Leukocyte detection method based on FBDM-YOLOv8s

Abstract

The invention aims to provide a white blood cell detection method based on FBDM-YOLOv8s, which comprises the following steps: A, constructing an improved YOLOv8s neural network model, and the improved YOLOv8s neural network model comprises a Backbond network, a Neck network and a Head network; the Backbond network takes a plurality of groups of Fasnet modules as a main framework, and finally, a group of BRA modules is arranged; b, the improved YOLOv8s neural network model is trained, and a trained improved YOLOv8s neural network model is obtained; and C, introducing a to-be-detected leukocyte image into the trained improved YOLOv8s neural network model, and sequentially processing the to-be-detected leukocyte image through a Backbond network, a Neck network and a Head network to obtain a detected image. According to the method, the edge detection precision can be improved while model parameters are optimized, the calculation speed is increased, and the problems that image cells are incomplete and two cells are mutually overlapped are better solved.

Description

Technical Field

[0001] The present invention relates to the fields of computer technology and medical image analysis technology, and in particular to a leukocyte detection method based on FBDM-YOLOv8s. Background Art

[0002] Cells are also commonly referred to as immune cells. As part of the immune system, they help the body fight infectious diseases and foreign pathogens. In the study of blood testing, accurate identification of white blood cells can provide important information for blood analysis. Using computer-assisted technology to effectively detect white blood cells in the blood can help pathologists diagnose various blood diseases more accurately. Normal white blood cells are divided into five categories: neutrophils, eosinophils, basophils, lymphocytes and monocytes. The differences between different cells, as well as the small number of white blood cells and the difficulty in finding them, make it a challenging task to accurately identify and detect normal white blood cells.

[0003] At present, hospitals usually use blood analyzers to count white blood cells. This equipment mainly uses physical or chemical schemes for counting, but it has certain limitations and can only perform technical analysis on a small number of white blood cells. For the diagnosis of major dangerous diseases, bone marrow smear cell morphology examination is usually required. In addition, blood analyzers can only be used for counting and cannot present cell images under a microscope for further judgment by clinicians. In comparison, manual counting methods are time-consuming and laborious, and the results of different examination physicians may be different. Therefore, it is necessary to improve the automation level of blood microscopy to better assist doctors in diagnosing major blood diseases.

[0004] With the rapid development of computer technology, deep learning has made breakthrough progress in the field of computer vision. More and more scholars are using deep learning methods to perform cell classification detection. Compared with traditional machine learning methods, deep learning methods can simplify the process of image processing and feature screening, making it more automated. However, there is a difficult problem in the detection of white blood cells in blood. Since there are too many types of cells and the differences between different types of cells are not large, it is of great significance to accurately classify multiple types of white blood cells. Summary of the invention

[0005] The present invention aims to provide a white blood cell detection method based on FBDM-YOLOv8s, which can improve the accuracy of edge detection while optimizing model parameters, speed up the calculation speed, and better solve the problems of incomplete image cells and overlapping of two cells.

[0006] The technical solution of the present invention is as follows:

[0007] The leukocyte detection method based on FBDM-YOLOv8s comprises the following steps:

[0008] A. Construct an improved YOLOv8s neural network model, wherein the improved YOLOv8s neural network model includes: Backbond network, Neck network, and Head network;

[0009] The Backbond network uses multiple groups of Fasternet modules as the main framework, and finally sets a group of BRA modules;

[0010] B. Train the improved YOLOv8s neural network model to obtain a trained improved YOLOv8s neural network model;

[0011] C. The white blood cell image to be detected is introduced into the trained improved YOLOv8s neural network model, and processed by the Backbond network, Neck network, and Head network in sequence to obtain the detected image;

[0012] Among them, the processing process in the Fasternet module is as follows:

[0013] The input result is processed by 3×3PCpnv layer, 1×1 convolution, BN layer, ReLU function, and 1×1 convolution in sequence, and the obtained result is added and fused with the input result to obtain the output result.

