CN116893162A - Rare anti-nuclear antibody karyotype detection method based on YOLO and attention neural network - Google Patents

Abstract

The present invention proposes a rare anti-nuclear antibody karyotype detection method based on YOLO and attention neural network. The process is that after image normalization processing, it is passed into the neural network for detection, thereby obtaining the rare anti-nuclear antibody karyotype. detection. The advantage of this method is that it is a customized and effective improvement and upgrade of the traditional target detection algorithm YOLO. According to the characteristics of clinical rare anti-nuclear antibody images, an attention mechanism is introduced on the basis of the YOLO target detection model, so that the model can improve rare anti-nuclear antibody The weight of the key features of the antibody karyotype image in the entire feature map improves the network's feature extraction capabilities for rare anti-nuclear antibody karyotype images and has the advantage of being more accurate. The implementation of this method can improve the efficiency of medical laboratory personnel in reading antinuclear antibody fluorescence images, avoid subjective differences in visual interpretation by different readers, and at the same time improve the current clinical situation due to insufficient understanding of rare ANA karyotypes by medical laboratory personnel. Missing detections are common.

Description

Rare anti-nuclear antibody karyotype detection method based on YOLO and attention neural network

Technical field

The invention belongs to the field of biomedicine, and specifically relates to a research method for rare anti-nuclear antibody karyotype detection based on the combination of YOLO target detection model and attention neural network model.

Background technique

Anti-nuclear antibody (ANA) is a serological marker for autoimmune diseases. Since 2007, the European League Against Rheumatism/American College of Rheumatology (EULAR/ACR) has After ANA was included in the main classification criteria of systemic lupus erythematosus (SLE), ANA has become one of the most important laboratory tests in the diagnosis of autoimmune diseases. It is included in the classification standards of various AIDs and plays an important role in the diagnosis, classification and prognosis of AID.

According to the international consensus on antinuclear antibody pattern (ICAP), there are as many as 30 categories of ANA fluorescence patterns. However, in fact, the number distribution of each category is uneven, and most of them are concentrated in granular, uniform and The clinical significance of this type of karyotype is now relatively clear. However, due to the extremely small number of some rare karyotypes, their clinical significance is not yet clear, and their understanding is still insufficient. Currently, karyotypes with a probability of occurrence less than 1% in ANA karyotypes are defined as rare ANA karyotypes. These karyotypes can be divided into 3 groups and 9 categories in total: cell cycle-related ones include NuMA type, spindle fiber type, and CENP-F. Type, intermediate type, PCNA type, centriole type; those related to the nucleus include multinuclear dot type and nuclear membrane type; those related to the cytoplasm are mainly Golgi type. Current clinical problems include: 1) Lack of knowledge: Since this part of the karyotype is very rare in clinical practice, inspectors have insufficient knowledge of the rare ANA karyotype, and laboratory routines do not report this pattern, so detection is often missed. ; 2) More complex: clinically rare ANA karyotype interpretation requires comprehensive judgment based on the situation of cell mitosis and interphase, and the cell division process is divided into four stages: prophase, chromosome aggregation; metaphase, chromosomes are pulled toward the poles Ascending; anaphase, the chromosomes begin to separate; telophase, the two chromosomes are completely separated. Although the shapes of the nuclei at each stage are similar, they all have different morphological characteristics. There is a certain complexity in visual interpretation; 3) Clinical value: Some rare ANA karyotypes are helpful in the diagnosis of certain types of autoimmune diseases. For example, nuclear membrane type is strongly related to autoimmune liver disease. Therefore, detecting rare ANA karyotypes is of great clinical value.

