identifying abnormal development of bone marrow hematopoietic cells based on neural networks. Background Art

[0002] Myelodysplastic syndrome (MDS) and myelodysplastic / myeloproliferative neoplasms (MDS / MPN) are malignant tumor diseases originating from hematopoietic stem / progenitor cells. Whether the proportion of dysplastic cells in the granulocyte, erythroid and megakaryocytic lineages is greater than 10% is of great significance for the diagnosis and prognosis of these two diseases. At present, the evaluation of three-lineage dysplasia mainly relies on classification under a manual microscope, that is, whether the granulocyte, erythroid and megakaryocytic lineages are accompanied by dysplasia is mainly judged under a manual microscope. It not only requires a lot of effort to learn and identify, but the identification of dysplasia during manual classification is easily introduced into the subjective factors. Therefore, the interpretation results are easily affected by subjective factors, thereby affecting the accuracy of the interpretation of dysplasia results.

[0003] At present, AI-assisted film reading systems are involved in various fields such as pathology and morphology, and some technologies have been applied in actual work. Although they have improved working methods and work efficiency, the functions of AI-assisted film reading are currently mostly used for initial screening and have not been well penetrated into the details of the application. Summary of the invention

[0004] In view of this, the present invention aims to propose a method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network, in order to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] A method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network comprises the following steps:

[0007] Acquire bone marrow cell images, divide the bone marrow cell images according to cell types to obtain multiple cell sample sets, and extract nucleus images from the bone marrow cell images in the cell sample sets to obtain multiple nucleus sample sets;

[0008] Constructing a neural network model, and using the cell sample set and the nucleus sample set to train the neural network model; wherein, in the process of training the neural network, the cell sample sets and the nucleus sample sets of different cell types are used to train the neural network separately;

[0009] According to the ROC curve of the prediction results of the neural network model, a threshold corresponding to the maximum Youden index is obtained, and an integral value for quantifying the prediction results is set, and a relationship function between the number of dysplastic cells and the integral value is established based on the obtained threshold;

[0010] Substitute the output results of the trained neural network model into the relationship function to obtain the abnormal recognition results of the current bone marrow cell image.

[0011] Furthermore, the process of acquiring bone marrow cell images, dividing the bone marrow cell images according to cell types, and obtaining multiple cell sample sets includes:

[0012] Scan the patient's bone marrow smear specimens with a high-throughput scanning device to obtain the corresponding bone marrow cell morphology images and screen out single cell images with clear cell structures;

[0013] All bone marrow cell morphology images are divided according to the custom cell types; wherein the custom cell types are specifically:

[0014] Normal granulocytes, granulocytes with somatic / cytoplasmic dysplasia and granulocytes with nuclear dysplasia, normal nucleated erythrocytes, nucleated erythrocytes with somatic / cytoplasmic dysplasia and nucleated erythrocytes with nuclear dysplasia, normal megakaryocytes, megakaryocytes with somatic dysplasia and megakaryocytes with nuclear dysplasia.

[0015] Furthermore, the obtained multiple cell sample sets and multiple nucleus sample sets of different cell types are divided into training sets, validation sets, and test sets respectively; wherein the ratios of the three sets, the training set, the validation set, and the test set, are 8:1:1, respectively.

[0016] Furthermore, the neural network model is a VGG classification neural network, and the process of constructing the neural network model includes:

[0017] Construct the convolution layer, maximum pooling layer, and first fully connected layer of the neural network model; wherein the convolutional neural network uses the convolution layer and the maximum pooling layer to extract the features of the bone marrow cell image, and the first fully connected layer flattens the features of the bone marrow cell image extracted by the convolution layer and the maximum pooling layer.

[0018] Furthermore, before executing the operation of using the cell sample set and the nucleus sample set to train the neural network model, the following operations are performed:

[0019] The cell sample set and the nucleus sample set are encapsulated into a data structure of a specified format, the encapsulated cell and nucleus sample sets are shuffled and randomly divided into multiple batches, and the multiple batches of cell and nucleus sample sets are transmitted to the neural network model in parallel.