[0014] The process of training the improved YOLOv8s neural network model in step B is as follows:

[0015] a. Obtain white blood cell images from the DML-LZWH dataset, preprocess the acquired white blood cell images, and divide them into a training set and a validation set;

[0016] b. Use the data of the training set to train the improved YOLOv8s neural network model to obtain the trained improved YOLOv8s neural network model;

[0017] c. Use the data of the validation set to test the improved YOLOv8s neural network model to obtain a trained improved YOLOv8s neural network model;

[0018] In the step a, the process of preprocessing the acquired white blood cell image is as follows:

[0019] A certain number of white blood cell microscope image data were randomly extracted from the DML-LZWH dataset, and the image data were randomly divided into a training set and a validation set, with the ratio of image data in the training set to that in the validation set being 1:1. All of the above images were from the CellaVision-DI-60 system, and all images were 360×367 pixels. They were peripheral blood films after Wright-Giemsa staining technology, and were taken with a digital camera connected to a microscope. The microscope used an achromatic lens with a magnification of 100 times, and the image storage format was JPG. Labeling of white blood cells was completed using labelimg under the guidance of the laboratory doctor.

[0020] The processing process in the Backbond network is as follows:

[0021] The input result is processed by the Embedding function, Fasternet module, Merging function, Fasternet module, Merging function, and Fasternet module in sequence to obtain a first result. The first result is processed by the Merging function and Fasternet module in sequence to obtain a second result. The second result is processed by the Merging function, Fasternet module, SPPF module, and BRA module in sequence to obtain a third result. The first result, the second result, and the third result are respectively input into the Neck network.

[0022] The processing process in the BRA module is as follows:

[0023] In the first stage, the input result is first divided into S×S regions, each region contains HW / S2 feature vectors, and then the region-level query Qr and key Kr are obtained through average pooling operation in each region, and then the adjacency matrix Ar between regions is calculated using Qr and Kr. Finally, the top-k operation is used to select the k most relevant regions of each region from Ar to obtain the routing index matrix Ir; then the routing index matrix Ir is used to collect the corresponding key-value pairs from the original key K and value V to obtain Kg and Vg, and finally the attention function is applied to the collected key-value pairs Kg and Vg to obtain the output result.

[0024] The processing process in the Neck network is as follows:

[0025] After upsampling, the third result is concatenated with the second result through the Concat function, and then processed by the first DSL4 module to obtain a first DSL4 processing result; after upsampling, the first DSL4 processing result is concatenated with the first result through the Concat function, and then processed by the second DSL4 module to obtain a second DSL4 processing result; the second DSL4 processing result and the first DSL4 processing result are concatenated through the Concat function, and then processed by the third DSL4 module to obtain a third DSL4 processing result; the third DSL4 processing result and the third result are concatenated through the Concat function, processed by the convolution layer, and input into the Head network;

[0026] The second DSL4 processing result is processed by the convolution layer and then input into the Head network;

[0027] The third DSL4 processing result is processed by the convolution layer and then input into the Head network.

[0028] The processing process in the DSL4 module is as follows:

[0029] After the input result is processed by the convolution layer, it is divided into a left part and a right part by the Split function. The right part is processed by the DBB module and the convolution layer in turn to obtain the first convolution result; the first convolution result is processed by the DBB module and the convolution layer in turn to obtain the second convolution result; the left part, the right part, the first convolution result, and the second convolution result are concatenated by the Concat function and resized by the convolution layer to obtain the output result.

[0030] The processing process in the DBB module is as follows:

[0031] The input results are divided into four paths. The first path is processed by 1*1 convolution and BatchNorm function in sequence to obtain the first path processing result; the second path is processed by 1*1 convolution, BatchNorm function, K*K convolution, BatchNorm function in sequence to obtain the second path processing result; the third path is processed by 1*1 convolution, BatchNorm function, average pooling layer, BatchNorm function to obtain the third path processing result; the fourth path is processed by K*K convolution and BatchNorm function to obtain the fourth path processing result;

[0032] The first processing result, the second processing result, the third processing result and the fourth processing result are added and fused, and then subjected to nonlinear processing to obtain an output result.