At present, the most traditional indirect immunofluorescence method (indirect immunofluorescence, IIF) is still regarded as the "gold standard", "reference method" and "preferred method" for ANA detection. However, the experimental results rely on naked eye interpretation, which has the disadvantages of incomplete human visual capture, easy fatigue, and missed minor lesions. As the scope of clinical application of ANA testing expands, clinical needs increase day by day, and the existing work model becomes increasingly difficult to meet daily clinical needs. With the advancement of deep learning represented by convolutional neural networks (CNN) in recent years, it has unique advantages in medical image processing. CNN is built to imitate the visual perception mechanism of living things and can perform supervised learning and unsupervised learning. The sharing of convolution kernel parameters in the hidden layer and the sparseness of inter-layer connections enable the convolutional neural network to perform tasks with a small amount of calculation. Learning gridded features is very suitable for processing complex images. ANA immunofluorescence karyotype images are complex and highly complex due to individual differences, but they exactly fit the advantages of CNN. The identification and location of rare ANA karyotypes is the prerequisite for the detection of rare ANA karyotypes.

In recent years, object detection algorithms in deep learning have made great progress. So far, target detection algorithms mainly include two categories. One is a two-stage detection algorithm headed by R-CNN, including Fast R-CNN, and the other is a single-stage detection algorithm, such as SSD and YOLO. This type of algorithm has achieved promising results in target detection tasks in various fields. However, due to the sparse number of rare ANA karyotypes and unclear cell characteristics, it is necessary to find some extremely subtle and easily overlooked key feature points to make an interpretation. Such features may appear at various stages in the cell division cycle, and there is a certain degree of complexity; Moreover, ANA detection of stromal human laryngeal cancer epithelial cell lines (HEP2 cells) still has the problem of dense stacking of cells, and conventional target detection algorithms have low efficiency in extracting such features.

The attention mechanism is a special structure embedded in deep learning models in recent years. It is used to learn and calculate the contribution of input data to output data, that is, to weight the input signal. This is very useful for highlighting the key feature points of rare ANA karyotypes. helpful. In the CNN architecture, in order to capture a large enough receptive field and capture semantic contextual information, the feature map adopts a gradual down-sampling method, using coarse spatial grid-level features to identify the location of the target object, and within the entire image range. To model the relationship between target objects, the basic function of the attention mechanism is to weight different areas of the image, and the most relevant parts will be assigned the greatest weight. These modules are trainable and applied to every part of the image, ensuring progressive weight learning and increased attention to critical areas.

Based on this, we considered introducing an attention mechanism based on the YOLO architecture to further improve the network's feature extraction capabilities for more complex information such as cell division periods. Therefore, for the specific task of detecting rare ANA karyotype key feature points, it is important to customize and effectively improve and upgrade the existing general target detection algorithm to improve the accuracy of the model. This is very important for the detection of rare ANA karyotypes. .

Problems with the existing technology include: 1. Clinically, inspectors have insufficient understanding of rare ANA karyotypes and often miss detections. Laboratories do not routinely report this pattern, which requires a lot of work in the model training phase. amount to collect training sets; 2. The number of rare ANA karyotypes is sparse, and the cell characteristics are not obvious. It is necessary to find some extremely subtle and easily overlooked key feature points to make an interpretation. For the extraction of such key features, the conventional target detection algorithm YOLO The capabilities are limited and existing algorithms need to be upgraded and improved.

Contents of the invention

The purpose of this invention is to solve the problem of missed detection caused by current clinical laboratory personnel's insufficient understanding of rare ANA karyotypes, and the problem that the existing target detection algorithm cannot complete target detection. It designs a method introduced on the basis of the YOLO architecture. Attention neural network method is used to detect rare ANA karyotypes.

The technical solution of the present invention has the following steps:

Step 1: Construct a rare ANA karyotype immunofluorescence image target detection data set, and label the data set;

Step 2: Data set preprocessing and partitioning;

Step 3: Use the deep convolutional neural network in the YOLO v5 basic network as a basis to extract features from the input image and build a YOLO target detection model; on top of this, use the effective feature layer and upsampling extracted from the Backbone backbone network of the basic network The final result is embedded with the attention mechanism, and a model based on YOLO and attention neural network is constructed;

Step 4: Set the training parameters, train the constructed model based on YOLO and attention neural network, obtain the best parameter model, and use the trained model to process the ANA immunofluorescence image data, and then obtain the target detection results of the image to be tested ;

Step 5: Clinical deployment of the trained rare ANA karyotype detection model.

