Bone marrow blood cell target identification and counting method based on improved single stage

Abstract

The invention specifically discloses a bone marrow blood cell target identification and counting method based on an improved single stage, and relates to the technical field of medical image processing. The method comprises the following steps: S1, image acquisition and data preparation; s2, constructing and preprocessing a data set, and dividing the preprocessed data set into a training set, a verification set and a test set; s3, constructing an initial model of an improved single-stage target recognition model; s4, training the initial model by using a training set and a verification set to obtain a bone marrow blood cell target recognition and counting model; and S5, identifying and positioning the bone marrow blood cells in the test set image by adopting a bone marrow blood cell target identification and counting model, outputting cell categories and target frame coordinates, and realizing counting according to the number of target frames. According to the method, the detection problem caused by large bone marrow cell scale change is solved, the classification precision is improved, and unification of high precision and high speed of bone marrow blood cell target identification and counting is realized.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an improved single-stage method for identifying and counting bone marrow blood cell targets. Background Technology

[0002] Bone marrow cell morphology examination is the gold standard for diagnosing hematologic disorders such as leukemia and anemia. In the clinical diagnostic process, doctors need to use a microscope to observe bone marrow smears, classify and count hundreds or even thousands of cells, and provide core evidence for disease diagnosis and condition assessment by accurately quantifying the proportion of each type of cell.

[0003] However, the current clinical bone marrow cell detection technology still has many shortcomings, which seriously restrict the efficiency and consistency of diagnosis. The specific manifestations are as follows: Manual microscopic examination is inefficient and highly subjective: Traditional bone marrow cell classification and counting relies on manual microscopic examination by doctors, which is not only time-consuming and labor-intensive, but also highly dependent on the operator's clinical experience and subjective judgment. Different doctors, or even the same doctor at different times, may have different judgments, making it difficult to guarantee the standardization and consistency of diagnosis.

[0004] Bone marrow cells have complex morphologies, making identification difficult: cells are densely distributed in bone marrow smears, and there are often overlapping and occlusion phenomena; at the same time, granulocytes at different developmental stages (such as neutrophils and metamyelocytes) have extremely high morphological similarity, and subtle structural differences are difficult to distinguish. These characteristics bring huge challenges to the automated identification of cells.

[0005] Limitations of the algorithm: Current mainstream cell detection algorithms struggle to balance accuracy and efficiency. Traditional two-stage target detection algorithms (such as Faster R-CNN) offer some detection accuracy, but their slow inference speed fails to meet the demands of real-time clinical diagnosis. Conventional single-stage algorithms (such as the original YOLO series) suffer from insufficient feature extraction capabilities for small and dense targets, leading to missed detections and false positives, making it difficult to achieve the accuracy requirements of clinical diagnosis.

[0006] In conclusion, clinical practice urgently needs an automated method for identifying and counting bone marrow cells that combines high reasoning efficiency with high detection accuracy, in order to overcome existing technological bottlenecks and promote the development of hematological disease diagnosis towards automation, standardization, and precision. Summary of the Invention

[0007] The purpose of this invention is to propose an improved single-stage bone marrow hematopoietic cell target identification and counting method. This method combines advanced data augmentation technology, multi-scale feature fusion mechanism and hybrid attention mechanism to achieve efficient and accurate analysis of bone marrow cells.

[0008] To achieve the above objectives, this invention proposes an improved single-stage bone marrow hematopoietic cell target identification and counting method, the specific steps of which are as follows: Step S1, Image Acquisition and Data Preparation: Using a high-resolution electron microscope at 100× magnification, and with the aid of a bone marrow cell morphology examination system, acquire full-area image data of bone marrow granulocyte smears processed by the standard Swiss staining method. Step S2: Construct and preprocess the dataset, and divide the preprocessed dataset into training set, validation set and test set; Step S3: Construct the initial model of the improved single-stage target recognition model. The initial model includes a core network structure with a multi-scale feature fusion mechanism BiFPN and a hybrid attention mechanism module CBAM. Step S4: Train the initial model using the training set and validation set to obtain the bone marrow blood cell target recognition and counting model; Step S5: Use the bone marrow blood cell target recognition and counting model to identify and locate bone marrow blood cells in the test set images, output cell type and target box coordinates, and count according to the number of target boxes.

[0009] Preferably, in step S2, the specific steps for constructing the dataset are as follows: Step S21: Select images with more than 200 cell units in the field of view as valid samples, and retain cell samples distributed at the edge of the image; Step S22: Using a three-blood expert cross-validation and voting counting mechanism, blood cells in the valid samples are labeled with categories to generate the initial dataset; Step S23: Preprocess the initial dataset, including image rotation, cropping, scaling, flipping, blending, padding with 0 pixels, and stitching, to expand the initial dataset.

