Research on Blood Cell Detection and Counting Based on YOLO-BC Algorithm

doi:10.21203/rs.3.rs-3992930/v1

Download PDF

Research Article

Research on Blood Cell Detection and Counting Based on YOLO-BC Algorithm

https://doi.org/10.21203/rs.3.rs-3992930/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Blood cell detection and counting have always been of great medical importance because healthcare professionals can identify possible abnormalities in patients based on the numbers of different types of blood cells. However, traditional counting methods have certain limitations, such as smaller blood cells being ignored or misclassified, and the accuracy of identifying scenes with occlusion and overlap between blood cells is low. To solve the problem of blood cells being missed or misidentified in existing models, this study introduces the improved YOLO-BC algorithm to solve the pixel-level differences of different categories of blood cells by combining efficient multi-scale attention and full-dimensional dynamic convolution models, thereby achieving fast and accurate identification and counting of blood cells. The BCCD (Blood Cell Count and Detection) dataset was used for related experiments and performed data augmentation. The mAP@50 score based on YOLO-BC is 3.1% higher than that of YOLOv8, the value of mAP@50:95 increases by 3.7%, and F1-score increases by 2% on the same dataset and iou parameters, where small objects such as platelets can also be clearly detected. YOLO-BC shows a certain degree of applicability for automated testing of blood cells by experimental results.

Blood cell detection

Object Detection

YOLOv8

EMSA

ODConv Automated

In the field of medicine, the diagnosis of patients' medical conditions can be made by detecting the quantity of three types of blood cells (red blood cells, white blood cells, and platelets) [1] in the human bloodstream system, as shown in in Fig. 1. For example, an excessive number of erythrocytes may indicate vascular obstruction, while a deficiency may suggest anemia. The quantity of leukocytes can provide indications of bacterial invasion [2].

Traditional detection methods require professionals to use devices such as microscope equipment, which consume significant manpower and resources [3]. Traditional image processing techniques for erythrocyte recognition, such as thresholding, edge detection, and morphological operations, have been widely used but may struggle with variations in erythrocyte morphology and the presence of overlapping or clustered cells. In contrast, machine learning-based approaches, particularly deep learning and convolutional neural networks (CNNs), have demonstrated high accuracy in erythrocyte recognition tasks even in the presence of challenging image features. A study explored a YOLO-Dense structure, which introduced dense connections into YOLO [4], achieving an mAP@50 (if there is more than 50% overlap, the detection is considered accurate) of 86% and outperforming Faster Region CNN (R-CNN).

Based on the research background mentioned above, a YOLO-BC (YOLOv8 for Blood Cell) algorithm which is an improved algorithm for detecting blood cells based on existing YOLOv8, is proposed by employing efficient multi-scale attention (EMSA) and omni-dimensional dynamic convolution (ODConv) in backbone. The improved YOLOv8 achieves higher F1-score than YOLOv8, and both mAP@50 and mAP@50:95 are superior to the YOLOv8. This study addresses scenarios of overlapping blood cells and improve accuracy in detecting small cell scenes by YOLO-BC, showing potential applicability and generalizability.

There are three main contributions of this study:

By using digital image processing methods such as flipping, scaling, and increasing brightness, the small-sample blood cell dataset was enhanced for insufficient cell samples and shapes, enriching the blood cell detection scenario and making it closer to actual requirements.

Effective Multi-Scale Attention Mechanism introduced into specific C2f module of YOLOv8, where EMSA enhances the feature maps generated by the backbone network by incorporating attention mechanisms across multiple scales.

Omni-Dimensional Dynamic Convolution used to replace the specified convolution module of YOLOv8, where it can adaptively combine convolution kernels from different dimensions, enhancing the capability of feature representation on different blood cells.