In the present invention, the step 1 is to construct a rare ANA karyotype target detection data set, specifically:

Collect samples that were tested for anti-nuclear antibodies at the Laboratory Department of Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine from April 2010 to December 2022, and the test results showed a rare ANA karyotype. Rare ANA karyotypes refer to karyotypes with a probability of occurrence less than 1% in antinuclear antibody karyotypes, which can be divided into 3 groups and 9 categories in total: 1) Cell cycle related, including NuMA type, spindle fiber type, CENP-F type , intermediate type, PCNA type, centriole type; 2) nucleus, including multinuclear dot type, nuclear membrane type; 3) cytoplasm, mainly Golgi type. The rare ANA karyotype images will be analyzed in accordance with the 2021 International Consensus on Antinuclear Antibody Pattern (ICAP) fluorescence karyotype classification standards by senior laboratory technicians who have been engaged in ANA fluorescence reading for more than 10 years and have deputy chief technician or above. Label annotation of category and location information. A total of 8465 rare ANA karyotype images were collected. The number of each karyotype is shown in Table 1. The database is owned by Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine.

After obtaining the data set, use the Labelimg annotation tool to label the image, mark the category and location information of the target in the image, and save it in the VOC data set format. Then it was randomly divided into training set, verification set and test set in a ratio of 8:1:1. The specific numbers are shown in Table 1.

Table 1 Composition of rare ANA karyotype target detection data set

In the present invention, the step 2 is data set preprocessing and division, specifically:

Before model training, data expansion of rare ANA karyotype image data is required to avoid a decrease in model detection accuracy due to insufficient sampling information. Data amplification preprocessing of images includes random cropping, horizontal flipping, vertical flipping, random rotation and changing the attributes of the image, where the attributes of the image include brightness, contrast, saturation or hue. The images are then normalized, and the images and corresponding annotation files are divided into training sets, verification sets and test sets according to a division ratio of 7:2:1.

In the present invention, step 3 is to use the deep convolutional neural network in the YOLO v5 basic network as the backbone network to extract features from the input image and build a YOLO target detection model, specifically:

The feature extraction network constructed with YOLO v5 as the main body includes the Input input end, Backbone backbone network, Neck neck network and Head output end; the Neck neck network part is embedded with the attention mechanism, and the attention mechanism includes both channel and spatial attention. The independent sub-module is mainly aimed at the effective feature layer extracted by the Backbone backbone network and the results after upsampling, thereby improving the ability to extract key features of rare antinuclear antibody karyotype images. The specific implementation steps are as follows:

4.1Input input terminal: Using Mosaic data enhancement, the 4 pictures are spliced according to random scaling, random cropping and random arrangement;

4.2 Backbone backbone network: First set the depth and width of the model; then use CSPDarknet53 as the architecture, which mainly includes Focus and Bottleneck structures; Focus is responsible for downsampling, and the Bottleneck structure includes CSP and SPP modules. The SPP module pair is extracted from the Bottleneck layer. After compressing the number of channels of feature maps of different scales, the multi-scale features are fused; the Bottleneck structure is the most basic structure of the Backbone backbone network. Its main component structure is the residual block, which combines the transmitted information through summation before continuing. pass downward;

4.3Neck neck network: Insert the FPN_PAN structure between the Backbone and the final Head output layer to form the basic Neck neck network; the horizontal axis of the FPN feature pyramid structure is regarded as the scale axis, and the scale-invariant feature variables are extracted, and then each Each pyramid feature map is evenly adjusted to the set high-resolution feature pyramid map, and finally the high-resolution feature pyramid map is connected to the extracted scale-invariant feature variables, which are used to detect rare ANA karyotypes at the output end of the YOLO model Head. feature area; due to the limited ability of the YOLO v5 basic Neck network to extract features of rare ANA karyotypes, in order to further improve the network's feature extraction capabilities for more complex information such as cell division stages, based on the original Neck network Embedding the attention mechanism, the attention mechanism includes two independent sub-modules of channel and spatial attention, which process and transmit the effective feature layer extracted by the backbone network and the upsampling results respectively. The YOLO v5 backbone network is embedded in the attention After the module, the network model structure is shown in Figure 2. For the attention module, the detailed implementation steps are as follows:

4.3.1 The principle and specific implementation process of the channel attention module is: take the input feature map F (H×W×C) (W and H are the width and height of the feature map respectively, and C is the number of channels of the feature map), After global maximum pooling and global average pooling based on width and height respectively, two 1×1×C feature maps are obtained, and then they are sent to a two-layer multi-layer perceptron (MLP) neural network respectively, and then The features output by MLP are added, and then passed through the sigmoid activation function to finally obtain the channel weight coefficient Mc (Formula 1). The weight coefficient Mc is processed and converted into a one-dimensional channel attention map A _T , and then the input feature map F and AT are multiplied at the pixel level (to represents), the channel-directed salient feature map F _T is obtained, and the calculation formula is as follows (Formula 3);

4.3.2 The principle and specific implementation process of the spatial attention module is: pass the feature map F through the maximum pooling layer and the average pooling layer to obtain two H×W×1 channel descriptions, and convert these The two descriptions are spliced together according to the channel, and then go through a 7×7 convolution layer and Sigmoid activation to obtain the spatial weight coefficient Ms (Formula 2);

4.3.3 Obtain the feature map with attention weight: Convert the spatial weight coefficient Ms obtained in the above steps into a one-dimensional spatial attention map A _K , and perform pixel-level comparison between the channel-directed salient feature map F _T and A _K. After multiplication and sequential merging, the feature map F _R with attention weight output is obtained (Formula 4);

4.4Head output: Replace the original loss function GIOU_Loss with SIoU based on the original network, and change the NMS of the prediction filter box to DIOU_nns to solve the regression convergence problem of the prediction box at different positions inside the real box due to dense stacking of image cells. .

In the present invention, the step 4 is to set training parameters, train the constructed model based on YOLO and attention neural network, and use the trained model to process the immunofluorescence image data, and then obtain the detection result of the image target to be tested, Specifically:

Read the rare ANA karyotype target detection data set, set the training parameters and start training, and get a curve of the loss value changing with training time. When the loss value converges, test the model. When the loss value does not converge, adjust the model parameters. Until the model converges, see Figure 3A and B, the model trained at this time is used as the final model for clinical deployment;

Since the ANA karyotype image uses the human laryngeal cancer epithelial cell line (HEP2 cells) as the detection matrix, the cells have the problem of dense stacking of cells. Therefore, DIOU_nns is used to optimize the area between the prediction box and the real box for convergence, so that NMS can be obtained The results are more reasonable and effective.

In the present invention, the step 5 is the model deployment of the rare anti-nuclear antibody karyotype attention neural network, including 5.1) model detection and 5.2 clinical visualization deployment:

The 5.1 model detection includes:

5.1.1: Read the image to be tested and perform data preprocessing;

5.1.2: Normalize the data of the target image to be detected;

5.1.3: Feed the normalized data into the model trained in step 4 above;

5.1.4: Perform model detection on the image to be tested to obtain the prediction results and prediction probability.

The 5.2 clinical visualization deployment is specifically:

Compile the trained model on the Windows platform through C++, configure the corresponding deep learning environment, visualize the model prediction results by compiling CMake, and then deploy the model on the fluorescence microscope CCD camera. The model can read clinical image pairs to be tested. It performs tests until it obtains predicted results and probabilities.

In view of the situation of missed detection caused by clinical laboratory personnel's insufficient understanding of rare ANA karyotypes, this invention innovatively proposes for the first time a method of introducing attention neural networks based on the YOLO architecture to achieve detection of rare ANA karyotypes. the purpose of detection.

