Cell image detection method based on improved YOLOv8

Abstract

The invention discloses a cell image detection method based on improved YOLOv8, which aims at the problems of low detection efficiency and accuracy of the existing cell micronucleus detection technology, and comprises the following steps: firstly, acquiring and preprocessing a cell image data set, and then, improving a YOLOv8 network architecture and parameters to obtain a cell micronucleus detection result. A C3STR module is introduced, an SPPF structure is replaced, a GAM Attention module is inserted, and a Wise-IoU frame loss function is used, so that an image detection model is constructed, the feature extraction capability of the model and the recognition capability of a small target feature map are optimized, and the granularity of a detection image of the detection model is improved; then, training optimization is conducted on the model through a test set, further verification is conducted through combination of a verification set and a reliability verification mechanism, a target image detection model is finally obtained, the model can respond to the test set to output an image detection result, and the accuracy and efficiency of micronucleus detection can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a cell image detection method based on improved YOLOv8. Background Art

[0002] Micronuclei refer to nucleosomes located in the cytoplasm that are independent of the main nucleus. They are stained the same as the main nucleus, and their diameter is generally 1 / 3-1 / 2 of the main nucleus. It is a manifestation of chromosome aberration in cells. It is a chromosome without centromere fragments or a whole chromosome lost due to spindle damage. In the late stage of cell division, it remains in the daughter cytoplasm to form one or several regular secondary nuclei. It can comprehensively display a variety of harmful endpoints such as DNA damage and repair, chromosome breakage or loss, and cell growth inhibition. The occurrence of various benign and malignant diseases is closely related to the damage of genetic material in the body's cells. By measuring micronuclei, it can reflect the degree of genetic damage caused by environmental mutagens and carcinogens to the body. It is of great significance in the risk screening and risk prediction of cancer.

[0003] Traditional micronucleus detection technology uses manual microscopy, which uses the human eye to identify and count cell micronuclei under a microscope. It takes a lot of time, and the quality of manual microscopy depends on the experience of the detector. Long-term work may lead to errors in the analysis results. With the rapid development of computer technology, a good foundation has been provided for micronucleus detection. Researchers use computer vision to detect cell micronuclei; image processing technology such as threshold segmentation and morphological analysis are used to detect micronucleus cells. This method has a good detection effect when the background is simple and the cells are evenly distributed. However, the detection results of samples with complex backgrounds and densely distributed cells are very poor, and the micronucleus detection task cannot be completed well.

[0004] In recent years, with the rapid development of machine learning in the medical field, new cell micronucleus visualization analysis and detection strategies based on deep learning have shown good application prospects, which can improve efficiency while reducing manpower waste. However, some existing micronucleus detection technologies still have serious problems of missed detection and false detection.

[0005] The above information disclosed in the Background section is only for enhancement of understanding of the background of the present application and therefore it may contain information that does not constitute the prior art that is already known to a person of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to address the problems existing in the existing cell micronucleus detection technology and propose a cell image detection method based on improved YOLOv8. First, a cell image data set is acquired and preprocessed. Then, an image detection model is constructed by improving the YOLOv8 network architecture and parameters to improve the granularity of the detection model's detection image. Then, the model is trained and optimized using a test set, and further verified by a verification set combined with a reliability verification mechanism. Finally, a target image detection model is obtained, which can respond to the test set to output image detection results, and can significantly improve the accuracy and efficiency of micronucleus detection.

[0007] In a first aspect, a technical solution provided in an embodiment of the present invention is a cell image detection method based on an improved YOLOv8, comprising the following steps: S1, obtain the cell image dataset and preprocess it to obtain the training set, validation set and test set; S2. Improve the network architecture and network parameters of YOLOv8 and build an image detection model based on the training set; S3. The optimized image detection model is verified through the verification set and in combination with the reliability verification mechanism to obtain a target image detection model; S4. The target image detection model outputs an image detection result in response to the test set.

[0008] Preferably, in S1, the step of obtaining a cell image dataset and preprocessing the dataset to obtain a training set, a validation set, and a test set comprises the following steps: The single-micronucleus dataset was constructed using normal monocytes and mono-micronucleus cells, and the double-micronucleus dataset was constructed using normal binucleus cells and binucleus cells; The cell image dataset was constructed using the single micronucleus dataset and the double micronucleus dataset; The elements in the cell image dataset are labeled and segmented according to the segmentation mechanism to obtain the training set, validation set and test set.