[0033] The Head network is composed of three MultiSEAM modules;

[0034] The result of the second DSL4 processing is processed by the convolution layer and input into the first MultiSEAM module, and then outputs the large-size target detection result after the first Detect detection;

[0035] The result of the third DSL4 processing is processed by the convolution layer and input into the second MultiSEAM module, and then outputs the medium-sized target detection result after the second Detect detection;

[0036] The third DSL4 processing result and the third result are concatenated by the Concat function, and the result processed by the convolution layer is input into the third MultiSEAM module, and the small-size target detection result is output by the third Detect detection;

[0037] The large-size target detection results, medium-size target detection results, and small-size target detection results are fused through multi-scale features to obtain the output result.

[0038] The processing process in the MultiSEAM module is as follows:

[0039] The input result is divided into three paths, each of which is processed by a GSMM module. Then, the processing results of the three GSMM modules are added and fused with the input result, and then processed by average pooling, and then processed by two layers of fully connected networks. The obtained results are multiplied and fused to obtain the output result.

[0040] Among them, the processing process in the GSMM module is as follows:

[0041] The input result is processed by the patch embedding module, GELU function, and BatchNorm function in sequence to obtain the first function processing result; the first function processing result is processed by the depthwise separable convolution, GELU function, and BatchNorm function in sequence to obtain the first function processing result; the first function processing result is added and fused with the second function processing result, and then processed by point-by-point convolution, GELU function, and BatchNorm function in sequence to obtain the output result;

[0042] Among them, the patches of the patch embedding modules in the three GSMM modules are set to 6, 7, and 8 respectively.

[0043] The full name of the BRA module is: BiLevelRoutingAttention;

[0044] The full name of DSL4 module is: DBBNCSPELAN4

[0045] The full name of DBB module is: Diverse Branch Block.

[0046] The beneficial effects of the present invention are as follows:

[0047] 1. The present invention uses a new improved YOLOv8s network model, which has the ability of fast learning and fast recognition detection of YOLOv8. Because the Fasternet module is used to replace the main framework of YOLOv8, and the attention mechanism BiLevelRoutingAttention (BRA) is introduced, the accuracy of edge detection can be improved while optimizing model parameters. The introduction of DBBNCSPELAN4 module can lightweight the model while realizing diversified feature extraction in multiple branches, reducing the number of parameters and speeding up the calculation speed. The integration of the detection head and the MultiSEAM module better solves the problem of incomplete image cells and overlapping of two cells.

[0048] 2. In view of the incomplete display of white blood cells in the picture and the connection of two white blood cell images, the present invention introduces the MultiSEAM module to effectively handle the detection of white blood cells under occlusion. However, the existing baseline YOLOv8s algorithm is difficult to detect them separately, and there will be false detections and missed detections. Figure 7 shown.

[0049] 3. The method provided by the present invention shows high performance in target detection of white blood cells on the DML-LZWH dataset, can effectively improve detection accuracy, and is helpful for the detection of blood type diseases.

[0050] 4. The target detection method of the present invention can effectively realize the target detection of five types of white blood cells, which can improve the work efficiency of doctors and also provide a guarantee for the screening and diagnosis of blood type diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the overall structure of the FBDM-YOLOv8s network model;

[0052] Figure 2 It is a structural diagram of the Fasternet module;

[0053] Figure 3 It is a schematic diagram of the structure of the BRA module (BiLevelRoutingAttention);

[0054] Figure 4 It is a schematic diagram of the structure of the DSL4 module (DBBNCSPELAN4);

[0055] Figure 5 It is a schematic diagram of the structure of the DBB module (Diverse Branch Block);

[0056] Figure 6It is a schematic diagram of the structure of MultiSEAM module and GSMM module;

[0057] Figure 7 This is the recognition error graph of the existing technical baseline YOLOv8s algorithm;

[0058] Figure 8 This is a diagram of white blood cell detection results, from left to right they are YOLOv3-tiny, YOLOv5, YOLOv6, YOLOv7, YOLOv8s, YOLOv9, YOLOv10, and the FBDM-YOLOv8s algorithm of the present invention. DETAILED DESCRIPTION

[0059] The present invention is specifically described below with reference to the accompanying drawings and embodiments.