[0020] Further, after executing the operation of training the neural network model using the cell sample set and the nucleus sample set, the following operations are performed:

[0021] The neural network model also includes a second fully connected layer to evaluate the prediction accuracy of the trained neural network model. When the prediction accuracy is stable, the output results of the neural network model trained using the cell sample set and the output results of the neural network model trained using the nucleus sample set are input into the second fully connected layer for data analysis to obtain abnormal prediction values ​​of the cell body, cytoplasm, and nucleus.

[0022] Furthermore, the process of obtaining the threshold value corresponding to the maximum Youden index according to the ROC curve of the prediction result of the neural network model includes:

[0023] The prediction accuracy of the trained neural network model is statistically analyzed to obtain the true positive rate and the false positive rate, and the true positive rate and the false positive rate are used as two coordinates to draw a ROC curve diagram of the prediction results of the neural network model; wherein the ROC curve diagram is a three-line ROC curve diagram including granulocytes, nucleated red blood cells, and megakaryocytes;

[0024] The maximum Youden index of the three-system prediction results was calculated respectively according to the three-system ROC curve diagram.

[0025] Furthermore, the process of setting the integral value of the quantified prediction result and establishing the relationship function between the number of dysplastic cells and the integral value based on the obtained threshold value includes:

[0026] The dysplasia integral value is obtained by calculating the product of the threshold and the number of cells with dysplasia identified by the neural network model, and judging whether the integral value of the current bone marrow cell image reaches the corresponding standard according to the WHO definition of bone marrow hematopoietic cell dysplasia;

[0027] If reached, it is determined that the patient's bone marrow cells corresponding to the current bone marrow cell image have abnormal development.

[0028] Compared with the prior art, the neural network-based method for identifying abnormal bone marrow hematopoietic cell development in the present invention has the following beneficial effects:

[0029] The method of the present invention can achieve an objective assessment of the proportion of dysplastic cells, thereby establishing a more objective and accurate diagnosis model for MDS and MDS / MPN;

[0030] The method of the present invention can make the interpretation criteria of trilineage dysplasia objective and quantitative, which not only improves the work efficiency of bone marrow morphology inspectors, but also improves the accuracy of interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0032] Figure 1 This is a flow chart of a method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to an embodiment of the present invention;

[0033] Figure 2 Schematic diagram of VGG classification neural network. DETAILED DESCRIPTION

[0034] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0035] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0036] Bone marrow smear specimens contain more than ten types of cells. Different types of cells are artificially divided into different differentiation stages according to the differentiation of cells. Granulocytes are divided into three types: neutrophils, eosinophils, and basophils, each of which is divided into six stages: primitive, early immature, intermediate immature, late immature, band nucleus, and lobed nucleus; nucleated red blood cells are divided into four stages: primitive, early immature, intermediate immature, and late immature; megakaryocytes are divided into five stages: primitive, immature, granular, platelet-producing, and naked nucleus. The types of developmental abnormalities involve It also affects the cell size (such as erythroid megaloblastic degeneration, giant lobed granulocytes, micromegakaryocytes, micromegakaryocytes), cytoplasmic abnormalities (erythroid cytoplasmic vacuoles, erythroid cytoplasmic pale areas, reduced or absent granulocyte cytoplasmic granules, pseudo-CH granules, Dühler bodies, etc.) and nuclear abnormalities (erythroid multinucleation, internuclear bridges, nuclear fragmentation, poor granulocyte nuclear lobing, drumstick bodies, abnormal nuclear chromatin condensation, pseudo-PH deformity, mononuclear megakaryocytes, binuclear megakaryocytes, multinuclear megakaryocytes).