[0009] Preferably, the segmentation mechanism segments the cell image dataset to obtain a training set, a validation set, and a test set in a ratio of 8:1:1.

[0010] Preferably, S2 comprises: The C3STR module is introduced into the backbone feature extraction network of YOLOv8. Specifically, the SwinTransformer module is embedded in the C3 structure, and the feature map extracted by the backbone network is used to interact the information between local features through the Swin Tran structure.

[0011] Preferably, the S2 further comprises: In the backbone feature extraction network of YOLOv8, the SPPF structure of the last layer is replaced by the SPPFCSPC structure.

[0012] Preferably, the S2 further comprises: Before the feature fusion operation of each feature map of different sizes, the GAM Attention module is inserted to realize the recognition of small target feature maps.

[0013] Preferably, the S2 further comprises: Replace the CIoU border loss function of YOLOv8 with the Wise-IoU border loss function.

[0014] As a preferred option, the Wise-IoU border loss function L WIoU It is expressed as follows: Among them, x, y, w, and h are the horizontal coordinate, vertical coordinate, width, and height of the anchor box respectively; x gt ,y gt , w gt ,h gt The horizontal coordinate, vertical coordinate, width and height of the target box respectively; W i is the width of the overlap between the target box and the anchor box, H ii is the height of the overlapping part of the target box and the prediction box; W g H is the width of the minimum bounding rectangle of the anchor box and the target box. g is the height of the minimum bounding rectangle of the anchor box and the target box; the symbol “*” represents Separated from the computational graph, when training a neural network, the role of the computational graph is to calculate the output of the forward propagation and the gradient of the back propagation.

[0015] Preferably, the evaluation index of the reliability verification mechanism includes one or more combinations of accuracy index, precision index, recall index, F1 index and mAP index.

[0016] Preferably, when one or more combinations of accuracy index, precision index, recall index, F1 index, and mAP index are used as the model reliability criterion, according to the bucket effect, the evaluation value of the index with the lowest value is used as the target evaluation value, and the comparison result between the target evaluation value and the set threshold is used as the model reliability judgment result; When the accuracy index, precision index, recall index, F1 index and mAP index are used as the model reliability criteria, the target evaluation value of the model reliability is obtained by weighted summing up the evaluation values ​​of each indicator and the corresponding weight ratio; the comparison result between the target evaluation value and the set threshold is used as the model reliability judgment result.

[0017] The present invention has at least the following substantial beneficial effects: (1) By improving the network architecture and parameters of YOLOv8, introducing the C3STR module, replacing the SPPF structure, inserting the GAM Attention module, and using the Wise-IoU border loss function, the model's feature extraction ability and the ability to recognize small target feature maps are optimized, thereby improving the accuracy and efficiency of micronucleus detection; (2) By constructing a cell image dataset containing normal monocytes, mononuclear cells, normal binuclear cells, and binuclear cells, and preprocessing and segmenting the dataset, the model can learn more diverse cell features during training, thereby enhancing the generalization ability and robustness of the model; (3) By introducing a reliability verification mechanism and using a combination of accuracy, precision, recall, F1, mAP, and other indicators as the evaluation criteria for model reliability, the performance of the model can be evaluated more comprehensively and accurately, ensuring that the obtained target image detection model has a reliable detection effect in practical applications.

[0018] The above invention content is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Other features, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings. The drawings are only for the purpose of illustrating preferred embodiments and are not to be considered as limiting the present invention. Also, the same reference symbols are used throughout the drawings to represent the same parts.

[0020] Figure 1 This is a flow chart of the cell image detection method based on improved YOLOv8 of the present invention.

[0021] Figure 2 This is a diagram of the improved YOLOv8 model structure in an embodiment of the present invention.

[0022] Figure 3 It is a structural diagram of SPPFCSPC in an embodiment of the present invention.

[0023] Figure 4 It is the GAM Attention mechanism module introduced in the YOLOv8 feature fusion part in the embodiment of the present invention.

[0024] Figure 5 This is a structural diagram of the spatial attention submodule of the GAM Attention mechanism in an embodiment of the present invention.