[0010] Preferably, in step S23, new synthetic image samples are formed through data augmentation using Mosaic. The specific steps are as follows: Step S231: Randomly select four different images from the dataset and randomly stitch them together; Step S232: Randomly select one of the target boxes and adjust the target box from its position in the original image to its new coordinates in the newly synthesized image; Step S233: Scale and crop the newly synthesized image to ensure that the target box is fully visible in the new image and adjusted to the correct size; Step S234: Use the new synthetic image as a new sample, and the adjusted target box position and size as labels to update the dataset.

[0011] Preferably, in step S3, the Neck part of the improved single-stage target recognition model network introduces a multi-scale feature fusion mechanism, BiFPN. The BiFPN mechanism achieves feature fusion through bidirectional and cross-layer connections, as detailed below: Introduce learnable weights at each fusion node. Numerical stabilization is achieved through fast normalization, and the calculation formula is as follows: in, This is the output feature map of the fusion node. For the input feature map, To prevent the minimum value where the denominator is zero, i , j For summation index.

[0012] Preferably, the CBAM module adopts a sequential application mechanism, including a channel attention module (CAM) and a spatial attention module (SAM): Channel Attention Module (CAM): Focuses on "what" is meaningful in feature channels; Spatial Attention Module (SAM): Focuses on "where" is meaningful in the feature space; The CBAM module is embedded after multiple convolutional blocks in the core network structure to give the network the ability to automatically select important features in the channel and spatial dimensions.

[0013] Preferably, in step S4, a combined loss function is used to guide model optimization during the initial model training process. The combined loss function is composed of a weighted sum of classification loss and localization loss. The classification loss uses binary cross-entropy loss to determine whether each target box contains blood cells and which category it belongs to. The localization loss uses GIOU localization loss to measure the difference between the predicted target box and the true target box, thereby improving the regression accuracy of the target box. The formula for calculating the combined loss function is as follows: ; in, For combined loss, For binary cross-entropy classification loss, Locate the loss for GIOU. These are the weighting coefficients.

[0014] Preferably, the initial model training process includes a frozen training phase and a unfrozen training phase: During the frozen training phase, the weight parameters of the backbone network are frozen, and only the parameters of the neck network and the detection head are updated. During the unfreeze training phase, the backbone network is unfrozen, and all parameters of the model are globally fine-tuned.

[0015] Therefore, this invention proposes an improved single-stage bone marrow hematopoietic cell target identification and counting method, which has the following beneficial effects: (1) The present invention ensures the accuracy of the annotation through expert cross-validation, and at the same time, through Mosaic data enhancement, it greatly enriches the background of the detected object and significantly improves the model's ability to identify small targets and occluded targets.

[0016] (2) By introducing the BiFPN mechanism, this invention achieves the effective fusion of deep semantic information and shallow detail information through a weighted bidirectional feature pyramid, thus solving the detection problem caused by the large scale variation of bone marrow cells.

[0017] (3) The present invention enables the model to focus on key distinguishing features such as cell nuclear texture and cytoplasmic granules through the CBAM attention mechanism, suppressing background noise and improving classification accuracy.

[0018] (4) This invention accelerates the model convergence speed and improves the regression accuracy of the prediction box by combining the training strategy of freezing and thawing and applying the GIOU loss function, and finally achieves the unity of high accuracy and high speed.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart of an improved single-stage bone marrow hematopoietic cell target identification and counting method according to the present invention; Figure 2 This is a network architecture diagram of the improved single-stage target recognition model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the BiFPN feature fusion node structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the CBAM hybrid attention mechanism module structure in an embodiment of the present invention. Detailed Implementation

[0021] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] Example like Figure 1 As shown, this invention provides an improved single-stage bone marrow blood cell target identification and counting method, the specific steps of which are as follows: Step S1: Image acquisition and data preparation.

[0024] Due to the minute morphology of bone marrow cells, extremely high image clarity is required. This embodiment utilizes an Olympus BX43 optical microscope equipped with a high-resolution CCD camera. The acquisition process was performed at 100× oil immersion magnification to ensure clear visibility of the cell nuclear chromatin structure and cytoplasmic granules. Simultaneously, bone marrow smears treated with standard Swiss staining were acquired using the BEION V4.90 bone marrow cell morphology examination system. To ensure data comprehensiveness, full-area image data was acquired, covering regions with different degrees of bone marrow hyperplasia.