In the method of the present invention, in view of the characteristics of rare ANA karyotype cells whose characteristics are not obvious and some extremely subtle and easily overlooked key feature points need to be found to make interpretations, the introduced attention neural network model feature extraction can combine low-level (spatial) and high-level Characteristics of (channel) networks.

Among them, low-level (spatial) information: low-level features after multiple downsampling. It can provide contextual semantic information of the target feature area in the entire image. This helps to discover the spatial location information of suspected rare ANA karyotypes throughout the image.

Among them, high-level (channel) information: high-resolution information directly transferred from the encoder to the decoder of the same height through the concatenate operation. The vectors output by low-level and high-level network features are fed to ReLU and Sigmoid activation, and the resulting attention weights provide support for capturing more refined features.

Compared with existing target detection technology, the present invention has the following technical effects:

(1) The present invention introduces an attention mechanism based on the YOLO architecture, and the jointly constructed neural network can efficiently extract the extremely subtle and easily overlooked key feature points of rare ANA karyotype cells, taking advantage of multiple extraction features, so that It can better adapt to complex and changeable clinical situations caused by individual patient differences;

(2) The YOLO framework loss function of the present invention adopts SIoU, which fully considers the spatial direction factors between the real bounding box and the predicted detection frame, which can make the target regression frame more stable and reduce the occurrence of missed detections and false detections. The NMS of the prediction filter box is changed to DIOU_nns to solve the regression convergence problem when the prediction box is at different positions inside the real box due to dense stacking of image cells.

(3) On the premise of not reducing the detection speed as much as possible, the present invention effectively improves the detection accuracy of rare ANA karyotypes and reduces the probability of false detection and missed detection. In clinical practice, due to insufficient understanding of rare ANA karyotypes by clinical laboratory personnel, This makes a strong contribution to the situation of false detection and missed detection.

The beneficial effects of the invention include: the invention belongs to the field of medical-related artificial intelligence technology and biomedicine. The invention can effectively reduce the workload of laboratory doctors, improve work efficiency, and reduce the phenomenon of false detection and missed detection in clinical practice. In the present invention, there is no need to match a large number of templates, the processing speed is fast, the accuracy is high, and it meets the requirements of laboratory doctors in actual work.

Description of the drawings

Figure 1A and B are image descriptions and clinical significance of rare ANA karyotypes. (Image description and clinical significance of the rare ANA karyotype (AC-26/25/14/27) is shown in Figure 1A), Figure 1B is AC-13/6/24/22.

Figure 2 shows the model network results and result output effects of the present invention.

Figure 3 is a model training diagram of the present invention, Figure 3A is a loss curve diagram, and Figure 3B is a model accuracy curve.

Figure 4 is a flow chart of the method of the present invention.

Figure 5 is an example of the image to be tested and the result prediction output of the model of the present invention. (Example of model test image and result prediction output

(Left side: image to be tested; right side: prediction output) as shown in Figure 5).

Detailed ways

The process and experimental methods for implementing the present invention will be further described in detail with reference to the following specific examples and drawings. Figure 3 is a flow chart of the method of the present invention, which can be divided into the following steps:

1. Data preprocessing: Preprocess the ANA cell slide image to be detected to obtain a normalized image.

2. Data expansion, random translation transformation and angle conversion of the image.

3. Model loading and prediction. After successfully loading the model, call the model prediction function to predict the image to be tested.

4. Result output. The output result of the network is a tensor with the dimension: K*K*[B*(5*N)], where K represents the number of grids divided into each picture to be tested, and B represents each The number of detection frames corresponding to the grid, C represents the coordinate information (x, y, h, w) and confidence of each detection frame, and N is the number of categories of the target karyotype. Among them, the coordinate information of the detection frame (x, y, h, w), x represents the x coordinate information of the detection frame; y represents the y coordinate information of the detection frame; h represents the height information of the detection frame; w represents the width information of the detection frame. The specific meaning of the expression K*K*[B*(5*N)] is: B is used to lock specific rare karyotype features falling at the center point of the grid, and each detection box corresponds to a score, Represents whether there is an object in the detection frame and the confidence level, where the confidence level is the predicted probability that the detection frame contains rare karyotype features, that is, the detection frame, and the score is the certainty of rare karyotype features in the grid. When judging the accuracy of model detection, the present invention uses SloU (Scylla Intersection over Union) to evaluate model performance. The calculation of SIoU is detailed in formula (4).

