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

Abstract

The application provides a rare anti-nuclear antibody nuclear type detection method based on YOLO and an attention neural network, and the flow is as follows: after image normalization processing, the image is transmitted into a neural network for detection, so that the rare anti-nuclear antibody nuclear type is detected. The advantage of the method is that the traditional target detection algorithm YOLO is customized, effectively improved and upgraded, the attention mechanism is introduced on the basis of the YOLO target detection model according to the characteristics of the rare anti-nuclear antibody image, the model can improve the weight of the key features of the rare anti-nuclear antibody nuclear type image in the whole feature map, improve the feature extraction capability of the network for the rare anti-nuclear antibody nuclear type image, and has the advantages of higher accuracy. The implementation of the method can improve the reading efficiency of the medical testing personnel for the anti-nuclear antibody fluorescence image, avoids the subjective difference of naked-eye judgment of different readers, and improves the current situation that the medical testing personnel are insufficient in understanding of the rare ANA nuclear type and the missed detection is relatively common.

Description

Technical Field

[0001] This invention belongs to the field of biomedicine, specifically relating to a research method for detecting rare antinuclear antibody karyotypes based on a combination of the YOLO target detection model and the attention neural network model. Background Technology

[0002] Antinuclear antibody (ANA), as a serological marker for autoimmune diseases, has become one of the most important laboratory tests in the diagnosis of autoimmune diseases since the European League Against Rheumatism / American College of Rheumatology (EULAR / ACR) included ANA in the main classification criteria for systemic lupus erythematosus (SLE) in 2007. It is incorporated into the classification criteria of various autoimmune diseases and plays an important role in their diagnosis, classification, and prognosis.

[0003] According to the international consensus on antinuclear antibody pattern (ICAP), there are as many as 30 existing ANA fluorescence patterns. However, the distribution of each pattern is uneven, with the vast majority concentrated in granular and homogeneous karyotypes, the clinical significance of which is relatively well understood. Some rare karyotypes, due to their extremely low number, have less clear clinical significance and require further understanding. Currently, karyotypes occurring in less than 1% of ANA karyotypes are defined as rare ANA karyotypes. These karyotypes can be divided into 3 groups with a total of 9 types: cell cycle-related types include NuMA, spindle fiber, CENP-F, intermediate, PCNA, and centriole types; nuclear-related types include multinucleated dot type and nuclear membrane type; and cytoplasmic-related types are mainly Golgi apparatus types. Current clinical problems include: 1) Limited awareness: Due to the rarity of this karyotype in clinical practice, laboratory personnel lack sufficient understanding of rare ANA karyotypes, and this pattern is not routinely reported in laboratories, often leading to missed detections; 2) Complexity: Clinical interpretation of rare ANA karyotypes requires comprehensive judgment based on the cell's mitotic and interphase states. Cell division is divided into four stages: prophase (chromosome aggregation), metaphase (chromosomes pulling to the poles), anaphase (chromosome separation), and telophase (two completely separated chromosomes). While the nuclei in each stage may resemble each other, they exhibit distinct morphological characteristics. Visual interpretation is complex; 3) Clinical value: Some rare ANA karyotypes are helpful in diagnosing certain autoimmune diseases, such as the nuclear membrane type, which is strongly correlated with autoimmune liver disease. Therefore, detecting rare ANA karyotypes is of significant clinical value.

[0004] Currently, the most traditional indirect immunofluorescence (IIF) method is still considered the "gold standard," "reference method," and "preferred method" for ANA detection. However, the results rely on visual interpretation, which has drawbacks such as incomplete human visual capture, susceptibility to fatigue, and the potential to miss minute lesions. As the clinical application of ANA detection expands and clinical demands increase, the existing working model is increasingly unable to meet daily clinical needs. With the advancements in deep learning in recent years, represented by convolutional neural networks (CNNs), they possess unique advantages in medical image processing. CNNs are constructed by mimicking biological visual perception mechanisms, enabling both supervised and unsupervised learning. The shared parameters of convolutional kernels within their hidden layers and the sparsity of inter-layer connections allow CNNs to learn gridded features with relatively low computational cost, making them ideal for processing complex images. ANA immunofluorescence karyotype images are highly complex due to individual differences, perfectly matching the advantages of CNNs. The identification and localization of rare ANA karyotypes is a prerequisite for the detection of rare ANA karyotypes.

