An ai-assisted diagnosis system for acute myeloid leukemia subtypes

Abstract

The application discloses an AI auxiliary diagnosis system for acute myeloid leukemia subtypes, comprising: a diagnosis model training module configured to train and test a detection network model using a first data set to obtain a trained detection network model; train and test a fusion model using a second data set to obtain a trained classification network model; a cell classification module configured to locate nucleated cells in images by combining multi-view images of a bone marrow smear to be diagnosed and the trained detection network model, and obtain rectangular bounding box results of all nucleated cells in each image; obtain the classification of nucleated cells according to single nucleated cells segmented automatically according to the rectangular box results and the trained fusion model; and a result diagnosis module configured to count the classified nucleated cells, calculate the proportion of each type of cell, and complete the diagnosis of acute myeloid leukemia subtypes according to the medical blood disease classification standard.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an AI-assisted diagnostic system for a subtype of acute myeloid leukemia. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Acute myeloid leukemia (AML) is a malignant disease of myeloid hematopoietic stem / progenitor cells, characterized primarily by the abnormal proliferation of primitive and immature myeloid cells in peripheral blood or bone marrow. AML is a type of leukemia with complex etiologies and poses a significant threat to human health. Epidemiological studies show that AML can occur at any age, is the most common acute leukemia in adults, and accounts for 30% of childhood leukemia; its incidence increases with age. According to the medical FAB morphological classification, AML can be divided into eight subtypes, M0-M7, each characterized by the abnormal proliferation of different cell types. Given the threat and complexity of AML, early and accurate diagnosis is crucial for its treatment.

[0004] Currently, the diagnosis of various leukemias, including AML, is based on bone marrow cell morphology examination. First, the doctor performs a bone marrow aspiration to obtain fluid. Then, the cells are stained with a staining agent, smeared onto a glass slide, and observed under an electron microscope. The technician first selects several suitable fields of view under low magnification, then adjusts to high magnification to count cells according to type and developmental stage, typically counting a total of 200 nucleated cells. Finally, the percentage of each cell type is calculated, and a diagnosis is made based on hematological classification standards. As can be seen, manual microscopic examination is time-consuming and labor-intensive, heavily relying on the technician's professional skills. Furthermore, subtle differences between cells can easily lead to misidentification, and subjective factors can cause variations in identification among different personnel.

[0005] With the development of digital imaging technology, high-resolution bone marrow microscopic images can be more easily digitally represented, promoting the application of image processing algorithms in intelligent cell morphology examination. In recent years, some studies have used traditional algorithms to segment individual cells, then extracted features such as color, texture, and geometry, which were then input into a classifier for classification. However, in actual bone marrow smears, cells exhibit aggregation and overlap, making it difficult for traditional segmentation methods to obtain single cells, introducing noise interference to traditional feature engineering, and thus resulting in inaccurate classification. With the development of deep learning, some more advanced automatic feature learning models have been applied to bone marrow cell classification, such as various convolutional networks (CNNs). Nevertheless, some problems remain in related research. First, the sample size used for network training is relatively small; second, the differences between cells at different differentiation stages are minimal, making network recognition difficult; third, CNN networks have weak generalization ability and perform poorly on unknown datasets; finally, some studies prioritize model accuracy while neglecting computational efficiency. In addition to cell classification, the quantity and proportion of cells must also be considered in the diagnosis of hematological diseases. In studies that also involve cell classification and counting, single-stage YOLO detection networks are mostly used. However, as the number of complex cell types increases, the annotation workload increases, and using only a single network can lead to missed detections and false positives. Due to these problems, many related studies have focused only on AML identification, while diagnosing more specific and diverse AML subtypes remains challenging.