In recent years, there have been significant advancements in research related to blood cell recognition. Since the birth of LeNet [5], a series of significant architectures and technologies have emerged. In the ImageNet image classification challenge of 2012 [6], breakthrough progress was made by introducing deep convolutional neural networks. In the 2014 ImageNet competition, VGGNet introduced a CNN architecture that enhances performance by using smaller convolutional kernels and deeper layers [6]. In the 2015 ImageNet competition, ResNet introduced residual connections, effectively addressing issues such as gradient vanishing and model degradation [7].

Specifically, for the task of medical cell counting, researchers have proposed a series of deep learning network architectures [8]. Faster R-CNN introduces a region proposal network (RPN) and ROI pooling layer, combining object localization and classification. YOLO [8] is renowned for its real-time object detection speed, transforming the object detection task into a regression problem to predict object positions and categories simultaneously. YOLO-Dense introduces Dense connections in YOLO, achieving better detection performance compared to Faster R-CNN with an mAP@50 score of 86% [9].

Given the relatively small scale of medical cell image datasets, data augmentation techniques are widely employed to increase data sample diversity. For instance, operations such as mirror flipping, rotation, scaling, and translation can be applied using affine transformations to augment the diversity of cell images, thereby improving model robustness and generalization ability.

This study delves into an in-depth investigation of various object detection algorithms for blood cells in their applications for blood cell recognition and counting, and build an improved YOLO-BC algorithm based on YOLOv8, by incorporating efficient multi-scale attention and omni-dimensional dynamic convolution models. Then, this study selects BCCD (Blood Cell Count and Detection) [9] as the dataset and performed data augmentation. Focusing on the two innovative points of EMSA and ODConv, YOLOv8-related ablation experiments are conducted in detail. Then, experimental comparisons are conducted to evaluate these algorithms in terms of performance metrics, including precision, recall, mAP, F1 score.

3.1. The design of YOLO-BC detection pipeline

For the original blood cell data set and YOLOv8-BC model, this study designed the training mechanism and model architecture as shown in Fig. 1. The original data set only included 364 picture samples. Through data augmentation, a data set with more comprehensive coverage of the scene was obtained, and then divided into a training set (765 images), a validation set (73 images) and a testing set (36 images). There were a total of 11,789 blood cell samples, including 10,031 red blood cells, 898 white blood cells, and 851 platelets. Finally, the YOLO-BC model is obtained after training on the dataset.

In terms of blood cell detection and counting, the YOLO-BC should be designed to automatically identify and count different types of blood cells in blood images, such as red blood cells, white blood cells, and platelets.

YOLO-BC performs image feature extraction and detection through a convolutional neural network (CNN), then divides the image into grids and predicts possible object bounding boxes and categories on each grid (Fig. 2 (a)). This enables YOLO-BC to maintain high detection speed and accuracy while processing large numbers of blood cells.

As shown in the Fig. 2 (b), taking white blood cell detection as an example, the YOLO-BC algorithm provides bounding boxes and corresponding judgment probabilities for the detection targets, and finally achieves the detection of corresponding types of blood cells. This process design plays an important role in the promotion of YOLO-BC in practical medical applications. Traditional medical diagnosis often relies on manual visual inspection and analysis, and the introduction of YOLO-BC visual detection software has the potential to provide doctors with a fast and accurate blood cell-related diagnosis method.

Taking the reading of microscope images as an example, doctors usually need to observe and analyze cells in a large number of blood samples and determine whether there are abnormalities. By using visual detection software developed by YOLO-BC pipeline, doctors only need to input microscope images into the system, and the software will automatically identify and locate blood cells, providing fast and accurate diagnostic results.

3.2. YOLO-BC algorithm

Based on the YOLOv8 model, incorporating the aforementioned EMSA and ODConv, this study designed and implemented an improved network architecture called YOLO-BC (YOLOv8 for Blood Cell), as illustrated in the Fig. 3 below.

EMSA is put to end of C2f module which is connected to second Concat structure of neck part, and the third Conv module is modified to ODConv in backbone part. The EMSA module can effectively aggregate multi-scale feature information, giving the network better perception capabilities at different scales. By placing the EMSA module at the end of the C2f module, the feature representation capability of the module can be enhanced and detection accuracy improved. By introducing the ODConv module into the backbone part, the network's ability to extract target features can be improved, thereby improving detection performance.