[0005] In recent years, object detection algorithms in deep learning have made significant progress. To date, object detection algorithms mainly fall into two categories: two-stage detection algorithms, spearheaded by R-CNN (including Fast R-CNN), and single-stage detection algorithms, such as SSD and YOLO. These algorithms have achieved remarkable results in object detection tasks across various fields. However, due to the scarcity of rare ANA karyotypes and their indistinct cellular features, it is necessary to find extremely subtle and easily overlooked key feature points for interpretation. These features may appear at various stages of the cell division cycle, adding complexity. Furthermore, ANA detection of the human laryngeal carcinoma epithelial cell line (HEP2 cells) presents the problem of dense cell stacking, resulting in low efficiency of conventional object detection algorithms in extracting such features.

[0006] Attention mechanisms are a special structure embedded in deep learning models in recent years. They are used to learn and calculate the contribution of input data to output data, i.e., weighting the input signal. This is very helpful in highlighting key feature points of rare ANA karyotypes. In CNN architectures, to capture a sufficiently large receptive field and semantic contextual information, feature maps are progressively downsampled, using coarse spatial grid-level features to identify the location of target objects and model the relationships between target objects across the entire image. The basic function of the attention mechanism is to weight different regions of the image, assigning the largest weight to the most relevant parts. These modules are trainable and applied to each part of the image, ensuring progressive weight learning and increasing attention to key regions.

[0007] Based on this, we considered introducing an attention mechanism into the YOLO architecture to further enhance the network's ability to extract features from more complex information such as cell division phase. Therefore, for the specific task of rare ANA karyotype feature point detection, we customized and effectively improved existing general object detection algorithms to enhance model accuracy, which is crucial for achieving rare ANA karyotype detection.

[0008] The existing technology has the following problems: 1. In clinical practice, laboratory personnel lack sufficient knowledge of rare ANA karyotypes, often resulting in missed detections. Laboratories do not routinely report this pattern, and a significant amount of work is required to collect training sets during the model training phase; 2. Rare ANA karyotypes are few in number and have indistinct cellular characteristics, requiring the identification of some extremely subtle and easily overlooked key feature points to make a judgment. Conventional target detection algorithms such as YOLO have limited capabilities in extracting such key features, and existing algorithms need to be upgraded and improved. Summary of the Invention

[0009] The purpose of this invention is to address the issue of missed detection due to insufficient understanding of rare ANA karyotypes among clinical laboratory personnel, as well as the inability of existing target detection algorithms to perform target detection. This invention proposes a method that introduces an attention neural network based on the YOLO architecture to detect rare ANA karyotypes.

[0010] The technical solution of this invention includes the following steps:

[0011] Step 1: Construct a rare ANA karyotype immunofluorescence image target detection dataset and annotate the dataset;

[0012] Step 2: Dataset preprocessing and partitioning;

[0013] Step 3: Use the deep convolutional neural network in the YOLO v5 base network as the basis to extract features from the input image and build a YOLO object detection model; on this basis, embed an attention mechanism into the effective feature layer extracted by the backbone network of the base network and the upsampled result to build a YOLO-based attention neural network model.

[0014] Step 4: Set training parameters, train the constructed YOLO-based and attention neural network model to obtain the optimal parameter model, and use the trained model to process ANA immunofluorescence image data to obtain the target detection results of the image to be tested.

[0015] Step 5: Clinical deployment of trained rare ANA karyotype detection models.