[0025] Figure 6 This is a structural diagram of the channel attention submodule of the GAM Attention mechanism in an embodiment of the present invention.

[0026] Figure 7 Schematic diagram of the WIoU loss function in an embodiment of the present invention.

[0027] Figure 8 This is a structural diagram of C3STR in an embodiment of the present invention.

[0028] Fig. 9 It is a structural diagram of the Swin Transformer module in an embodiment of the present invention.

[0029] Fig.10 The figure shows the comparison of the detection results of the improved model and the original model in the embodiment of the present invention.

[0030] Fig.11 Confusion matrix comparison between the improved model and the original model in the embodiment of the present invention DETAILED DESCRIPTION

[0031] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0032] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the operations (or steps) as sequential processes, many of the operations (or steps) therein can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but can also have additional steps not included in the drawings; the process can correspond to a method, function, procedure, subroutine, subprogram, etc. Embodiment 1:

[0033] like Figure 1As shown, the cell image detection method based on improved YOLOv8 includes the following steps: S1. Obtain a cell image dataset and preprocess it to obtain a training set, a validation set, and a test set.

[0034] Specifically, S1 includes the following steps: The single-micronucleus dataset was constructed using normal monocytes and mono-micronucleus cells, and the double-micronucleus dataset was constructed using normal binucleus cells and binucleus cells; The cell image dataset was constructed using the single micronucleus dataset and the double micronucleus dataset; The elements in the cell image dataset are labeled and segmented according to the segmentation mechanism to obtain the training set, validation set and test set.

[0035] Furthermore, the segmentation mechanism segments the cell image dataset to obtain a training set, a validation set, and a test set in a ratio of 8:1:1.

[0036] In this embodiment, the purpose of step S1 is to construct a cell image data set that is properly annotated and segmented so as to be used for training and evaluating a cell image recognition model. The constructed data set is essential for developing an algorithm that can accurately identify and classify cells (such as normal monocytes, mono-micronucleated cells, normal binucleated cells, and binucleated cells). Among them, images of normal monocytes and mono-micronucleated cells are collected. These images may come from cell samples under a microscope and are obtained by high-resolution imaging technology. Similarly, images of normal binucleated cells and binucleated cells are collected. The mono-micronucleated data set and the binucleated data set are merged to form a more comprehensive cell image data set. This data set contains four different types of cell images. Each image in the data set is annotated to indicate the type of cell in the image (normal mononucleated, mono-micronucleated, normal binucleated, binucleated). Annotation may involve manual labeling or the use of automated tools to assist. According to the ratio of 8:1:1, the annotated data set is divided into a training set, a validation set, and a test set. For example, if there are 1000 images, 800 are used for training, 100 are used for validation, and 100 are used for testing.

[0037] Through the above technical means, this embodiment includes four different types of cell images, and the data set is diverse, which helps the model learn to identify the characteristics of different cell types. By dividing the data set into a training set, a validation set, and a test set, it can be ensured that the model will not overfit during the training process, and at the same time can be effectively evaluated on unseen data. Using annotated data sets for training can improve the efficiency of the model in learning cell features, because the model can learn how to distinguish different types of cells directly from the annotation information.

[0038] As a specific example of this embodiment, used to further explain this embodiment, assume that there is an initial data set containing 1000 cell images, wherein each cell type (normal mononucleus, monomicronucleus, normal binucleus, binucleus) has 250 images. After segmentation at a ratio of 8:1:1, the distribution of images of each cell type in the training set, validation set, and test set is as follows: Training set: 200 images of each cell type, a total of 800 images. Validation set: 25 images of each cell type, a total of 100 images. Test set: 25 images of each cell type, a total of 100 images. In this way, it can be ensured that the model is exposed to representative samples of various cell types during training and evaluation, thereby improving its accuracy and reliability in practical applications.

[0039] like Figure 2 As shown in S2, the network architecture and network parameters of YOLOv8 are improved and the image detection model is constructed in combination with the training set.

[0040] Furthermore, if Figure 8 As shown in Figure 2, the C3STR module is introduced into the backbone feature extraction network of YOLOv8. Specifically, Fig. 9 As shown in the figure, the Swin Transformer module is embedded in the C3 structure, and the feature map extracted by the backbone network is used to interact the information between local features through the SwinTran structure; by introducing the C3STR module and replacing the SPPF structure, the model can extract richer and more representative features, which helps to improve the accuracy of detection.