[0006] In summary, existing AML diagnostic methods are mainly based on bone marrow cell morphology examination, which has the following shortcomings: 1) Manual microscopic examination is time-consuming and labor-intensive, tests the professional level of the examiner, is highly subjective, and is prone to identification errors; 2) Diagnostic methods based on traditional image algorithms are difficult to obtain single-cell images for cell clusters in bone marrow smears, and manual feature extraction is cumbersome and the classification is not accurate enough; 3) Diagnostic methods based on deep learning have problems such as insufficient training data and class imbalance, weak network generalization ability, difficulty in detecting multiple cell types with a single network, and lack of specific diagnosis of multiple subtypes. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes an AI-assisted diagnostic system for acute myeloid leukemia (AML) subtypes. First, a lightweight target detection network is used to detect nucleated cells in bone marrow microscopic images. Then, the individual cells obtained are input into a lightweight fusion model for specific type identification. Finally, by combining multiple fields of view under high magnification, the number of each type of cell detected is statistically analyzed, the proportions are calculated, and the diagnosis of the AML subtype is completed according to disease classification standards.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides an AI-assisted diagnostic system for a subtype of acute myeloid leukemia, comprising:

[0010] The dataset acquisition module is configured to: acquire bone marrow microscopic images and bone marrow single-cell images; annotate the bone marrow microscopic images to form a first dataset, and annotate the bone marrow single-cell images to form a second dataset;

[0011] The diagnostic model training module is configured to: train and test the detection network model using the first dataset to obtain the trained detection network model; and train and test the fusion model using the second dataset to obtain the trained classification network model.

[0012] The cell classification module is configured to: combine multi-view images of the bone marrow smear to be diagnosed with a trained detection network model to locate nucleated cells in the images and obtain rectangular bounding boxes of all nucleated cells in each image; and automatically segment individual nucleated cells based on the detected rectangular bounding boxes and the trained fusion model to obtain the classification results of nucleated cells.

[0013] The diagnostic module is configured to: count the nucleated cells after classification, calculate the proportion of each cell type, and complete the diagnosis of the subtype of acute myeloid leukemia according to the medical classification standards for hematological diseases.

[0014] In a second aspect, the present invention provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, complete an AI-assisted diagnostic method for a subtype of acute myeloid leukemia, the method comprising:

[0015] Obtain bone marrow microscopic images and bone marrow single-cell images; annotate the bone marrow microscopic images to form the first dataset, and annotate the bone marrow single-cell images to form the second dataset;

[0016] The detection network model was trained and tested using the first dataset to obtain the trained detection network model; the fusion model was trained and tested using the second dataset to obtain the trained classification network model.

[0017] By combining multi-view images of the bone marrow smear to be diagnosed with a trained detection network model, nucleated cells in the images are located, and rectangular bounding boxes of all nucleated cells in each image are obtained. Based on the detected rectangular bounding boxes, individual nucleated cells are automatically segmented and combined with the trained fusion model to obtain the classification results of nucleated cells.

[0018] After classifying the nucleated cells, the proportion of each cell type is calculated, and the diagnosis of the subtype of acute myeloid leukemia is completed according to the medical classification standards for hematological diseases.

[0019] Thirdly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete an AI-assisted diagnostic method for a subtype of acute myeloid leukemia, the method comprising:

[0020] Obtain bone marrow microscopic images and bone marrow single-cell images; annotate the bone marrow microscopic images to form the first dataset, and annotate the bone marrow single-cell images to form the second dataset;

[0021] The detection network model was trained and tested using the first dataset to obtain the trained detection network model; the fusion model was trained and tested using the second dataset to obtain the trained classification network model.

[0022] By combining multi-view images of the bone marrow smear to be diagnosed with a trained detection network model, nucleated cells in the images are located, and rectangular bounding boxes of all nucleated cells in each image are obtained. Based on the detected rectangular bounding boxes, individual nucleated cells are automatically segmented and combined with the trained fusion model to obtain the classification results of nucleated cells.