[0025] Step S2: Dataset construction and preprocessing, dividing the preprocessed dataset into training, validation, and test sets. The dataset construction steps are as follows: Step S21, Sample Screening: Remove images with severely ruptured cells or extremely poor staining, and select images containing at least 200 cell units within the field of view; pay special attention to preserving cells at the image edges to train the model's ability to handle truncated targets.

[0026] Step S22, Ground Truth: Cross-validation was conducted by three senior hematology experts. First, two experts independently annotated the data. For inconsistent results, a third expert arbitrated the data. Finally, a voting mechanism was used to determine the cell type (e.g., blast cells, promyelocytes, neutrophils, etc.) and location.

[0027] Step S23: Preprocess the initial dataset, including image rotation, cropping, scaling, flipping, blending, padding with 0 pixels, and stitching, to expand the initial dataset.

[0028] The specific steps for creating new synthetic image samples using data augmentation Mosaic are as follows: Step S231: Randomly select four different images from the dataset and randomly stitch them together; Step S232: Randomly select one of the target boxes and adjust the target box from its position in the original image to its new coordinates in the newly synthesized image; Step S233: Scale and crop the newly synthesized image to ensure that the target box is fully visible in the new image and adjusted to the correct size; Step S234: Use the new synthetic image as a new sample, and the adjusted target box position and size as labels to update the dataset.

[0029] The updated dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

[0030] Step S3: Construct the initial model of the improved single-stage target recognition model.

[0031] This embodiment builds a model based on a single-stage detection architecture (such as an improved version of the YOLO series architecture). Figure 2 As shown, the improvements mainly lie in the Neck and Attention sections.

[0032] Introducing BiFPN for multi-scale feature fusion: Traditional FPN (Feature Pyramid Network) typically only has a top-down unidirectional information flow, while PANet, although adding a bottom-up path, has a high computational cost; for example Figure 3 As shown, this invention employs multi-scale feature fusion BiFPN, specifically as follows: Structural optimization: Nodes with only one input edge were removed, and a path was added that jumps directly from the input node to the output node, achieving a higher level of feature reuse.

[0033] Weighted fusion: Considering that feature maps of different resolutions contribute differently to the output, this invention introduces learnable weights at each fusion node. Numerical stabilization is achieved through fast normalization, and the calculation formula is as follows: in, This is the output feature map of the fusion node. For the input feature map, To prevent the minimum value where the denominator is zero, i , j For the summation index, , .

[0034] Embedded Hybrid Attention Mechanism CBAM: like Figure 4 As shown, a CBAM module is inserted after the output feature layer of the backbone network; the CBAM module adopts a sequential application mechanism, including a channel attention module (CAM) and a spatial attention module (SAM): Channel Attention (CAM): Compresses spatial dimensions through global average pooling and max pooling, and learns channel weights through MLP to focus on "what" is meaningful in feature channels.

[0035] Spatial Attention (SAM): Based on the output of CAM, pooling is performed in the channel dimension, and a spatial attention map is generated through convolutional layers to focus on "where" is meaningful in the feature space.

[0036] Step S4: Train the initial model using the training set and validation set to obtain the bone marrow blood cell target recognition and counting model; During the initial model training process, a combined loss function is used to guide model optimization. The combined loss function consists of a weighted sum of classification loss and localization loss. The classification loss uses binary cross-entropy loss to determine whether each target box contains blood cells and which category it belongs to. The localization loss uses GIOU localization loss to measure the difference between the predicted target box and the true target box, thereby improving the regression accuracy of the target box. The formula for calculating the combined loss function is as follows: ; in, For combined loss, For binary cross-entropy classification loss, Locate the loss for GIOU. These are the weighting coefficients.

[0037] The initial model training process includes a frozen training phase and a unfrozen training phase: During the frozen training phase (first 50 epochs), the weight parameters of the backbone network are frozen, and only the parameters of the neck network and the detection head are updated; During the unfreeze training phase (last 100 epochs), the backbone network is unfrozen, and all parameters of the model are globally fine-tuned using a small learning rate.

[0038] Step S5: Use the bone marrow blood cell target recognition and counting model to identify and locate bone marrow blood cells in the test set images, output cell type and target box coordinates, and count according to the number of target boxes.

[0039] Experimental results verify: Tested on a dataset containing 1501 bone marrow smear images, the improved method of this invention achieved significant advantages: the mAP (mean accuracy) reached 94.5%, an improvement of 6.2% compared to the original single-stage model's 88.3%; at the same time, the inference time per image was only 0.04 seconds, comparable to the original model, which significantly improved accuracy while ensuring inference speed, fully meeting the real-time requirements of clinical auxiliary diagnosis.