[0041] like Figure 3 As shown in the figure, further, the SPPF structure of the last layer in the backbone feature extraction network of YOLOv8 is replaced by the SPPFCSPC structure; the SPPFCSPC structure is an improved spatial pyramid pooling structure that can more effectively capture multi-scale features and improve the robustness of feature extraction.

[0042] like Figure 4 As shown in Figure 1, before the feature fusion operation of each feature map of different sizes, the GAM Attention module is inserted to realize the recognition of small target feature maps. The structure of the GAM Attention module is mainly composed of the spatial attention mechanism and the channel attention mechanism, which is improved on the basis of CBAM; specifically, Figure 6 As shown in the figure, the channel attention mechanism submodule uses 3D permutation to retain the three-dimensional information, and then amplifies the cross-dimensional channel-space dependency through two layers of MLP. Figure 5As shown in the figure, in the spatial attention submodule, two convolution levels are used to fuse spatial information in order to focus on spatial information. The GAM Attention module combines spatial attention and channel attention, which can better focus on and weight important feature information. The feature fusion part of YOLOv8 introduces the GAM Attention mechanism, which is placed before the feature fusion operation at three scales. The performance of target detection is improved by improving the feature information of feature maps at three scales. GAM Attention has multiple levels of attention mechanisms, which can realize feature modeling at different spatial scales, and can effectively improve the detection performance in small target detection tasks. After inserting the GAM Attention module, the model's recognition ability for small targets (such as micronucleus cells) is enhanced, reducing missed detections and false detections.

[0043] Furthermore, if Figure 7 As shown in the figure, the CIoU border loss function of YOLOv8 is replaced with the Wise-IoU border loss function; using the Wise-IoU border loss function to replace the original CIoU loss function makes the model more accurate in locating cells and improves the accuracy of border prediction.

[0044] Specifically, the Wise-IoU border loss function L WIoU It is expressed as follows: Among them, x, y, w, and h are the horizontal coordinate, vertical coordinate, width, and height of the anchor box respectively; x gt ,y gt , w gt ,h gt The horizontal coordinate, vertical coordinate, width and height of the target box respectively; W i is the width of the overlap between the target box and the anchor box, H ii is the height of the overlapping part of the target box and the prediction box; W g H is the width of the minimum bounding rectangle of the anchor box and the target box. g is the height of the minimum bounding rectangle of the anchor box and the target box; the symbol “*” represents Separated from the computational graph, when training a neural network, the role of the computational graph is to calculate the output of the forward propagation and the gradient of the back propagation.

[0045] S3, verifying the optimized image detection model through the verification set and combining the reliability verification mechanism to obtain the target image detection model; Specifically, the evaluation indicators of the reliability verification mechanism include one or more combinations of accuracy indicator, precision indicator, recall rate indicator, F1 indicator, and mAP indicator.

[0046] Furthermore, when one or more combinations of accuracy, precision, recall, F1, and mAP are used as criteria for model reliability, according to the bucket effect, the evaluation value of the indicator with the lowest value is used as the target evaluation value, and the comparison result between the target evaluation value and the set threshold is used as the model reliability judgment result.

[0047] Specifically, the accuracy index is the ratio of the number of correct predictions to the total number of samples, and its formula is: Among them, TP represents the number of samples whose true label and predicted label are both positive; FP represents the number of samples whose true label is negative and predicted to be positive; TN represents the number of samples whose true label is negative and predicted to be positive; FN represents the number of samples whose true label is negative and predicted to be positive.

[0048] Specifically, the precision index formula is: Specifically, the recall rate indicator formula is: Specifically, the F1 index, also known as the F1 score, is a measure of the classification problem, and the formula is: Then calculate the average value of all categories, recorded as F1, the formula is: Specifically, the mAP (mean Average Precision) indicator is the average value of AP in each column. The calculation formula of AP is: Where n represents the number of monitoring points, Pinterp represents the value of the accuracy when the recall rate is r, and mAP can be calculated based on AP. The formula is:

[0049] S4. The target image detection model outputs image detection results in response to the test set.