[0060] Example 1

[0061] The leukocyte detection method based on FBDM-YOLOv8s of this embodiment includes the following steps:

[0062] A. Construction Figure 1 The improved YOLOv8s neural network model shown, the improved YOLOv8s neural network model includes: Backbond network, Neck network, Head network;

[0063] The Backbond network uses multiple groups of Fasternet modules as the main framework, and finally sets a group of BRA modules;

[0064] B. Train the improved YOLOv8s neural network model to obtain a trained improved YOLOv8s neural network model;

[0065] C. The white blood cell image to be detected is introduced into the trained improved YOLOv8s neural network model, and processed by the Backbond network, Neck network, and Head network in sequence to obtain the detected image;

[0066] Among them, Figure 2 As shown, the processing process in the Fasternet module is as follows:

[0067] The input result is processed by 3×3PCpnv layer, 1×1 convolution, BN layer, ReLU function, and 1×1 convolution in sequence, and the obtained result is added and fused with the input result to obtain the output result.

[0068] The process of training the improved YOLOv8s neural network model in step B is as follows:

[0069] a. Obtain white blood cell images from the DML-LZWH dataset, preprocess the acquired white blood cell images, and divide them into a training set and a validation set;

[0070] b. Use the data of the training set to train the improved YOLOv8s neural network model to obtain the trained improved YOLOv8s neural network model;

[0071] c. Use the data of the validation set to test the improved YOLOv8s neural network model to obtain a trained improved YOLOv8s neural network model;

[0072] In the step a, the process of preprocessing the acquired white blood cell image is as follows:

[0073] A certain number of white blood cell microscope image data were randomly extracted from the DML-LZWH dataset, and the image data were randomly divided into a training set and a validation set, with the ratio of image data in the training set to that in the validation set being 1:1. All of the above images were from the CellaVision-DI-60 system, and all images were 360×367 pixels. They were peripheral blood films after Wright-Giemsa staining technology, and were taken with a digital camera connected to a microscope. The microscope used an achromatic lens with a magnification of 100 times, and the image storage format was JPG. Labeling of white blood cells was completed using labelimg under the guidance of the laboratory doctor.

[0074] The processing process in the Backbond network is as follows:

[0075] The input result is processed by the Embedding function, Fasternet module, Merging function, Fasternet module, Merging function, and Fasternet module in sequence to obtain a first result. The first result is processed by the Merging function and Fasternet module in sequence to obtain a second result. The second result is processed by the Merging function, Fasternet module, SPPF module, and BRA module in sequence to obtain a third result. The first result, the second result, and the third result are respectively input into the Neck network.

[0076] like Figure 3 As shown, the processing process in the BRA module is as follows:

[0077] In the first stage, the input result is first divided into S×S regions, each region contains HW / S2 feature vectors, and then the region-level query Qr and key Kr are obtained through average pooling operation in each region, and then the adjacency matrix Ar between regions is calculated using Qr and Kr. Finally, the top-k operation is used to select the k most relevant regions of each region from Ar to obtain the routing index matrix Ir; then the routing index matrix Ir is used to collect the corresponding key-value pairs from the original key K and value V to obtain Kg and Vg, and finally the attention function is applied to the collected key-value pairs Kg and Vg to obtain the output result.

[0078] The processing process in the Neck network is as follows:

[0079] After upsampling, the third result is concatenated with the second result through the Concat function, and then processed by the first DSL4 module to obtain a first DSL4 processing result; after upsampling, the first DSL4 processing result is concatenated with the first result through the Concat function, and then processed by the second DSL4 module to obtain a second DSL4 processing result; the second DSL4 processing result and the first DSL4 processing result are concatenated through the Concat function, and then processed by the third DSL4 module to obtain a third DSL4 processing result; the third DSL4 processing result and the third result are concatenated through the Concat function, processed by the convolution layer, and input into the Head network;

[0080] The second DSL4 processing result is processed by the convolution layer and then input into the Head network;

[0081] The third DSL4 processing result is processed by the convolution layer and then input into the Head network.