[0023] After classifying the nucleated cells, the proportion of each cell type is calculated, and the diagnosis of the subtype of acute myeloid leukemia is completed according to the medical classification standards for hematological diseases.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] (1) This invention uses a lightweight target detection network to detect nucleated cells in bone marrow microscopic images; the obtained individual cells are input into a lightweight fusion model for specific type identification; by combining multiple fields of view under high magnification of the patient, the number of each type of cell is counted, the proportion is calculated, and the diagnosis of AML subtype is completed according to the disease classification criteria. The cell classification and counting method using a multi-stage, multi-model fusion method can improve the detection accuracy compared to a single network, and is more suitable for situations where the differences between multiple cell types are small; it efficiently realizes the identification and diagnosis of specific AML subtypes, rather than just stopping at the level of identifying the broad category of AML.

[0026] (2) This invention uses a lightweight CNN-ViT fusion model for cell classification, which fully integrates the structural characteristics of CNN and ViT respectively, and can achieve good performance with fewer parameters and less computation.

[0027] (3) The method proposed in this invention is closer to the conventional testing and diagnostic process, providing a certain reference for the diagnosis of more hematological diseases and their subtypes in the future.

[0028] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1 This is an overall flowchart of AML subtype diagnosis provided in an embodiment of the present invention;

[0031] Figure 2 This is an architecture diagram of the YOLOv7 network provided in an embodiment of the present invention;

[0032] Figure 3 This is an architecture diagram of the CNN-ViT fusion model provided in an embodiment of the present invention. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0037] Example 1

[0038] like Figure 1 As shown, from the perspective of bone marrow cell morphology examination and based on a multi-stage, multi-model fusion algorithm in deep learning, this embodiment provides an AI-assisted diagnostic system for a subtype of acute myeloid leukemia, including:

[0039] The data acquisition module is configured to: acquire bone marrow microscopic images and bone marrow single-cell images; annotate the bone marrow microscopic images to form a first dataset D1, and annotate the bone marrow single-cell images to form a second dataset D1;

[0040] In this embodiment, the acquisition process of the first dataset D1 is as follows: 208 bone marrow microscopic images were collected from various channels such as hospitals and the Internet, covering various common blood diseases. Each image contains 10-100 nucleated cells, and the Labelme tool was used to label all nucleated cells to obtain their label information: category (number of categories is 1) and rectangular bounding box (center point coordinates, width and height), thus forming the first dataset D1.

[0041] The second dataset D1 was obtained as follows: one part was segmented from bone marrow microscopic images, and the other part was obtained from a public database.

[0042] The images, totaling 6011, were categorized by professional physicians into 11 classes: primitive / early erythroblasts (PEN), intermediate / late erythroblasts (PON), primitive granulocytes (MYB), early / intermediate / late myelocytes (PMM), aberrant promyelocytes (ABP), band neutrophils (NGB), segmented neutrophils (NGS), eosinophils (EOS), basophils (BAS), lymphocytes (LYT), and monocytes (MON). These images form the second dataset, D2.

[0043] The data preprocessing module is configured to partition the first dataset D1 and the second dataset and perform data augmentation, respectively.

[0044] In this embodiment, the partitioning and enhancement process of the first dataset D1 is as follows:

[0045] The dataset D1 was divided into a training set and a test set in an 8:2 ratio. Online data augmentation techniques were used on the training set, namely, random color space transformation, random translation, random horizontal flipping, and Mosaic enhancements were applied in batches during the training iterations.

[0046] The partitioning and augmentation process of the second dataset D2 is as follows:

[0047] First, due to the class imbalance problem in the original dataset D2, this embodiment uses offline data augmentation to expand the dataset. Various augmented images are saved to folders using methods including random horizontal / vertical flipping, random 90-degree rotation, random Gaussian blur, and random color space transformation. Each cell category's images are augmented to over 2000, resulting in a total of 22525 images.

[0048] Next, the dataset D2 was divided into training, validation, and test sets in a 7:1:2 ratio. Online data augmentation techniques were used during training, including random scaling and cropping, random horizontal / vertical flipping, random color space transformation, and random reduction of color channel bits.

[0049] The nucleated cell detection module is configured to: train and test the detection network model using a first dataset to obtain a trained detection network model; and train and test the fusion network model using a second dataset to obtain a trained classification network model.