[0050] In this embodiment, a validation set containing various scenes and targets has been prepared in step S1 to ensure that it is consistent with the training set and the test set in data distribution. Assume that the validation set contains 25 images of each cell type, for a total of 100 images. Different lighting conditions, background complexity, and target size are covered. The model is evaluated using multiple evaluation indicators such as accuracy, precision, recall, F1 index, and mAP.

[0051] It is understandable that when one or more combinations of accuracy, precision, recall, F1, and mAP are used as the model reliability criterion, according to the barrel effect, the evaluation value of the lowest indicator is used as the target evaluation value, and the comparison result between the target evaluation value and the set threshold is used as the model reliability evaluation result; for example: according to the barrel effect, the lowest indicator is selected as the key criterion for model reliability, assuming that the set threshold is: accuracy ≥ 90%, precision ≥ 85%, recall ≥ 80%, F1 ≥ 85%, and mAP ≥ 90%. Apply the optimized image detection model to the validation set and calculate the various evaluation indicators. Assume that the indicators obtained are: accuracy 92%, precision 87%, recall 78%, F1 84%, and mAP 91%. According to the barrel effect, the recall rate of 78% is the lowest indicator and is lower than the set threshold of 80%, so the model needs to be further optimized. Adjust the parameters or structure of the model according to the validation results to improve the recall rate. Repeat the validation process until all indicators reach or exceed the set threshold.

[0052] Next, prepare a test set: prepare a test set that is independent and identically distributed with the training set and validation set. Assume that the test set contains 25 images of each cell type, for a total of 100 images, also covering different scenes and targets. Apply the verified target image detection model to the test set. The model detects each image and outputs the detection results.

[0053] Through the methods or steps of the above embodiments, Fig.10 , Fig.11 It can be seen that the target image detection model performs better than the original model on the test set, with higher detection accuracy, precision, recall and mAP value. Through the evaluation of the reliability verification mechanism, the target image detection model reached or exceeded the set threshold on all the set evaluation indicators, indicating that it has high reliability. The model can also maintain good performance on unseen test sets, further verifying its generalization ability. The target image detection model can be applied to actual cell image detection tasks, providing strong support for biomedical research. Embodiment 2:

[0054] Most of the technical contents are the same as those of Example 1, except that Example 2 is different in that: the reliability verification mechanism of the model is different; specifically: when the accuracy index, precision index, recall rate index, F1 index and mAP index are used as the model reliability criteria, the target evaluation value of the model reliability is obtained by weighted summing up the evaluation values ​​of each indicator and the corresponding weight ratio; the comparison result of the target evaluation value and the set threshold is used as the model reliability judgment result.

[0055] It can be understood that when the accuracy index, precision index, recall rate index, F1 index and mAP index are used as the model reliability criteria, the target evaluation value of the model reliability is obtained by weighted summing up the evaluation values ​​of each indicator and the corresponding weight ratio; the comparison result between the target evaluation value and the set threshold is used as the model reliability judgment result. When evaluating the model performance, in order to better evaluate the improved trained model, the above evaluation indicators are used for comprehensive evaluation; the corresponding weight ratio is set according to the importance of each indicator: mAP is the most core indicator to measure the overall performance of the model, which combines precision and recall, and comprehensively evaluates the average detection accuracy of each category, which can more comprehensively reflect the overall performance of the model in the target detection task, and its weight is 0.6; precision represents the actual correct sample ratio among all samples predicted as positive, reflecting the quality of the model's positive sample recognition; recall represents the detection rate of real samples, and the two evaluation indicators can better show the model's detection of task targets, and their weights are 0.25 respectively; F1 score is an indicator that balances precision and recall; better controls the detection of task targets; its weight is 0.05; accuracy represents the overall correct rate of the model in all detections, but because the detection task includes position regression (bounding boxprediction) and classification, the accuracy rate alone cannot reflect the model performance well, and its weight is 0.05; accuracy rate = 90%, precision rate = 85%, recall rate = 80%, F1 index = 85%, mAP = 90%; according to the indicator evaluation value and the corresponding weight value, the reliability score of the model is: G = 0.6*0.9+0.15*0.85+0.15*0.8+0.05*0.9+0.05*0.85=0.875; when the set threshold is 0.8; G ≥ 0.8, it can be seen that the reliability of the model is qualified at this time; if G is lower than the set threshold of 80%; the model needs to be further optimized.