[0037] Reduced or absent cytoplasmic granules, pseudo-CH granules, and Dühle bodies can be seen in granulocytes at the four stages of medullary, late medullary, band-shaped, and lobed nuclei; poor nuclear lobing and abnormal nuclear chromatin aggregation are often seen in granulocytes at the late medullary, band-shaped, and lobed nuclei stages; drumstick bodies and pseudo-PH deformities are often seen in granulocytes at the lobed nucleus stage; giant lobed granulocytes are seen in granulocytes at the lobed nucleus stage; megaloblastic changes, multinuclei, internuclear bridges, nuclear fragmentation, and cytoplasmic pale areas of the erythroid system are often seen in nucleated erythrocytes at the medullary and late medullary stages; erythroid cytoplasmic vacuoles are often seen in nucleated erythrocytes at the early medullary and medullary stages; megakaryotic developmental abnormalities (micromegakaryocytes, micromegakaryocytes, mononuclear megakaryocytes, binuclear megakaryocytes, and multinuclear megakaryocytes) are concentrated in granular and platelet-producing megakaryocytes.

[0038] Such a wide variety of developmental abnormalities not only need to be differentiated from normal cell morphology, but different types of developmental abnormalities may also coexist and affect the determination of developmental abnormalities. For example, the granulocyte system may simultaneously have poor nuclear lobulation + abnormal nuclear chromatin condensation, pseudo-PH deformity + reduced or absent cytoplasmic granules, reduced or absent cytoplasmic granules + Dühler bodies, and poor nuclear lobulation + reduced or absent cytoplasmic granules + Dühler bodies.

[0039] Based on the complexity and judgment difficulty of the above-mentioned human bone marrow cell images, the present invention proposes a method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network, which includes the following steps:

[0040] Step 1: Obtain bone marrow cell images, divide the bone marrow cell images according to cell types to obtain multiple cell sample sets, and extract nucleus images from the bone marrow cell images in the cell sample sets to obtain multiple nucleus sample sets.

[0041] The process of obtaining bone marrow cell images in step 1, dividing the bone marrow cell images according to cell types, and obtaining multiple cell sample sets includes:

[0042] Scan the patient's bone marrow smear specimens with a high-throughput scanning device to obtain the corresponding bone marrow cell morphology images, and manually screen out single cell images with clear cell structures;

[0043] In order to enable the VGG neural network constructed later to better capture the structural features of different parts of cells, all bone marrow cell morphology images are divided according to custom cell types; wherein the custom cell types are specifically:

[0044] Normal granulocytes, granulocytes with somatic / cytoplasmic dysplasia and granulocytes with nuclear dysplasia, normal nucleated erythrocytes, nucleated erythrocytes with somatic / cytoplasmic dysplasia and nucleated erythrocytes with nuclear dysplasia, normal megakaryocytes, megakaryocytes with somatic dysplasia and megakaryocytes with nuclear dysplasia.

[0045] The multiple nucleus sample sets obtained by extracting nucleus images from bone marrow cell images in the cell sample set specifically include:

[0046] The nuclei in the cell sample set were cut out and saved as 6 folders: normal granulocyte nuclei, granulocyte nuclei with nuclear dysplasia, normal nucleated erythrocyte nuclei, nucleated erythrocyte nuclei with nuclear dysplasia, normal megakaryocyte nuclei, and megakaryocyte nuclei with nuclear dysplasia.

[0047] Preferably, the obtained multiple cell sample sets and multiple nucleus sample sets of different cell types are divided into a training set, a validation set, and a test set, respectively; wherein the ratio of the three sets, the training set, the validation set, and the test set, is 8:1:1, respectively.

[0048] Step 2: construct a neural network model, and use the cell sample set and the nucleus sample set to train the neural network model; wherein, in the process of training the neural network, the neural network is trained separately using the cell sample set and the nucleus sample set of different cell types.

[0049] The neural network model disclosed in step 2 is specifically: VGG classification neural network;

[0050] The process of building a neural network model includes:

[0051] Construct the convolutional layer, maximum pooling layer, first fully connected layer, and second fully connected layer of the neural network model;

[0052] The convolution layer is responsible for the convolution operation in the neural network. It can extract the features in each picture and help the neural network to intelligently and dynamically identify the target. The features extracted by the convolution layer are used as the input information of the fully connected layer. The neurons in the first fully connected layer apply a nonlinear activation function and output σ(Wx+b), where x represents the input feature, σ represents the nonlinear activation function, W represents the weight, and b represents the bias. Then, the ReLU activation function is selected, that is, f(x)=max(0,x), and the second fully connected layer is entered. Finally, the softmax activation function is used, that is, Output a classification result, where c represents the number of categories.