Then, YOLOv8 model has made significant improvements in its network structure. The backbone is primarily used for feature extraction, and YOLOv8 replaces the CSP (Cross-Stage Partial) module in YOLOv5 with a lightweight C2f (Context to Features) module, enhancing feature representation through dense residual structures. The tail of the network utilizes a SPPF (Fast Spatial Pyramid Pooling layer) to increase receptive field and capture features at different levels in the scene. The neck section is mainly used for feature fusion, utilizing path aggregation networks and the C2f module to fuse feature maps of different scales from the three stages of the backbone.

In blood cell images, the presence of occlusion and overlap between blood cells adds to the challenge of object detection. Thus, an attention model was presented in the Fig. 3 to overcome the drawback on blood cell detection of YOLOv8, such as repetitive counting on overlapping blood cell and low recognition accuracy in small target blood cell recognition, etc.

3.3. Efficient Multi-Scale Attention

The EMSA (Efficient Multi-Scale Attention) mechanism is a novel attention module that has gained considerable attention recently [9, 10]. Its core idea is to reduce computational costs while preserving information on each channel.

By reshaping a portion of the channels into the batch dimension, the EMSA mechanism can process channel information more efficiently, as is shown in the Fig. 4. By grouping the channel dimension into multiple sub-features, the module ensures an even distribution of spatial semantic features in each feature group. Assume that the input feature map is X, there are m samples in total, each sample has n features, and the feature map is divided into G groups. For each group g, calculate the mean ${\mu }_{g}$ and variance ${\sigma }_{g}$ of the features within the group, defined as [10]:

$${\mu }_{g}=\sum {X}_{i}/(n/GHW)$$

$${{\sigma }_{g}}^{2}=\sum {({X}_{i}-{\mu }_{g})}^{2}/(n/GHW)$$

Intuitively, H and W represent the height and width of the feature map respectively.

Then, it normalizes the features ${X}_{i}$ within the group as:

$${Y}_{i}=({X}_{i}-{\mu }_{g})/\sqrt{{{\sigma }_{g}}^{2}+\epsilon }$$

where ε is a small positive number to prevent the denominator from being zero. Finally, the normalized feature Y is restored to its original shape.

This grouping improves the expression capability of different blood cells’ pixel features and reduces computational costs, enhancing model efficiency.

The EMSA mechanism encodes global information and adjusts channel weights in each parallel branch to further improve channel weight calibration. Taking advantage of the benefits provided by the novel EMSA structure, the C2f module of YOLOv8 is improved by introducing a multi-scale attention mechanism that is capable of addressing differences in pixel-level features across different blood cell categories, especially for small targets like platelets, and modifying the backbone P4/16 after C2f accordingly.

On the other hand, EMSA mechanism utilizes cross-dimensional interaction further to aggregate the output features from two parallel branches. This allows the module to capture pairwise relationships at the pixel level of blood cells, better understanding the connections between information from different positions in the image. This cross-dimensional interaction enhances feature consistency and accuracy, further enhancing the expressive power of the model. In the backbone, EMSA is added before the bottleneck module (Fig. 5).

3.4. Omni-Dimensional Dynamic Convolution

ODConv (Omni-Dimensional Dynamic Convolution) is [11] a full-dimensional dynamic convolution method that extends and expands upon CondConv [12]. It considers the dynamics of multiple dimensions, such as spatial domain, input channels, and output channels, and improves model performance through parallel strategies and multi-dimensional attention mechanisms to learn complementary attention.

The core idea of ODConv is to decompose the convolution operation into multiple sub-operations and introduce dynamic weights for feature representation. By using matrix decomposition techniques [13], it decomposes the convolution kernel into the product of two low-rank matrices, reducing the number of parameters and computations, and improving model efficiency.