[0056] The specific implementation described above is a preferred implementation of the cell image detection method based on the improved YOLOv8 of the present invention, and is not intended to limit the specific implementation scope of the present invention. The scope of the present invention includes but is not limited to this specific implementation. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A cell image detection method based on improved YOLOv8, characterized by: The steps include: S1, obtain the cell image dataset and preprocess it to obtain the training set, validation set and test set; S2. Improve the network architecture and network parameters of YOLOv8 and build an image detection model based on the training set; S3, verifying the optimized image detection model through the verification set and combining the reliability verification mechanism to obtain the target image detection model; S4. The target image detection model outputs image detection results in response to the test set.

2. The cell image detection method based on improved YOLOv8 according to claim 1, characterized in that: In S1, the step of obtaining a cell image dataset and preprocessing the dataset to obtain a training set, a validation set, and a test set includes the following steps: The single-micronucleus dataset was constructed using normal monocytes and mono-micronucleus cells, and the double-micronucleus dataset was constructed using normal binucleus cells and binucleus cells; The cell image dataset was constructed using the single micronucleus dataset and the double micronucleus dataset; The elements in the cell image dataset are labeled and segmented according to the segmentation mechanism to obtain the training set, validation set and test set.

3. The cell image detection method based on improved YOLOv8 according to claim 2, characterized in that: The segmentation mechanism segments the cell image dataset to obtain a training set, a validation set, and a test set in a ratio of 8:1:

1.

4. The cell image detection method based on improved YOLOv8 according to claim 1, characterized in that: The S2 includes: The C3STR module is introduced into the backbone feature extraction network of YOLOv8. Specifically, the SwinTransformer module is embedded in the C3 structure, and the feature map extracted by the backbone network is used to interact the information between local features through the Swin Tran structure.

5. The cell image detection method based on improved YOLOv8 according to claim 4, characterized in that: The S2 further includes: In the backbone feature extraction network of YOLOv8, the SPPF structure of the last layer is replaced by the SPPFCSPC structure.

6. The cell image detection method based on improved YOLOv8 according to claim 4 or 5, characterized in that: The S2 further includes: Before the feature fusion operation of each feature map of different sizes, the GAM Attention module is inserted to realize the recognition of small target feature maps.

7. The cell image detection method based on improved YOLOv8 according to claim 4 or 5, characterized in that: The S2 further includes: Replace the CIoU border loss function of YOLOv8 with the Wise-IoU border loss function.

8. The cell image detection method based on improved YOLOv8 according to claim 7, characterized in that: Wise-IoU border loss function L WIoU It is expressed as follows: Among them, x, y, w, and h are the horizontal coordinate, vertical coordinate, width, and height of the anchor box respectively; x gt ,y gt , w gt ,h gt The horizontal coordinate, vertical coordinate, width and height of the target box respectively; W i is the width of the overlap between the target box and the anchor box, H i is the height of the overlapping part of the target box and the prediction box; W g H is the width of the minimum bounding rectangle of the anchor box and the target box. g is the height of the minimum bounding rectangle of the anchor box and the target box; the symbol "*" represents Separated from the computational graph, when training a neural network, the role of the computational graph is to calculate the output of the forward propagation and the gradient of the back propagation.

9. The cell image detection method based on improved YOLOv8 according to claim 1, characterized in that: The evaluation indicators of the reliability verification mechanism include one or more combinations of accuracy indicator, precision indicator, recall rate indicator, F1 indicator, and mAP indicator.

10. The cell image detection method based on improved YOLOv8 according to claim 9, characterized in that: When one or more combinations of accuracy, precision, recall, F1, and mAP are used as the model reliability criterion, according to the bucket effect, the evaluation value of the lowest indicator is used as the target evaluation value, and the comparison result between the target evaluation value and the set threshold is used as the model reliability judgment result; When the accuracy index, precision index, recall index, F1 index and mAP index are used as the model reliability criteria, the target evaluation value of the model reliability is obtained by weighted summing up the evaluation values ​​of each indicator and the corresponding weight ratio; the comparison result between the target evaluation value and the set threshold is used as the model reliability judgment result.

Images