[0050] The specific steps of training and testing the detection network model using the first dataset include:

[0051] In this embodiment, the detection network model adopts the YOLOv7-tiny network, and its overall network structure is as follows: Figure 2 As shown.

[0052] Before training and testing the detection network model, the images in the first dataset are first scaled up, for example, to 640*640*3.

[0053] The process of training and testing the detection network model using the first dataset is as follows:

[0054] Feature extraction is performed through modules such as CBS and ELAN in the backbone, followed by feature fusion through the neck part. Finally, the three heads output the predicted bounding box information (x, y, w, h), foreground confidence score obj, and class probability cls_score at different scales, in the form of h′×w′×3×(x, y, w, h, obj, cls_score), where h′×w′ represents the feature map size and 3 represents the number of initial anchor boxes generated for each feature map pixel.

[0055] The optimized loss functions used in this training process include: bounding box loss, confidence loss, and classification loss;

[0056] The bounding box loss uses GIOU loss(L) giou The calculation is as follows:

[0057]

[0058] In the formula, B pred B represents the predicted area. true B represents the actual bounding box area, while B c This represents the area of ​​the frame corresponding to the longest diagonal of the two frames.

[0059] Both confidence loss and classification loss use BCE loss(Lbce The calculation is as follows:

[0060]

[0061] In the formula, y i p represents the true label of the sample, which is 0 or 1. i This represents the predicted probability (between 0 and 1).

[0062] During training, the network was pre-trained using the COCO dataset with a batch size of 64. The SGD optimizer was used with an initial learning rate of 0.01 and an L2 decay coefficient of 0.0005. The training run consisted of 500 epochs, employing a warm-up (3 epochs) + cosine annealing learning strategy. For testing, Non-Maximum Suppression (DIOUNMS) was used for bounding box selection, with a confidence threshold of 0.35 and an NMS threshold of 0.45. The final test accuracy reached 99%.

[0063] The training and testing of the fusion network model using the second dataset specifically includes:

[0064] In this embodiment, the fusion network model adopts a lightweight CNN-ViT fusion model;

[0065] Because this embodiment involves a large number of bone marrow cell types with minimal differences between some cells, using a single CNN network for classification is ineffective. Therefore, the ViT model, which can process global information and has strong generalization capabilities, is introduced to construct a lightweight CNN-ViT fusion model. The CNN uses the MobileNetV3-large network, while the ViT uses the ViT-r-s16-p8 network. The individual CNN and ViT networks are trained separately on dataset D2. The class probabilities output by both are then fused using two methods: maximum vote and average vote. The resulting fusion model has only 10.3M parameters and 0.67G of FLOPs (computational cost), which does not incur significant computational overhead.

[0066] The architecture of the CNN-ViT fusion model is as follows: Figure 3 As shown, the CNN branch mainly consists of stacked depth-wise (3*3) convolutions, point-wise (1*1) convolutions, and SE modules; the ViT branch mainly consists of patch embedding and multi-head self-attention. Both ultimately output class probabilities through fully connected layers and a sigmoid function.

[0067] Both CNN and ViT employ the following common training strategies: loading pre-trained parameters from the ImageNet dataset, converting class labels to one-hot encoding, using BCE loss as the loss function, employing an input image resolution of 224*224, using the SGD optimizer, setting the first batch size when training the CNN, and setting the second batch size when training the ViT network, with different iteration counts for training the CNN and ViT networks. During testing, if the predicted probability is lower than a set threshold, the cell is classified as another category.

[0068] Specifically, the initial learning rate was 0.01, the L2 decay coefficient was 0.0005, and a cosine annealing strategy was used. The difference was that the batch size for training the CNN was set to 64, with 100 epochs, while the batch size for ViT was set to 128, with 80 epochs. During testing, if the predicted probability was lower than 0.3, the cell was classified as "other". Table 1 shows the performance of the single model and the two fusion methods on the test set.