[0016] In this invention, step 1 involves constructing a rare ANA karyotype target detection dataset, specifically:

[0017] Samples with rare ANA karyotypes were collected from the Department of Laboratory Medicine at Xinhua Hospital, affiliated with Shanghai Jiao Tong University School of Medicine, between April 2010 and December 2022, after antinuclear antibody (ANA) testing. Rare ANA karyotypes are defined as karyotypes occurring in less than 1% of ANA karyotypes, and can be divided into 3 groups with 9 subtypes: 1) Cell cycle-related, including NuMA, spindle fiber, CENP-F, intermediate, PCNA, and centriole types; 2) Nuclei, including multinucleated punctate and nuclear membrane types; 3) Cytoplasm, primarily Golgi apparatus types. Senior laboratory technicians with over 10 years of experience in ANA fluorescence slide reading and holding the rank of Associate Chief Technician or higher, labeled the rare ANA karyotype images according to the 2021 International Consensus on Antinuclear Antibody Pattern (ICAP) classification standard for fluorescence karyotypes. 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 with Shanghai Jiao Tong University School of Medicine.

[0018] After obtaining the dataset, the images were labeled using the Labelimg annotation tool to mark the category and location information of the targets in the images, and saved in VOC dataset format. Then, the dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio, with the specific numbers shown in Table 1.

[0019] Table 1. Composition of the Rare ANA Kernel Type Target Detection Dataset

[0020] Classification ICAP number Training set (sheets) Validation set (sheets) Test set (sheets) NuMA (Nuclear mitotic apparatus protein 1) cell-cycle related AC-26 760 84 74 Spindle fiber type (kinesin-5) cell-cycle related AC-25 710 80 72 CENP-F (Centromere proteins F) cell-cycle related AC-14 793 100 94 Midbody cell-cycle related AC-27 469 54 46 PCNA (Proliferating cell nuclear antigen) cell-cycle related AC-13 926 100 86 Multiple nuclear dots nuclear AC-6 575 66 62 Centriole cell-cycle related AC-24 749 74 71 Gorky's physique cytoplasmatic AC-22 1002 90 107 Nuclear envelope type nuclear AC-11 / 12 1000 110 91 total 6984 778 703

[0021] In this invention, step 2 is dataset preprocessing and partitioning, specifically as follows:

[0022] Before model training, data augmentation is required on rare ANA karyotype image data to avoid reduced model detection accuracy due to insufficient sampling information. Image augmentation preprocessing includes random cropping, horizontal flipping, vertical flipping, random rotation, and altering image attributes, including brightness, contrast, saturation, and hue. Subsequently, the images are normalized, and the images and corresponding annotation files are divided into training, validation, and test sets according to a 7:2:1 ratio.

[0023] In this invention, step 3 involves using a deep convolutional neural network in the YOLO v5 base network as the backbone network to extract features from the input image and construct a YOLO target detection model. Specifically:

[0024] The YOLO v5-based feature extraction network consists of an Input layer, a Backbone network, a Neck network, and a Head output layer. The Neck network incorporates an attention mechanism, which includes two independent sub-modules: channel attention and spatial attention. This attention mechanism primarily targets the effective feature layers extracted from the Backbone network and the upsampled results, thereby improving the extraction capability of key features from rare antinuclear antibody karyotype images. The specific implementation steps are as follows:

[0025] 3.1 Input: Mosaic data augmentation is used, and four images are stitched together by random scaling, random cropping, and random arrangement;

[0026] 3.2 Backbone Network: First, the depth and width of the model are set; then, CSPDarknet53 is used as the architecture, mainly including Focus and Bottleneck structures; Focus is responsible for downsampling, and the Bottleneck structure includes CSP and SPP modules. The SPP module compresses the number of channels of feature maps of different scales extracted from the Bottleneck layer and then fuses the multi-scale features; the Bottleneck structure is the most basic structure of the Backbone network, and its main component is the residual block, which combines the transmitted information by summing and then continues to pass it down;