ODConv acts as a plug-and-play operation that can be conveniently embedded into existing YOLOv8 models to address the low detection accuracy issue in RBC by selectively replacing parts of the Conv modules in the existing structure of YOLOv8. And this study further optimizes the existing ODConv to version 2, by connecting a batch normalization layer afterwards and selecting the SiLU activation function, then it makes adjustments to Conv module of YOLOv8 P3/8 pixel level, including spatial size, input channel size and output channel size. And all the following results are using ODConv version2 in YOLO-BC.

3.5. Dataset

Data augmentation was performed on the original BCCD dataset [13, 14], thereby improving the model's ability to generalize over unseen data. Each image has a 50% probability of being horizontally flipped. This transformation aids the model in learning to recognize objects from different perspectives, enhancing its ability to generalize and potentially improving detection performance. Similarly, each image has a 50% chance of undergoing a vertical flip.

This augmentation can particularly help in cases where objects may appear upside down or in different vertical orientations in real-world scenarios. And a portion of the images ranging from 0 to 15% is randomly cropped, which can help the model focus on various parts of an image. Furthermore, the exposure of the images is varied within a range of -20% to + 20%. In total, 874 clinical blood cell image samples were used as the dataset. Among them, 765 samples were randomly selected for the training set, 73 samples for the validation set, and 36 samples for the test set.

In the training set split by BCCD, the number of platelet instances is 739, the number of red blood cell instances is 8814, and the number of white blood cell instances is 789 in total (Fig. 6 (a)). At the same time, normalized statistics were performed on the position of the detection box from 0 to 1, and it was found that the length and width of the detection frame were mainly concentrated around 0.2. Here, the x and y axis represent relative positions of all the detection boxes in blood cell dataset, so they have no units (Fig. 6 (c)), and the width and height axis represent relative size of the detection boxes (Fig. 6 (d)).

Label correlogram can facilitate the detection of patterns or correlations in the spatial arrangement of object annotations, wherein such analysis involves distinct classes and scales. This technique can uncover instances where particular classes exhibit high co-occurrence within a single image or where specific classes tend to manifest at particular scales with greater frequency. The figure below is a label correlation plot for a blood cell dataset in xywh space (Fig. 7), for each label's x, y, width and height variables of detection box, their corresponding relationships are displayed. And the size distribution of the blood cell data set is relatively balanced and is more suitable for training.

3.6. Statistics

In this study, accuracy (Precision), recall rate (Recall), F1 score (F1), mean average precision (mAP) [15, 16], number of parameters (Params), computational complexity (Flops), and frames per second (FPS) [17] were used as evaluation metrics for detection performance. Below is a detailed introduction of these metrics along with their corresponding formulas:

(1) Precision

Precision represents the ratio of true positives in the samples predicted as positive (the blood cell’s actual class), indicating the model accuracy for detection. Its formula is [17]:

$$Precision=\frac{TP}{TP+FP} \left(3\right)$$

In Formula (3), TP represents the number of true positives and FP represents the number of false positives.

(2) Recall

Recall represents the proportion of true positives detected among all positive samples, which is used to measure the recall rate of the model. Its formula is:

$$Recall=\frac{TP}{TP+FN} \left(4\right)$$

In Formula (4), TP represents the number of true positives, and FN represents the number of false negatives.

(3) F1 score

Combines Precision and Recall evaluating the performance of the model comprehensively. Its formula is:

$$F1=\frac{2\times Precision\times Recall}{Precision+Rrcall} \left(5\right)$$

(4) mAP@50 and mAP@50–95

mAP (Mean Average Precision) is a frequently employed measure for evaluating performance of blood cell object detection, measuring the matching degree between the inference results of blood cells and the truth labels, and the overall formula is:

$$mAP=\frac{1}{n}\sum _{i=1}^{n}{AP}_{i} \left(6\right)$$

In Formula (6), n represents the number of object categories, and ${AP}_{i}$ represents the Average Precision of the i-th category, where mAP@50 and mAP@50–90 are the mAP values, where the threshold of IoU is set to 0.5 and 0.5–0.95 respectively.