[0069] Table 1. Comparison test results of models for nucleated cell classification

[0070]

[0071] The evaluation metrics used include accuracy, precision, recall, F1 score, and test time, calculated using the following formulas:

[0072]

[0073]

[0074]

[0075]

[0076] In the formula, TP, FP, TN, and FN represent the correctly predicted positive example, the incorrectly predicted negative example, the correctly predicted negative example, and the incorrectly predicted positive example, respectively.

[0077] The cell classification module is configured to: combine multi-view images of bone marrow smears from patients to be diagnosed with a trained detection network model to detect and locate nucleated cells in the images, and obtain rectangular bounding boxes of all nucleated cells in each image; automatically segment individual nucleated cells based on the detected rectangular bounding boxes, input them into the trained fusion model, and select the average voting method with relatively good fusion results to obtain the classification results of nucleated cells;

[0078] The process of acquiring multi-field images of bone marrow smears from patients to be diagnosed is as follows: Under a high-power microscope (70-100x is acceptable), select multiple fields of view in the bone marrow smear (ensuring the total number of cells is greater than 200) and use them as input images for the diagnostic model. The selected fields of view are areas with uniform cell distribution, mostly located at the body-tail junction of the medullary membrane.

[0079] The diagnostic module is configured to: count nucleated cells after detection and classification, calculate the proportion of each cell type, and complete the diagnosis of AML subtypes according to the medical classification standards for hematological diseases.

[0080] The procedure involves examining over 200 nucleated cells across multiple fields of view, counting each cell type, and calculating their percentage within the total cell count. Then, based on medical classification standards for hematological disorders, specifically the relevant cell proportion range, the specific subtype of AML in the patient is diagnosed.

[0081] The diagnostic criteria used in this embodiment are based on the 2016 WHO Classification of Tumors of Hematopoietic and Lymphoid Tissues.

[0082] This invention obtained digital panoramic bone marrow smears from 10 AML patients for testing, including M1-M5 subtypes, with an accuracy rate of 100%.

[0083] In further embodiments, the following is also provided:

[0084] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0085] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0086] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0087] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0088] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0089] Those skilled in the art will recognize that the units, i.e., algorithm steps, of the various examples described in connection with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0090] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An AI-assisted diagnostic system for a subtype of acute myeloid leukemia, characterized in that, include: The dataset acquisition module is configured to acquire bone marrow microscopic images and bone marrow single-cell images. The first dataset was formed by annotating bone marrow microscopic images, and the second dataset was formed by annotating bone marrow single-cell images. The diagnostic model training module is configured to: train and test the detection network model using the first dataset to obtain the trained detection network model; and train and test the fusion model using the second dataset to obtain the trained classification network model. The fusion model is constructed by using a parallel dual-branch lightweight CNN-ViT fusion model, where the CNN branch uses the MobileNetV3-large network and the ViT branch uses the ViT-r-s16-p8 network. The CNN branch and the ViT branch are trained separately on the second dataset. Then, the first class probability output by the CNN branch and the second class probability output by the ViT branch are fused by maximum voting or average voting to obtain the final cell class probability. The cell classification module is configured to: combine multi-view images of the bone marrow smear to be diagnosed with a trained detection network model to locate nucleated cells in the images and obtain rectangular bounding boxes of all nucleated cells in each image; and automatically segment individual nucleated cells based on the detected rectangular bounding boxes and the trained fusion model to obtain the classification results of nucleated cells. The diagnostic module is configured to: count the nucleated cells after classification, calculate the proportion of each cell type, and complete the diagnosis of the subtype of acute myeloid leukemia according to the medical classification standards for hematological diseases.

2. The AI-assisted diagnostic system for a subtype of acute myeloid leukemia as described in claim 1, characterized in that, The system also includes a data preprocessing module, which is configured to partition the first dataset and the second dataset and perform data augmentation, respectively.