[0027] 3.3 Neck Network: An FPN_PAN structure is inserted between the Backbone and the final Head output layer to form the basic Neck network. The horizontal axis of the FPN feature pyramid structure is considered the scale axis. Scale-invariant feature variables are extracted, and each pyramid feature map is uniformly adjusted to a set high-resolution feature pyramid map. Finally, the high-resolution feature pyramid map is concatenated with the extracted scale-invariant feature variables for detecting feature regions of rare ANA karyotypes at the YOLO model's Head output. Since the YOLO v5 basic Neck network has limited ability to extract features of rare ANA karyotypes, to further improve the network's feature extraction ability for more complex information such as cell division phase, an attention mechanism is embedded in the original Neck network. The attention mechanism includes two independent sub-modules: channel attention and spatial attention. These sub-modules process and pass on the effective feature layers extracted by the backbone network and the upsampled results, respectively. After embedding the attention module, the network model structure of the YOLO v5 backbone network is shown in [see image]. Figure 2 For the attention module, the detailed implementation steps are as follows:

[0028] 3.3.1 The principle and specific implementation process of the channel attention module are as follows: 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) is processed by global max pooling based on width and global average pooling based on height, respectively, to obtain two 1×1×C feature maps. These two maps are then fed into a two-layer multilayer perceptron (MLP) neural network. The features output by the MLP are then summed, and finally, the channel weight coefficients are obtained by passing the sigmoid activation function. Mc (Formula 1). The weighting coefficients... Mc Processing is converted into a one-dimensional channel attention map A T Then, the input feature map F is multiplied by AT at the pixel level (in order to...). (represented), resulting in the salient feature map F of the channel direction. T The calculation formula is as shown in (Formula 3);

[0029] 3.3.2 The principle and specific implementation process of the spatial attention module are as follows: The feature map F is passed through the max pooling layer and the average pooling layer to obtain two H×W×1 channel descriptions. These two descriptions are then concatenated together according to the channels, and then passed through a 7×7 convolutional layer and Sigmoid activation to obtain the spatial weight coefficients. Ms (Formula 2);

[0030] 3.3.3 Obtaining a feature map with attention weights: The spatial weight coefficients obtained in the above steps... Ms Processing is converted into a one-dimensional spatial attention map A K The salient feature map F of the channel direction T With A K After pixel-level multiplication and sequential merging, a feature map F with attention weights is obtained. R (Formula 4);

[0031] ( ) = ( MLP ( AugPool ( + MLP ( MaxPool (

[0032] (1)

[0033] ( ) = ( [( AugPool ( , MaxPool (

[0034] (2)

[0035] (3)

[0036] (4)

[0037] 3.4 Head Output: The original loss function GIOU_Loss is replaced with SIoU on the basis of the original network, and the NMS of the predicted selection box is changed to DIOU_nns to solve the regression convergence problem when the predicted box is in different positions inside the real box due to the dense stacking of image cells.

[0038] In this invention, step 4 involves setting training parameters, training the constructed YOLO-based and attention neural network model, and using the trained model to process immunofluorescence image data to obtain the target detection result of the image under test. Specifically:

[0039] Read the rare ANA keratogram object detection dataset, set the training parameters, and start training. A curve showing the loss value changing over training time is obtained. Once the loss value converges, the model is tested. If the loss value does not converge, the model parameters are adjusted until the model converges. See [link to documentation]. Figure 3 A and B, the model trained at this point will be used as the final model for clinical deployment;

[0040] Since the ANA karyotype images use the human laryngeal cancer epithelial cell line (HEP2 cells) as the detection matrix, there is a problem of dense cell stacking. Therefore, DIOU_nns is used to optimize the area between the predicted box and the ground truth box for convergence, making the results obtained by NMS more reasonable and effective.

[0041] In this invention, step 5 is the deployment of a rare antinuclear antibody karyotype attention neural network model, including 5.1) model detection and 5.2 clinical visualization deployment:

[0042] The model detection described in section 5.1 includes:

[0043] 5.1.1: Read the image to be tested and perform data preprocessing;

[0044] 5.1.2: Normalize the data of the target image to be detected;

[0045] 5.1.3: Input the normalized data into the model that has been trained in step 4 above;

[0046] 5.1.4: Perform model detection on the image to be tested to obtain the prediction results and prediction probabilities.

[0047] The specific deployment of clinical visualization as described in section 5.2 is as follows:

[0048] The trained model is compiled in C++ on the Windows platform, the corresponding deep learning environment is configured, and the model prediction results are visualized by compiling CMake. Then, the model is deployed on a fluorescence microscope CCD camera. The model can read clinical images to be tested and perform detection until the prediction results and probabilities are obtained.