4.1. Experimental Environment

The experiment was conducted using the pytorch 2.0.0 training framework [18] on a linux operating system environment with an NVDIA GeForce RTX 3090 graphics card with 24GB of memory [19]. The model training parameters were set to 100 epochs, and the optimizer used Adam with an initial learning rate of 0.01 and a momentum of 0.937.

4.2. Training

The training process of the improved model is shown in the Fig. 8 below, where the x-axis represents the number of epoches and the y-axis represents the corresponding indicator value. The YOLO-BC model params, Flops and FPS are 3072484, 8.3G and 22.4 f/s respectively.

For the purpose of comparison, the YOLO-BC model have the default YOLOv8 hyperparameters where batch size and the number of epochs is 16 and 100 respectively. At the end of the training, all indicators [20] of the model converge, where value of mAP@50 is 0.927 and mAP@50:95 is 0.653 on train dataset. But the true performance of the blood cell detection model needs to be evaluated on a test set in order to truly reflect the robustness of the YOLO-BC model, and experimental results of train set is the reference of improvements in model.

It indicates that when the IoU is 0.5, the YOLO-BC model shows a small improvement compared to YOLOv8 on training dataset of blood cells (Fig. 9, (b), (c) and (d) compared to (a)). In contrast, as the IoU increases, the model exhibits significant enhancement on mAP, albeit with relatively lower value. Both recall and precision also experience varying degrees of improvement.

4.3. Ablation Environment

An ablation experiment on test dataset was designed to validate the effectiveness of the improved YOLOv8 model. The experimental results are shown in Table 1. The detection accuracy of YOLO-BC, as measured by mAP@50 and mAP@50:95, reached 0.901 and 0.62 respectively.

Table 1

Ablation experiment on test dataset of BCCD
Model	P/%	R/%	mAP@50	mAP@50:95	FPS
YOLOv8	79.9	87.5	87	58.3	20.6
YOLOv8 + EMSA	78.9	87.2	88	59.8	22.6
YOLOv8 + ODConv	81.4	86.5	88.2	61.1	25.9
YOLO-BC	80.9	88.9	90.1	62	22.4

These values were the highest among multiple models, demonstrating the significant advantages of the improved model (Table 1). From the Precision-Recall comparison (Fig. 10), the YOLOv8 + ODConv model is slightly ahead in Precision, while YOLOv8 + EMSA is slightly inferior in Recall. However, in terms of evaluating mAP, the most important indicator for blood cell target detection, the YOLO-BC model achieved the highest scores on both indicator by the comparison of mAP@50 and mAP@50:95. Ablation experiments demonstrate the effectiveness of these two improvements, EMSA and ODConv (Fig. 11).

Compared with YOLOv8, the separate improvements of EMSA and ODConv have brought about an increase in the core indicator mAP@50. Considering the different types of blood cells, the mAP@50 of Platelets has increased from 0.855 to 0.873, the mAP of RBC has increased from 0.782 to 0.834, and the mAP of WBC has increased from 0.974 to 0.994 (Fig. 12). The final YOLO-BC model mAP@50 increased from 0.87 to 0.901 compared to before the improvement on YOLOv8.

4.4. Comparison experiment

In order to further improve the superiority of the YOLO-BC model, this study selected different versions of mainstream object detection models, including SSD [21, 22], Faster-RCNN [23, 24], YOLOv5 [25, 26, 27], for comparative experiments. The results are shown in Table 2.

As is shown in Table 2, YOLO-BC achieves the highest F1 score (85%) and mAP@50 (90.1%), which indicates a better overall performance in blood cell object detection in all.

Table 2

Comparison experiment of different models.
Model	F1	P/%	R/%	[email protected]
SSD	66.3	80.4	64.9	75.7
Faster-RCNN	56.3	53.9	72.8	67.2
YOLOv5	81.9	78.8	85.8	87
YOLOv8	83	79.9	87.5	87
YOLO-BC	85	80.9	88.9	90.1

Furthermore, it has the highest precision (80.9%) and recall (89%) among all the models, which demonstrates its excellent ability to both detect and identify objects accurately.