[0053] Among them, the convolutional neural network uses convolutional layers and maximum pooling layers to extract the features of bone marrow cell images. The first fully connected layer flattens the features of the bone marrow cell images extracted by the convolutional layer and the maximum pooling layer. The purpose of the maximum pooling layer is to reduce the number of features of the convolutional layer, thereby reducing the complexity of the model and avoiding overfitting of the model.

[0054] Specifically, the Flatten function is used for flattening; the Flatten function is a function of numpy.ndarray.flatten, that is, it returns a one-dimensional array, which can realize the flattening of a multidimensional vector into a one-dimensional vector. In the entire convolutional neural network, the convolution layer is used to learn and extract features from the data, and the maximum pooling layer passes the learning information to the fully connected layer. The fully connected layer learns how to make accurate predictions. Therefore, we input the data into the convolutional neural network, and the convolutional neural network will automatically learn and extract the best features for classification.

[0055] Each sample in the six folders of normal granulocytes, granulocytes with abnormal cell body / cytoplasm development, normal nucleated erythrocytes, nucleated erythrocytes with abnormal cell body / cytoplasm development, normal megakaryocytes, and megakaryocytes with abnormal cell body development will get a predicted value after the VGG classification neural network trained in the above steps. The difference between the predicted value and the true value is the loss. In order to minimize the difference, we choose Where m is the number of samples, k is the number of categories, and y ij represents the true value of the jth label of the i-th sample, Represents the predicted value of the jth label of the i-th sample.

[0056] Preferably, before executing the operation of using the cell sample set and the nucleus sample set to train the neural network model, the following operations are first performed:

[0057] The cell sample set and the nucleus sample set are encapsulated into a data structure recognizable by Python through Dataset, the encapsulated cell sample set and the nucleus sample set are shuffled and randomly divided into multiple batches, and the multiple batches of cell sample sets and the nucleus sample sets are passed into the neural network model in parallel through Dataloader.

[0058] Preferably, after executing the operation of training the neural network model using the cell sample set and the nucleus sample set, the following operations are performed:

[0059] Use the evaluate function to evaluate the prediction accuracy of the trained neural network model When the prediction accuracy is stable, the output results of the neural network model trained using the cell sample set and the output results of the neural network model trained using the nucleus sample set are input into the second fully connected layer for data analysis to obtain the abnormal prediction values ​​of the cell body, cytoplasm, and nucleus. The current cell is judged to be abnormal based on the abnormal prediction values. If the abnormal prediction values ​​of the cell body, cytoplasm, and nucleus of the current cell are all normal, the current cell is judged to be normal, otherwise the current cell is considered to be abnormally developed.

[0060] Among them, the second fully connected layer is used to summarize the abnormal prediction values, and the prediction results of the neural network model obtained by training the cell sample set and the neural network model obtained by training the nucleus sample set are summarized in the second fully connected layer. If there is an abnormal prediction result obtained by the neural network model, the second fully connected layer can output the judgment result of the cell abnormality.

[0061] In this example, the image is scaled to 224×224, the neural network learning rate is set to 0.001, the number of classifications of the neural network classifier is set to correspond to the number of categories of the image, the batch size of the training set and the validation set is set to 32, the training set is trained for 300 epochs, and the classification accuracy of the neural network is verified every 20 epochs. The validation set images are input into the neural network, and the classification accuracy is calculated based on whether the classification is normal or not. When the accuracy of the neural network no longer increases or reaches 100%, the training of the neural network can be stopped.

[0062] Epoch in model training refers to the process of forward propagation and backward propagation performed once on the entire training set when training the model. Epoch is an important indicator to measure the training process. An epoch means that each sample in the training set has been processed by the model at least once. The size of epoch can vary depending on the problem. Common values ​​are 10, 100, 200, etc. The size of epoch directly affects the training time and performance of the model. Smaller epochs may cause underfitting of the model, while larger epochs may cause overfitting. Therefore, choosing the appropriate epoch size is an important hyperparameter adjustment task. Here, the value of epoch is 300.