[0049] This invention addresses the issue of missed detection due to insufficient understanding of rare ANA karyotypes among clinical laboratory personnel. It innovatively proposes a method that introduces an attention neural network based on the YOLO architecture to detect rare ANA karyotypes.

[0050] In the method of this invention, for rare ANA karyotype cells with indistinct features, it is necessary to find some extremely subtle and easily overlooked key feature points to make a judgment. The attention neural network model feature extraction introduced can combine the features of low-level (spatial) and high-level (channel) networks.

[0051] Low-level (spatial) information refers to low-level features after multiple downsampling. It provides contextual semantic information about the target feature region within the entire image. This helps in discovering the spatial location information of suspicious rare ANA karyotypes within the entire image.

[0052] High-level (channel) information refers to high-resolution information that is directly passed from the encoder to the decoder of the same height after a concatenation operation. The vectors output from the low-level and high-level network features are fed into ReLU and Sigmoid activations, and the resulting attention weights support the capture of more refined features.

[0053] Compared with existing target detection technologies, the present invention has the following technical advantages:

[0054] (1) This invention introduces an attention mechanism on the basis of the YOLO architecture. The neural network constructed together can efficiently extract key feature points that are extremely subtle and easily ignored in rare ANA karyotype cells, giving full play to the advantages of multiple feature extraction, and making it better adaptable to the complex and varied situations caused by individual differences in patients in clinical practice.

[0055] (2) The YOLO framework loss function of this invention adopts SIoU, which fully considers the spatial orientation factors between the real bounding box and the predicted detection box, making the target regression box more stable and reducing the occurrence of missed detections and false detections. The NMS of the predicted screening box is changed to DIOU_nns to solve the regression convergence problem when the predicted box is at different positions inside the real box due to the dense stacking of image cells;

[0056] (3) This invention effectively improves the detection accuracy of rare ANA karyotypes without reducing the detection speed as much as possible, reduces the probability of false positives and false negatives, and makes a strong contribution to the situation of false positives and false negatives caused by insufficient understanding of rare ANA karyotypes by clinical laboratory personnel.

[0057] The beneficial effects of this invention include: Belonging to the fields of medical-related artificial intelligence technology and biomedicine, this invention can effectively reduce the workload of laboratory physicians, improve work efficiency, and reduce the occurrence of false positives and false negatives in clinical practice. This invention does not require matching with a large number of templates, has a fast processing speed, and high accuracy, meeting the requirements of laboratory physicians in their actual work. Attached Figure Description

[0058] Figure 1A Image description and clinical significance of rare ANA karyotype B. Figure 1A It is AC-13 / 6 / 24 / 22. Figure 1B It is AC-13 / 6 / 24 / 22.

[0059] Figure 2 This describes the model network results and output effects of this invention.

[0060] Figure 3 This is the model training diagram of the present invention. Figure 3 A is the loss curve. Figure 3 B is the model accuracy curve.

[0061] Figure 4 This is a flowchart of the method of the present invention.

[0062] Figure 5 This is an example of the model's output of the image to be tested and the predicted result.

[0063] The present invention will be further described in detail below with reference to the specific embodiments and accompanying drawings, including the process and experimental methods for implementing the present invention. Figure 3 The method flowchart of this invention can be divided into the following steps:

[0064] Data preprocessing: The images of the ANA cell slides to be detected are preprocessed to obtain normalized images.

[0065] Data augmentation involves performing random translation and angle transformations on the image.

[0066] Model loading and prediction: After the model is successfully loaded, the model prediction function is called to predict the image to be tested.