4.5. Visualization

After completing the ablation experiment and comparison experiment, the feature map visualization [28, 29] is conducted on YOLO-BC detection for blood cells. And the feature map visualization refers to the visual display of the intermediate layer feature map of the neural network during the inference process, so as to better understand the working principle of the model and the learned features [30].

As shown in the Fig. 13, it can be found that YOLO-BC fully extracts the pixel features of various types of blood cells and clearly obtains and displays the outline of different blood cells. However, YOLO-BC still needs to process various features to detect and classify blood cells. Through the processing of multi-layer neural network structure, YOLO-BC will fully learn the pixel characteristics of blood cells.

In order to evaluate the performance of the proposed algorithm more intuitively, the detection results of YOLOv8 and YOLO-BC on the test dataset are displayed (Fig. 14). It can be observed that YOLOv8 exhibits lower confidence in detecting RBCs in the same region, leading to a severe degradation in the counting accuracy of RBCs, as shown in corresponding positions at both ends of the dotted line (Fig. 14). In contrast, the improved YOLO-BC achieves precise detection and counting of the same blood cell targets, and it can be confirmed that YOLO-BC is capable of more accurately localizing and recognizing blood cells.

Considering the actual needs of medical staff for blood cell diagnosis of the subject, for example, if they need to count only the number of red blood cells to determine diseases related to red blood cell abnormalities [31, 32, 33], they can flexibly specify the category label of the YOLO-BC algorithm to only detect red blood cells and counting, as is illustrated in Fig. 15. At the same time, medical staff can better focus on the types of blood cells they want to observe, avoiding visual interference caused by simultaneous detection.

And it is noteworthy that YOLO-BC can also detect small targets such as platelets more accurately, indicating that the model has certain scalability in general cell diagnosis and has the potential to detect small objects in microfluidic images.

In this study, an improved blood cell image detection and counting method called YOLO-BC based on the YOLOv8 model is proposed. EMSA is introduced before the bottleneck of the backbone understanding the spatial relationships between information in different locations of image, and ODConv is selectively substituted for the P3/8 pixel Conv module to enhance the model's ability to capture and leverage omni-dimensional context. Experimental results demonstrate that the improved model achieves a superior detection accuracy of 0.901 for mAP@50 and 0.62 for mAP@50:95.

And YOLO-BC tackles the pixel-level differences in blood cells of different categories, addressing the limitations of traditional counting methods by incorporating EMSA and ODConv module. These advancements contribute to overcoming challenges such as missed detections, false positives, and repeated counting in complex scenarios.

By comparing SSD, Faster R-CNN, YOLOv5 and YOLOv8, YOLO-BC achieves the highest F1 score (85%) and mAP@50 (90.1%), which indicates a better overall performance in object detection. Additionally, it has the highest precision (80.9%) and recall (89%) among all the models, which demonstrates its excellent ability to both detect and identify objects accurately. Therefore, YOLO-BC shows significant advantages in accuracy compared to other widely used models, making it a promising option for object detection applications. In future work, data augmentation methods [34, 35, 36], implementation of lightweight models [37, 38], and fine-tuning of the model will be pursued. Data enhancement can improve the shortcomings of incomplete coverage of scenes with a small number of samples, and the lighter YOLO-BC model will have faster inference speed, greatly speeding up the diagnosis of blood cells by medical workers, and is also convenient for deployment on some portable medical devices.

Author Contribution

Zeyu Liu: Methodology, software, writing original draft. Dan Yuan: Supervision, project administration, review and editing.Guohun Zhu: Supervison, investigation,validation, review and editing.All authors reviewed the manuscript.