[0063] At this point, the neural network model obtained after training can identify the developmental abnormalities in any part of the cell body, cytoplasm, and nucleus in the received bone marrow cell images, and determine whether the cells have developmental abnormalities.

[0064] Step 3: Obtain the threshold corresponding to the maximum Youden index based on the ROC curve of the prediction results of the neural network model, set the integral value used to quantify the prediction results, and establish a relationship function between the number of dysplastic cells and the integral value based on the obtained threshold.

[0065] Preferably, the process of obtaining the threshold value corresponding to the maximum Youden index according to the ROC curve of the prediction result of the neural network model includes:

[0066] The prediction accuracy of the trained neural network model is statistically analyzed to obtain the true positive rate TPR and the false positive rate FPR.

[0067] When analyzing the ROC curve, the area under the curve (AUC) can be used to evaluate the performance of the model that generates the curve. A large AUC means that the model can distinguish different categories more accurately. A small AUC means that the prediction accuracy made by the model is not high and the prediction results are often wrong. An ROC falling on the diagonal means that the accuracy of the model's prediction results is no higher than the accuracy of random predictions.

[0068] The true positive rate and the false positive rate are used as two coordinates to draw a ROC curve diagram of the prediction results of the neural network model; wherein the ROC curve diagram is a three-line ROC curve diagram including granulocytes, nucleated red blood cells, and megakaryocytes;

[0069] Specifically, the roc_curve function is executed to obtain the false positive rate and the true positive rate to draw the ROC curve of the three systems, and the maximum Youden index is obtained according to the ROC curve (wherein the maximum Youden index = TPR + (1-FPR) - 1), and the obtained maximum Youden index is used as the k value required for subsequent steps.

[0070] The process of setting the integral value for quantifying the prediction result and establishing the relationship function between the number of dysplastic cells and the integral value based on the obtained threshold value includes:

[0071] The dysplasia integral value is obtained by calculating the product of the threshold and the number of cells with dysplasia identified by the neural network model. Specifically, the number of cells with dysplasia identified is the x-axis, the dysplasia integral value is the y-axis, and the relationship y=kx is established, where the k value is the threshold corresponding to the maximum Youden index for dysplasia identification of cells of each lineage.

[0072] According to the WHO (World Health Organization) definition of bone marrow hematopoietic cell dysplasia (the current WHO definition of bone marrow hematopoietic cell dysplasia is that if 100 cells of a lineage are counted separately and the proportion of dysplastic cells is ≥ 10%, it can be considered that the cells of the lineage have dysplasia), determine whether the integral value of the current bone marrow cell image meets the corresponding standard, that is, whether the integral value is ≥ 10;

[0073] If reached, it is determined that the patient's bone marrow cells corresponding to the current bone marrow cell image have abnormal development. For example, if 100 images of the granulocyte lineage are identified and the converted integral value of the number of abnormal development cells is ≥ 10, it indicates that the granulocyte cells have abnormal development.

[0074] Step 4: Substitute the output result of the trained neural network model into the relationship function to obtain the abnormal recognition result of the current bone marrow cell image.

[0075] Those of ordinary skill in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0076] In the several embodiments provided in the present application, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the division of the units described above is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The above-mentioned units may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.

[0077] 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 replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network, characterized in that: The steps include: Acquire bone marrow cell images, divide the bone marrow cell images according to cell types to obtain multiple cell sample sets, and extract nucleus images from the bone marrow cell images in the cell sample sets to obtain multiple nucleus sample sets; Constructing a neural network model, and using the cell sample set and the nucleus sample set to train the neural network model; wherein, in the process of training the neural network, the cell sample sets and the nucleus sample sets of different cell types are used to train the neural network separately; According to the ROC curve of the prediction results of the neural network model, a threshold corresponding to the maximum Youden index is obtained, and an integral value for quantifying the prediction results is set, and a relationship function between the number of dysplastic cells and the integral value is established based on the obtained threshold; Substitute the output results of the trained neural network model into the relationship function to obtain the abnormal recognition results of the current bone marrow cell image.