[0082] like Figure 4 As shown, the processing process in the DSL4 module is as follows:

[0083] After the input result is processed by the convolution layer, it is divided into a left part and a right part by the Split function. The right part is processed by the DBB module and the convolution layer in turn to obtain the first convolution result; the first convolution result is processed by the DBB module and the convolution layer in turn to obtain the second convolution result; the left part, the right part, the first convolution result, and the second convolution result are concatenated by the Concat function and resized by the convolution layer to obtain the output result.

[0084] like Figure 5 As shown, the processing process in the DBB module is as follows:

[0085] The input results are divided into four paths. The first path is processed by 1*1 convolution and BatchNorm function in sequence to obtain the first path processing result; the second path is processed by 1*1 convolution, BatchNorm function, K*K convolution, BatchNorm function in sequence to obtain the second path processing result; the third path is processed by 1*1 convolution, BatchNorm function, average pooling layer, BatchNorm function to obtain the third path processing result; the fourth path is processed by K*K convolution and BatchNorm function to obtain the fourth path processing result;

[0086] The first processing result, the second processing result, the third processing result and the fourth processing result are added and fused, and then subjected to nonlinear processing to obtain an output result.

[0087] The Head network is composed of three MultiSEAM modules;

[0088] The result of the second DSL4 processing is processed by the convolution layer and input into the first MultiSEAM module, and then outputs the large-size target detection result after the first Detect detection;

[0089] The result of the third DSL4 processing is processed by the convolution layer and input into the second MultiSEAM module, and then outputs the medium-sized target detection result after the second Detect detection;

[0090] The third DSL4 processing result and the third result are concatenated by the Concat function, and the result processed by the convolution layer is input into the third MultiSEAM module, and the small-size target detection result is output by the third Detect detection;

[0091] The large-size target detection results, medium-size target detection results, and small-size target detection results are fused through multi-scale features to obtain the output result.

[0092] like Figure 6 As shown, the processing process in the MultiSEAM module is as follows:

[0093] The input result is divided into three paths, each of which is processed by a GSMM module. Then, the processing results of the three GSMM modules are added and fused with the input result, and then processed by average pooling, and then processed by two layers of fully connected networks. The obtained results are multiplied and fused to obtain the output result.

[0094] Among them, the processing process in the GSMM module is as follows:

[0095] The input result is processed by the patch embedding module, GELU function, and BatchNorm function in sequence to obtain the first function processing result; the first function processing result is processed by the depthwise separable convolution, GELU function, and BatchNorm function in sequence to obtain the first function processing result; the first function processing result is added and fused with the second function processing result, and then processed by point-by-point convolution, GELU function, and BatchNorm function in sequence to obtain the output result;

[0096] Among them, the patches of the patch embedding modules in the three GSMM modules are set to 6, 7, and 8 respectively.

[0097] Example 2

[0098] The white blood cell target detection algorithm studied is mainly used in the field of medical impact and detection, and has high requirements for the real-time performance and accuracy of the algorithm. Therefore, the performance of the one-stage algorithm is compared with the algorithm FBDM-YOLOv8s of the present invention under the same conditions. The one-stage algorithm is mainly based on the YOLO series algorithm. The algorithm selected for comparison in the present invention is the detection network of various versions of the YOLO series.

[0099] In this embodiment, YOLOv3-tiny, YOLOv5s, YOLOv6, YOLOv7, YOLOv9, and YOLOv10 are compared with the algorithm FBDM-YOLOv8s of the present invention using the same data set, as shown in Table 1.