[0040] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0041] Therefore, this invention provides an improved single-stage bone marrow blood cell target identification and counting method. By introducing the BiFPN mechanism to fuse deep and shallow features, the problem of cell-scale variation is solved. The CBAM attention mechanism is used to focus on key identification features to improve classification accuracy. At the same time, a training strategy combining freezing and thawing and the GIOU loss function are adopted to accelerate model convergence and improve the regression accuracy of the prediction box. Finally, high accuracy and high speed are achieved in bone marrow blood cell identification and counting.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying and counting bone marrow hematopoietic cells based on an improved single-stage approach, characterized in that, The specific steps are as follows: Step S1, Image Acquisition and Data Preparation: Using a high-resolution electron microscope at 100× magnification, and with the aid of a bone marrow cell morphology examination system, acquire full-area image data of bone marrow granulocyte smears processed by the standard Swiss staining method. Step S2: Construct and preprocess the dataset, and divide the preprocessed dataset into training set, validation set and test set; Step S3: Construct the initial model of the improved single-stage target recognition model. The initial model includes a core network structure with a multi-scale feature fusion mechanism BiFPN and a hybrid attention mechanism module CBAM. Step S4: Train the initial model using the training set and validation set to obtain the bone marrow blood cell target recognition and counting model; Step S5: Use the bone marrow blood cell target recognition and counting model to identify and locate bone marrow blood cells in the test set images, output cell type and target box coordinates, and count according to the number of target boxes.

2. The improved single-stage bone marrow hematopoietic cell target identification and counting method according to claim 1, characterized in that, In step S2, the specific steps for constructing the dataset are as follows: Step S21: Select images with more than 200 cell units in the field of view as valid samples, and retain cell samples distributed at the edge of the image; Step S22: Using a three-blood expert cross-validation and voting counting mechanism, blood cells in the valid samples are labeled with categories to generate the initial dataset; Step S23: Preprocess the initial dataset, including image rotation, cropping, scaling, flipping, blending, padding with 0 pixels, and stitching, to expand the initial dataset.

3. The improved single-stage bone marrow hematopoietic cell target identification and counting method according to claim 2, characterized in that, In step S23, new synthetic image samples are generated through data augmentation using Mosaic. The specific steps are as follows: Step S231: Randomly select four different images from the dataset and randomly stitch them together; Step S232: Randomly select one of the target boxes and adjust the target box from its position in the original image to its new coordinates in the newly synthesized image; Step S233: Scale and crop the newly synthesized image to ensure that the target box is fully visible in the new image and adjusted to the correct size; Step S234: Use the new synthetic image as a new sample, and the adjusted target box position and size as labels to update the dataset.

4. The improved single-stage bone marrow hematopoietic cell target identification and counting method according to claim 3, characterized in that, In step S3, the improved single-stage target recognition model network introduces a multi-scale feature fusion mechanism, BiFPN, into its neck portion. The BiFPN mechanism achieves feature fusion through bidirectional and cross-layer connections, as detailed below: Introduce learnable weights at each fusion node. Numerical stabilization is achieved through fast normalization, and the calculation formula is as follows: in, This is the output feature map of the fusion node. For the input feature map, To prevent the minimum value where the denominator is zero, i , j For summation index.

5. The improved single-stage bone marrow hematopoietic cell target identification and counting method according to claim 4, characterized in that, The CBAM module adopts a sequential application mechanism, including the channel attention module CAM and the spatial attention module SAM.

6. The improved single-stage bone marrow hematopoietic cell target identification and counting method according to claim 5, characterized in that, In step S4, during the initial model training process, a combined loss function is used to guide model optimization. The combined loss function consists of a weighted sum of classification loss and localization loss. The classification loss uses binary cross-entropy loss to determine whether each target box contains blood cells and which category it belongs to. The localization loss uses GIOU localization loss to measure the difference between the predicted target box and the true target box, thereby improving the regression accuracy of the target box. The formula for calculating the combined loss function is as follows: ； in, For combined loss, For binary cross-entropy classification loss, Locate the loss for GIOU. These are the weighting coefficients.

7. The improved single-stage bone marrow hematopoietic cell target identification and counting method according to claim 6, characterized in that, The initial model training process includes a frozen training phase and a unfrozen training phase: During the frozen training phase, the weight parameters of the backbone network are frozen, and only the parameters of the neck network and the detection head are updated. During the unfreezing training phase, the backbone network is unfrozen, and all parameters of the model are globally fine-tuned.

Images