3. The AI-assisted diagnostic system for a subtype of acute myeloid leukemia as described in claim 1, characterized in that, The construction process of the detection network model is as follows: Feature extraction is performed using the CBS and ELAN modules in the backbone, followed by feature fusion through the neck section. Finally, the predicted bounding box information, foreground confidence score, and class probability at different scales are output through three heads.

4. The AI-assisted diagnostic system for a subtype of acute myeloid leukemia as described in claim 1, characterized in that, The optimized loss functions used during the training of the detection network model include: bounding box loss, confidence loss, and classification loss.

5. The AI-assisted diagnostic system for a subtype of acute myeloid leukemia as described in claim 1, characterized in that, The process of training and testing the individual CNN and ViT networks on the second dataset is as follows: During training, pre-trained parameters from the second dataset are loaded, class labels are converted into one-hot encoded forms, BCE loss is used as the loss function, and SGD optimizer is used. The first batch size is set when training the CNN, and the second batch size is set when training the ViT network. The number of iterations when training the CNN is different from the number of iterations when training the ViT network. During testing, if the predicted probability is lower than the set threshold, the cell is identified as another class.

6. The AI-assisted diagnostic system for a subtype of acute myeloid leukemia as described in claim 1, characterized in that, The medical classification standard for hematological diseases adopts the "2016 WHO Classification of Tumors of Hematopoietic and Lymphoid Tissues".

7. An electronic device, characterized in that, The system includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform an AI-assisted diagnostic method for a subtype of acute myeloid leukemia, the method comprising: Obtain bone marrow microscopic images and bone marrow single-cell images; annotate the bone marrow microscopic images to form the first dataset, and annotate the bone marrow single-cell images to form the second dataset; The detection network model was trained and tested using the first dataset to obtain the trained detection network model; the fusion model was trained and tested using the second dataset to obtain the trained classification network model. By combining multi-view images of bone marrow smears to be diagnosed with a trained detection network model, nucleated cells are located in the images, resulting in rectangular bounding boxes of all nucleated cells in each image. Individual nucleated cells are automatically segmented based on the detected bounding boxes, and the trained fusion model is used to obtain the classification results of the nucleated cells. The fusion model is constructed using a parallel dual-branch lightweight CNN-ViT fusion model, where the CNN branch uses a MobileNetV3-large network and the ViT branch uses a ViT-r-s16-p8 network. The CNN branch and the ViT branch are trained separately on a second dataset. Then, the first class probability output by the CNN branch and the second class probability output by the ViT branch are fused using either maximum voting or average voting to obtain the final cell class probability. After classifying the nucleated cells, the proportion of each cell type is calculated, and the diagnosis of the subtype of acute myeloid leukemia is completed according to the medical classification standards for hematological diseases.

8. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete an AI-assisted diagnostic method for a subtype of acute myeloid leukemia, the method comprising: Obtain bone marrow microscopic images and bone marrow single-cell images; annotate the bone marrow microscopic images to form the first dataset, and annotate the bone marrow single-cell images to form the second dataset; The detection network model was trained and tested using the first dataset to obtain the trained detection network model; the fusion model was trained and tested using the second dataset to obtain the trained classification network model. By combining multi-view images of bone marrow smears to be diagnosed with a trained detection network model, nucleated cells are located in the images, resulting in rectangular bounding boxes of all nucleated cells in each image. Individual nucleated cells are automatically segmented based on the detected bounding boxes, and the trained fusion model is used to obtain the classification results of the nucleated cells. The fusion model is constructed using a parallel dual-branch lightweight CNN-ViT fusion model, where the CNN branch uses a MobileNetV3-large network and the ViT branch uses a ViT-r-s16-p8 network. The CNN branch and the ViT branch are trained separately on a second dataset. Then, the first class probability output by the CNN branch and the second class probability output by the ViT branch are fused using either maximum voting or average voting to obtain the final cell class probability. After classifying the nucleated cells, the proportion of each cell type is calculated, and the diagnosis of the subtype of acute myeloid leukemia is completed according to the medical classification standards for hematological diseases.

Images