E. K. Sackmann, A. L. Fulton, and D. J. Beebe, The present and future role of microfluidics in biomedical research, Nature, vol. 507, no. 7491, 2014, pp. 181-189.
Yousif T Y E, Impact of abnormal leukocyte count in the pathophysiology of sickle cell anemia, Journal of Blood Medicine, 2022: 673-679.
İnce and F. Demet, et al, The comparison of automated urine analyzers with manual microscopic examination for urinalysis automated urine analyzers and manual urinalysis, Practical laboratory medicine, 2016, pp. 14-20.
X. Wang and J. Liu, Tomato anomalies detection in greenhouse scenarios based on YOLO-Dense, Frontiers in Plant Science, 2021, 12: 634103.
Y. LeCun, L. Bottou, Y. Bengio, and P. Haner, Gradient-Based Learning Applied to Document Recognition, Proceedings of the IEEE, 1998, 86: 2278-2324.
K. Alex, I. Sutskever, and G. E. Hinton, Imagenet classification with deep convolutional neural networks, Advances in neural information processing systems 25, 2012.
Shafiq M, Gu Z, Deep residual learning for image recognition: A survey, Applied Sciences, 2022, 12(18): 8972.
Cong X, Li S, Chen F, et al., A Review of YOLO Object Detection Algorithms based on Deep Learning, Frontiers in Computing and Intelligent Systems, 2023, 4(2): 17-20.
Tian Zhong, Research on Blood Cell Recognition and Counting Based on Improved YOLO v7, Advances in Applied Mathematics, 2023, 12(3), 1083-1089.
J. Terven, and D. Cordova-Esparza, A comprehensive review of YOLO: From YOLOv1 to YOLOv8 and beyond, arXiv, 2023, 2304.00501.
D. Ouyang et al, Efficient Multi-Scale Attention Module with Cross-Spatial Learning, ICASSP 2023, Rhodes Island, Greece, 2023, pp. 1-5.
C. Li, A. Zhou, and A. Yao, Omni-dimensional dynamic convolution, arXiv, 2022, 2209.07947.
Hsu D, Kakade S M, Zhang T, Robust matrix decomposition with sparse corruptions, IEEE Transactions on Information Theory, 2011, 57(11): 7221-7234.
Kutlu H, Avci E, Özyurt F, White blood cells detection and classification based on regional convolutional neural networks, Medical hypotheses, 2020, 135: 109472.
Flach P, Kull M, Precision-recall-gain curves: PR analysis done right, Advances in neural information processing systems, 2015, 28.
Roy A M, Bhaduri J, A deep learning enabled multi-class plant disease detection model based on computer vision, Ai, 2021, 2(3): 413-428.
Lee Y, Hwang J, Lee S, et al. An energy and GPU-computation efficient backbone network for real-time object detection, Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. 2019: 0-0.
Wu P, PyTorch 2.0: The Journey to Bringing Compiler Technologies to the Core of PyTorch (Keynote), Proceedings of the 21st ACM/IEEE International Symposium on Code Generation and Optimization. 2023: 1-1.
Oakden T, Kavakli M, Performance Analysis of RTX Architecture in Virtual Production and Graphics Processing, 2022 IEEE 42nd International Conference on Distributed Computing Systems Workshops (ICDCSW). IEEE, 2022: 215-220.
Bolton T, Bass J, Gaber T, et al., Comparing Object Recognition Models and Studying Hyperparameter Selection for the Detection of Bolts, International Conference on Applications of Natural Language to Information Systems. Cham: Springer Nature Switzerland, 2023: 186-200.
Wang Q, Bi S, Sun M, et al., Deep learning approach to peripheral leukocyte recognition, PloS one, 2019, 14(6): e0218808.
Shah R, Shastri J, Bohara M H, et al., Detection of different types of blood cells: A comparative analysis, 2022 IEEE international conference on distributed computing and electrical circuits and electronics (ICDCECE). IEEE, 2022: 1-5.