[0100] Table 1 Comparison results of the DMC-Unet network model of the present invention and multiple related networks

[0101]

[0102] It can be seen from the data in the table that YOLOv7 has a poor effect in white blood cell target detection, with an mAP of only 0.951, which does not meet the actual application requirements in the medical field. The detection accuracy of YOLOv5 and YOLOv6 is lower than that of the algorithm of the present invention. The network structure of YOLOv5 is simpler, so the detection speed is faster. Compared with the advanced YOLOv9 and YOLOv10 algorithms, although the number of parameters is greatly reduced, the amount of calculation is greatly increased compared with the algorithm of the present invention, which increases the calculation time, and the mAP is lower than that of the algorithm of the present invention. By comparison, the algorithm of the present invention performs best in Precision, Recall, mAP and F1 scores. Compared with the basic YOLOv8s, the mAP is increased by 2.1%. Although the number of parameters has increased, the amount of calculation is greatly reduced, so the detection speed is improved.

[0103] (3) From Figure 8It can be seen that the data set composed of five types of normal white blood cells is used to evaluate the detection results of YOLOv3-tiny, YOLOv5s, YOLOv6, YOLOv7, YOLOv8s, YOLOv9, YOLOv10 algorithms and the FBDM-YOLOv8s algorithm of the present invention. The same parameters are used for testing. The comparison algorithm has false detection, missed detection, low precision, etc. The FBDM-YOLOv8s algorithm of the present invention can accurately detect some obscured cells while improving the detection accuracy.

[0104] The above description is a detailed description of the preferred feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modified changes completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A leukocyte detection method based on FBDM-YOLOv8s, characterized in that: The following steps are involved: A. Construct an improved YOLOv8s neural network model, wherein the improved YOLOv8s neural network model includes: Backbond network, Neck network, and Head network; The Backbond network uses multiple groups of Fasternet modules as the main framework, and finally sets a group of BRA modules; B. Train the improved YOLOv8s neural network model to obtain a trained improved YOLOv8s neural network model; C. The white blood cell image to be detected is introduced into the trained improved YOLOv8s neural network model, and processed by the Backbond network, Neck network, and Head network in sequence to obtain the detected image; Among them, the processing process in the Fasternet module is as follows: The input result is processed by 3×3PCpnv layer, 1×1 convolution, BN layer, ReLU function, and 1×1 convolution in sequence, and the obtained result is added and fused with the input result to obtain the output result.

2. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 1, characterized in that: The process of training the improved YOLOv8s neural network model in step B is as follows: a. Obtain white blood cell images from the DML-LZWH dataset, preprocess the obtained white blood cell images, and divide them into a training set and a validation set; b. Use the data of the training set to train the improved YOLOv8s neural network model to obtain the trained improved YOLOv8s neural network model; c. Use the data of the validation set to test the improved YOLOv8s neural network model to obtain a trained improved YOLOv8s neural network model.

3. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 2, characterized in that: In the step a, the process of preprocessing the acquired white blood cell image is as follows: A certain number of white blood cell microscope image data were randomly extracted from the DML-LZWH dataset, and the image data were randomly divided into a training set and a validation set, with the ratio of image data in the training set to that in the validation set being 1:

1. All of the above images were from the CellaVision-DI-60 system, and all images were 360×367 pixels. They were peripheral blood films after Wright-Giemsa staining technology, and were taken with a digital camera connected to a microscope. The microscope used an achromatic lens with a magnification of 100 times, and the image storage format was JPG. Labeling of white blood cells was completed using labelimg under the guidance of the laboratory doctor.

4. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 1, characterized in that: The processing process in the Backbond network is as follows: The input result is processed by the Embedding function, Fasternet module, Merging function, Fasternet module, Merging function, and Fasternet module in sequence to obtain a first result. The first result is processed by the Merging function and Fasternet module in sequence to obtain a second result. The second result is processed by the Merging function, Fasternet module, SPPF module, and BRA module in sequence to obtain a third result. The first result, the second result, and the third result are respectively input into the Neck network.

5. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 4, characterized in that: The processing process in the BRA module is as follows: In the first stage, the input result is first divided into S×S regions, each region contains HW / S² feature vectors, and then the region-level query Qr and key Kr are obtained through average pooling operation in each region, and then the adjacency matrix Ar between regions is calculated using Qr and Kr. Finally, the top-k operation is used to select the k most relevant regions of each region from Ar to obtain the routing index matrix Ir; then the routing index matrix Ir is used to collect the corresponding key-value pairs from the original key K and value V to obtain Kg and Vg, and finally the attention function is applied to the collected key-value pairs Kg and Vg to obtain the output result.

6. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 5, characterized in that: The processing process in the Neck network is as follows: After upsampling, the third result is concatenated with the second result through a Concat function, and then processed by the first DSL4 module to obtain a first DSL4 processing result; After upsampling, the first DSL4 processing result is concatenated with the first result through a Concat function, and then processed by the second DSL4 module to obtain a second DSL4 processing result; The second DSL4 processing result and the first DSL4 processing result are concatenated by a Concat function, and then processed by a third DSL4 module to obtain a third DSL4 processing result; The third DSL4 processing result and the third result are concatenated by the Concat function, processed by the convolution layer, and input into the Head network; The second DSL4 processing result is processed by the convolution layer and then input into the Head network; The third DSL4 processing result is processed by the convolution layer and then input into the Head network.

7. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 6, characterized in that: The processing process in the DSL4 module is as follows: After the input result is processed by the convolution layer, it is divided into a left part and a right part by the Split function. The right part is processed by the DBB module and the convolution layer in turn to obtain the first convolution result; the first convolution result is processed by the DBB module and the convolution layer in turn to obtain the second convolution result; the left part, the right part, the first convolution result, and the second convolution result are concatenated by the Concat function and resized by the convolution layer to obtain the output result.

8. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 7, characterized in that: The processing process in the DBB module is as follows: The input results are divided into four paths. The first path is processed by 1*1 convolution and BatchNorm function in sequence to obtain the first path processing result; the second path is processed by 1*1 convolution, BatchNorm function, K*K convolution, BatchNorm function in sequence to obtain the second path processing result; the third path is processed by 1*1 convolution, BatchNorm function, average pooling layer, BatchNorm function to obtain the third path processing result; the fourth path is processed by K*K convolution and BatchNorm function to obtain the fourth path processing result; The first processing result, the second processing result, the third processing result and the fourth processing result are added and fused, and then subjected to nonlinear processing to obtain an output result.

9. The leukocyte detection method based on FBDM-YOLOv8s as claimed in claim 6, characterized in that: The Head network is composed of three MultiSEAM modules; The result of the second DSL4 processing is processed by the convolution layer and input into the first MultiSEAM module, and then outputs the large-size target detection result after the first Detect detection; The result of the third DSL4 processing is processed by the convolution layer and input into the second MultiSEAM module, and then outputs the medium-sized target detection result after the second Detect detection; The third DSL4 processing result and the third result are concatenated by the Concat function, and the result processed by the convolution layer is input into the third MultiSEAM module, and the small-size target detection result is output by the third Detect detection; The large-size target detection results, medium-size target detection results, and small-size target detection results are fused through multi-scale features to obtain the output result.

10. The leukocyte detection method based on FBDM-YOLOv8s according to claim 9, characterized in that: The processing process in the MultiSEAM module is as follows: The input result is divided into three paths, each of which is processed by a GSMM module. Then, the processing results of the three GSMM modules are added and fused with the input result, and then processed by average pooling, and then processed by two layers of fully connected networks. The obtained results are multiplied and fused to obtain the output result. Among them, the processing process in the GSMM module is as follows: The input result is processed by the patch embedding module, GELU function, and BatchNorm function in sequence to obtain the first function processing result; the first function processing result is processed by the depthwise separable convolution, GELU function, and BatchNorm function in sequence to obtain the first function processing result; the first function processing result is added and fused with the second function processing result, and then processed by point-by-point convolution, GELU function, and BatchNorm function in sequence to obtain the output result; Among them, the patches of the patch embedding modules in the three GSMM modules are set to 6, 7, and 8 respectively.

Images