2. The method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to claim 1, characterized in that: The process of acquiring bone marrow cell images, dividing the bone marrow cell images according to cell types, and obtaining multiple cell sample sets includes: Scan the patient's bone marrow smear specimens with a high-throughput scanning device to obtain the corresponding bone marrow cell morphology images and screen out single cell images with clear cell structures; All bone marrow cell morphology images are divided according to the custom cell types; wherein the custom cell types are specifically: Normal granulocytes, granulocytes with somatic / cytoplasmic dysplasia and granulocytes with nuclear dysplasia, normal nucleated erythrocytes, nucleated erythrocytes with somatic / cytoplasmic dysplasia and nucleated erythrocytes with nuclear dysplasia, normal megakaryocytes, megakaryocytes with somatic dysplasia and megakaryocytes with nuclear dysplasia.

3. The method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to claim 1, characterized in that: The obtained multiple cell sample sets and multiple nucleus sample sets of different cell types are divided into a training set, a validation set, and a test set respectively; wherein the ratio of the three sets of the training set, the validation set, and the test set is 8:1:1 respectively.

4. The method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to claim 1, characterized in that: The neural network model is a VGG classification neural network, and the process of building the neural network model includes: Construct the convolution layer, maximum pooling layer, and first fully connected layer of the neural network model; wherein the convolutional neural network uses the convolution layer and the maximum pooling layer to extract the features of the bone marrow cell image, and the first fully connected layer flattens the features of the bone marrow cell image extracted by the convolution layer and the maximum pooling layer.

5. The method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to claim 1, characterized in that: Before executing the operation of using the cell sample set and the nucleus sample set to train the neural network model, first perform the following operations: The cell sample set and the nucleus sample set are encapsulated into a data structure of a specified format, the encapsulated cell and nucleus sample sets are shuffled and randomly divided into multiple batches, and the multiple batches of cell and nucleus sample sets are transmitted to the neural network model in parallel.

6. The method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to claim 1, characterized in that: After executing the operation of training the neural network model using the cell sample set and the nucleus sample set, perform the following operations: The neural network model also includes a second fully connected layer to evaluate the prediction accuracy of the trained neural network model. When the prediction accuracy is stable, the output results of the neural network model trained using the cell sample set and the output results of the neural network model trained using the nucleus sample set are input into the second fully connected layer for data analysis to obtain abnormal prediction values ​​of the cell body, cytoplasm, and nucleus.

7. The method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to claim 1, characterized in that: The process of obtaining the threshold value corresponding to the maximum Youden index according to the ROC curve of the prediction result of the neural network model includes: The prediction accuracy of the trained neural network model is statistically analyzed to obtain the true positive rate and the false positive rate, and the true positive rate and the false positive rate are used as two coordinates to draw a ROC curve diagram of the prediction results of the neural network model; wherein the ROC curve diagram is a three-line ROC curve diagram including granulocytes, nucleated red blood cells, and megakaryocytes; The maximum Youden index of the three-system prediction results was calculated respectively according to the three-system ROC curve diagram.

8. The method for identifying abnormal development of bone marrow hematopoietic cells based on a neural network according to claim 1, characterized in that: The process of setting the integral value for quantifying the prediction result and establishing the relationship function between the number of dysplastic cells and the integral value based on the obtained threshold value includes: The dysplasia integral value is obtained by calculating the product of the threshold and the number of cells with dysplasia identified by the neural network model, and judging whether the integral value of the current bone marrow cell image reaches the corresponding standard according to the WHO definition of bone marrow hematopoietic cell dysplasia; If reached, it is determined that the patient's bone marrow cells corresponding to the current bone marrow cell image have abnormal development.

Citation Information

Patent Citations

  • Automatic identification method and system for bone marrow cell image in continuous maturation stage, and medium

    CN113011306A

  • Artificial intelligence analysis method and system for bone marrow cell morphology

    CN118506361A