Yang S, Fang B, Tang W, et al., Faster R-CNN based microscopic cell detection, 2017 international conference on security, pattern analysis, and cybernetics (SPAC). IEEE, 2017: 345-350.
Wen T, Wu H, Du Y, et al., Faster R-CNN with improved anchor box for cell recognition, Mathematical Biosciences and Engineering, 2020, 17(6): 7772-7786.
Zhang Z, Deng Z, Wu Z, et al., An Improved EIoU-Yolov5 Algorithm for Blood Cell Detection and Counting, 2022 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI). IEEE, 2022: 989-993.
Nayer M M A, Rahaman K M A, Hossen M N, DCBC DeepL: Detection and Counting of Blood Cells Employing Deep Learning and YOLOv5 Model, Artificial Intelligence and Data Science: First International Conference, ICAIDS 2021, Hyderabad, India, December 17–18, 2021, Revised Selected Papers. Springer Nature, 2022: 203.
Luong D T, Anh D D, Thang T X, et al., Distinguish normal white blood cells from leukemia cells by detection, classification, and counting blood cells using YOLOv5, 2022 7th National Scientific Conference on Applying New Technology in Green Buildings (ATiGB). IEEE, 2022, 156-160.
Chen Y, Zhao Z, Yu Y, et al., Understanding IFRA for Detecting Synchronous Machine Winding Short Circuit Faults Based on Image Classification and Smooth Grad-CAM++, IEEE Sensors Journal, 2022, 23(3): 2422-2432.
Menikdiwela M, Nguyen C, Li H, et al., CNN-based small object detection and visualization with feature activation mapping, 2017 international conference on image and vision computing New Zealand (IVCNZ). IEEE, 2017: 1-5.
Ma W, Wu Y, Cen F, et al., Mdfn: Multi-scale deep feature learning network for object detection, Pattern Recognition, 2020, 100: 107149.
Georgatzakou H T, Antonelou M H, Papassideri I S, et al. Red blood cell abnormalities and the pathogenesis of anemia in end‐stage renal disease, PROTEOMICS–Clinical Applications, 2016, 10(8): 778-790.
Krishnevskaya E, Molero M, Ancochea Á, et al. New-Generation Ektacytometry Study of Red Blood Cells in Different Hemoglobinopathies and Thalassemia, Thalassemia Reports, 2023, 13(1): 70-76.
Peng S, Li W, Ke W. Association between red blood cell distribution width and all-cause mortality in unselected critically ill patients: analysis of the mimic-iii database, Frontiers in Medicine, 2023, 10: 1152058.
Bhuiyan M, Islam M S. A new ensemble learning approach to detect malaria from microscopic red blood cell images, Sensors International, 2023, 4: 100209.
Alomar K, Aysel H I, Cai X. Data augmentation in classification and segmentation: A survey and new strategies, Journal of Imaging, 2023, 9(2): 46.
Hao X, Zhu Y, Appalaraju S, et al. Mixgen: A new multi-modal data augmentation, Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2023: 379-389.
Wu Y, Han Q, Jin Q, et al, LCA-YOLOv8-Seg: An Improved Lightweight YOLOv8-Seg for Real-Time Pixel-Level Crack Detection of Dams and Bridges, Applied Sciences, 2023, 13(19): 10583.
Du Y, Liu X, Yi Y, et al., Optimizing Road Safety: Advancements in Lightweight YOLOv8 Models and GhostC2f Design for Real-Time Distracted Driving Detection, Sensors, 2023, 23(21): 8844

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Research on Blood Cell Detection and Counting Based on YOLO-BC Algorithm

Status:

Version 1

Abstract

Figures

1 Introduction

2 Related Works

3 Methods

3.1. The design of YOLO-BC detection pipeline

3.2. YOLO-BC algorithm

3.3. Efficient Multi-Scale Attention

3.4. Omni-Dimensional Dynamic Convolution

3.5. Dataset

3.6. Statistics

4 Experiments and results

4.1. Experimental Environment

4.2. Training

4.3. Ablation Environment

4.4. Comparison experiment

4.5. Visualization

5 Conclusions

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1