Among them, Δ represents distance cost and angle cost, specifically the distance between the center point of the predicted bounding box and the center point of the real box, and the angle formed by the connecting line and the vertical line of the height of the two points; Ω represents the shape loss. (Shape cost), specifically the similarity of the shape of the predicted bounding box and the real box; IoU is specifically the IoU value between the predicted bounding box and the real box. The degree of certainty is represented by SIoU _{Predicted/True} , which represents the IoU value between the predicted box and the real box.

Figure 2 is a schematic diagram of the results of using the method of the present invention to detect rare ANA karyotype images detected in clinical practice. Figure 5 is an example rendering of the prediction of the rare ANA karyotype discovered. It can be seen that the method of the present invention can effectively detect rare ANA karyotypes, reduce the missed detection and misdiagnosis rate of laboratory doctors, and improve clinical work efficiency.

Claims (6)

1. A rare anti-nuclear antibody karyotype detection method based on YOLO and attention neural network, which is characterized in that the method includes: Step 1: Construct a rare anti-nuclear antibody (ANA) karyotype immunofluorescence image target detection data set, and label the data set; Step 2: Data set preprocessing and partitioning; Step 3: Use YOLO v5 as the basic network to build a target detection model, and embed the attention mechanism in the effective feature layer extracted from the Backbone backbone network part of the basic network structure and the upsampled results, thereby building a model based on YOLO and attention neural network model; Step 4: Use the target detection data set to train the built YOLO and attention neural network model to obtain the best parameter model, and use the unlabeled ANA karyotype immunofluorescence images collected clinically to evaluate the trained model. Test, and then obtain the target detection results of the image to be tested; Step 5: Clinically deploy the trained YOLO and attention neural network model. 2. The rare anti-nuclear antibody karyotype detection method of YOLO and attention neural network as claimed in claim 1, characterized in that the step (1) is specifically: Collect immunofluorescence images of rare ANA karyotypes in clinical testing to create a data set; use the image annotation tool Labelimg to label the characteristic areas of rare ANA karyotypes with category and location information, and classify rare ANA karyotypes as NuMA, CENP-F, Kinesin-5, Midbody, PCNA, Multiple nuclear dots, Centriole, Golgi, Nuclearenvelope, a total of nine types of tags, generate XML format annotation files corresponding to various types of nuclear images, and build a target detection data set. 3. The rare anti-nuclear antibody karyotype detection method of YOLO and attention neural network as claimed in claim 1, characterized in that the step (2) is specifically: Data amplification preprocessing of images includes random cropping, horizontal flipping, vertical flipping, random rotation and changing the attributes of the image, where the attributes of the image include brightness, contrast, saturation or hue, and then normalizing the image. And the images and corresponding annotation files are divided into training sets, verification sets and test sets according to the division ratio of 7:2:1. 4. The rare anti-nuclear antibody karyotype detection method of YOLO and attention neural network as claimed in claim 1, characterized in that the step (3) is specifically: The feature extraction network constructed with YOLO v5 as the main body includes the Input input end, Backbone backbone network, Neck neck network and Head output end; the Neck neck network part is embedded with the attention mechanism, and the attention mechanism includes both channel and spatial attention. The independent sub-module is mainly aimed at the effective feature layer extracted by the Backbone backbone network and the results after upsampling, thereby improving the ability to extract key features of rare antinuclear antibody karyotype images. The specific implementation steps are as follows: 4.1 Input terminal: Using Mosaic data enhancement, the 4 pictures are spliced according to random scaling, random cropping and random arrangement; 4.2 Backbone backbone network: First set the depth and width of the model; then use CSPDarknet53 as the architecture, which mainly includes Focus and Bottleneck structures; Focus is responsible for downsampling, and the Bottleneck structure includes CSP and SPP modules. The SPP module pair is extracted from the Bottleneck structure. After compressing the number of channels of feature maps of different scales, the multi-scale features are fused; the Bottleneck structure is the most basic structure of the Backbone backbone network. Its main component structure is the residual block, which combines the transmitted information through summation before continuing. pass downward; 4.3Neck neck network: Insert the FPN_PAN structure between the Backbone and the final Head output layer to form the basic Neck neck network; the horizontal axis of the FPN feature pyramid structure is regarded as the scale axis, and the scale-invariant feature variables are extracted, and then each Each pyramid feature map is evenly adjusted to the set high-resolution feature pyramid map, and finally the high-resolution feature pyramid map is connected to the extracted scale-invariant feature variables, which are used to detect rare ANA karyotypes at the output end of the YOLO model Head. feature area; because the YOLO v5 basic Neck network has limited ability to extract rare ANA karyotype features, the attention mechanism is embedded on the original basis. The specific implementation steps are as follows: After the input feature map, that is, the corrected H×W×C, feature vector is processed through three layers of convolution, a basic feature map F of H×W×48 is obtained, where W and H are the width and height of the feature map respectively. , C is the number of channels of the feature map; then use the channel attention module to convert the input feature map F into a one-dimensional channel attention map A _T , and then multiply the input feature map F and A _T at the pixel level to obtain represents, the channel-directed salient feature map F _T is obtained, calculated as formula (1), and then the spatial attention module is used to convert the input feature map F _T into a one-dimensional spatial attention map A _K , and finally F _T and A _K are Perform pixel-level multiplication to obtain the output feature map F _R , calculated as formula (2); the channel attention feature map F _T and the spatial attention map A _K obtained in the above steps are sequentially merged to obtain features with attention weights picture; 4.4Head output: Replace the loss function with SIoU, which is used to correct the spatial direction factor between the real bounding box and the predicted detection frame, making the target regression frame more stable; the NMS of the prediction screening frame is changed to DIOU_nns to solve the problem of dense image cells. Model of regression convergence problem when predicted boxes appear at different positions inside the real box when stacked. 5. The rare anti-nuclear antibody karyotype detection method of YOLO and attention neural network as claimed in claim 1, characterized in that the step (4) is specifically: After setting the training parameters, start model training and obtain a curve of loss value changing with training time. When the loss value converges, test the model. When the loss value does not converge, adjust the model parameters until the model has the highest accuracy and is well trained. model; SloU is used to evaluate model performance. The calculation of SIoU is detailed in formula (3): Among them, Δ represents distance cost and angle cost, specifically the distance between the center point of the predicted bounding box and the center point of the real box, and the angle formed by the connecting line and the vertical line of the height of the two points; Ω represents the shape loss. (Shape cost), specifically the similarity between the shape of the predicted bounding box and the real box; IoU is specifically the IoU value between the predicted bounding box and the real box, and the confidence is represented by SIoU _{Predicted/True} , which represents the difference between the predicted box and the real box IoU value between; because the ANA karyotype image uses the human laryngeal cancer epithelial cell line as the detection matrix, the cells have the problem of dense stacking of cells, so DIOU_nns is used to optimize the area between the prediction box and the real box for convergence, so that NMS The results obtained are more reasonable and effective. 6. The rare anti-nuclear antibody karyotype detection method of YOLO and attention neural network as claimed in claim 1, characterized in that the step (5) is specifically: Compile the trained model on the Windows platform through C++, configure the corresponding deep learning environment, visualize the model prediction results by compiling CMake, and then deploy the model on the fluorescence microscope CCD camera. The model can read clinical image pairs to be tested. It performs detection until the predicted results and predicted probabilities are obtained.