[0067] The network output is a tensor with dimensions K*K*[B*(5*N)], where K represents the number of grids in each test image, B represents the number of bounding boxes in each grid, C represents the coordinates (x, y, h, w) and confidence score of each bounding box, and N is the number of target kernel types. The coordinates (x, y, h, w) of the bounding boxes are: x represents the x-coordinate, y represents the y-coordinate, h represents the height, and w represents the width. Specifically, the expression K*K*[B*(5*N)] means: B is used to identify specific rare kernel features falling at the center point of the grid, and each bounding box corresponds to a score representing whether the bounding box contains an object and its confidence score. The confidence score is the predicted probability that the bounding box contains rare kernel features, and the score represents the confidence level of the rare kernel features in the grid. When judging the accuracy of model detection, this invention uses SloU (Scylla Intersection over Union) to evaluate model performance. The specific calculation of SIoU is shown in formula (4).

[0068]

[0069] in Representing distance cost and angle cost, specifically the distance between the center point of the predicted bounding box and the center point of the ground truth box, and the angle formed by the line connecting the two points and the perpendicular line between their heights; The shape cost represents the similarity between the predicted bounding box and the ground truth bounding box; the IoU represents the IoU value between the predicted bounding box and the ground truth bounding box. The power is determined by SIoU. Predicted / True This represents the IoU value between the predicted bounding box and the ground truth bounding box.

[0070] Figure 2 This is a schematic diagram illustrating the results of detecting rare ANA karyotype images in clinical settings using the method of this invention. Figure 5 The image shown is an example of the prediction made based on the discovered rare ANA karyotype. It can be seen that the method of this invention can effectively detect rare ANA karyotypes, reduce the rate of missed detection and misdiagnosis by laboratory physicians, and improve clinical work efficiency.

Claims

1. A method for detecting rare antinuclear antibody karyotypes based on YOLO and attention neural networks, characterized in that, The method includes: Step 1: Construct a rare antinuclear antibody nuclear type immunofluorescence image target detection dataset and annotate the dataset; Step 2: Dataset preprocessing and partitioning; Step 3: Use YOLO v5 as the base network to build a target detection model, and embed an attention mechanism into the effective feature layer extracted from the backbone network part of the base network structure and the upsampled result, thereby building a YOLO-based attention neural network model; The YOLO v5-based feature extraction network consists of an Input layer, a Backbone network, a Neck network, and a Head output layer. The Neck network incorporates an attention mechanism, which includes two independent sub-modules: channel attention and spatial attention. This attention mechanism primarily targets the effective feature layers extracted from the Backbone network and the upsampled results, thereby improving the extraction capability of key features from rare antinuclear antibody karyotype images. The specific implementation steps are as follows: 3.1 Input: Mosaic data augmentation is used, and four images are stitched together by random scaling, random cropping, and random arrangement; 3.2 Backbone Network: First, the depth and width of the model are set; then, CSPDarknet53 is used as the architecture, mainly including the Focus and Bottleneck structures; the Focus is responsible for downsampling, and the Bottleneck structure includes CSP and SPP modules. The SPP module compresses the number of channels of the feature maps of different scales extracted from the Bottleneck structure and then fuses the multi-scale features; the Bottleneck structure is the most basic structure of the Backbone network, and its main component is the residual block, which combines the transmitted information by summing and then continues to pass it down; 3.3 Neck Network: An FPN_PAN structure is inserted between the Backbone and the final Head output layer to form the basic Neck network. The horizontal axis of the FPN feature pyramid structure is considered the scale axis. Scale-invariant feature variables are extracted, and each pyramid feature map is uniformly adjusted to a set high-resolution feature pyramid map. Finally, the high-resolution feature pyramid map is concatenated with the extracted scale-invariant feature variables to detect feature regions of rare ANA karyotypes at the YOLO model's Head output. Since the YOLO v5 basic Neck network has limited ability to extract features of rare ANA karyotypes, an attention mechanism is embedded on top of it. The specific implementation steps are as follows: The input feature map, i.e., the corrected H×W×C, is processed by three convolutional layers to obtain a basic feature map F of H×W×48, where W and H are the width and height of the feature map, respectively, and C is the number of channels in the feature map. Then, a channel attention module is used to transform the input feature map F into a one-dimensional channel attention map A. T Then, the input feature map F and A are compared. T Perform pixel-level multiplication, to This indicates that the salient feature map F of the channel direction is obtained. T The calculation is as shown in formula (1), and then the spatial attention module is used to process the input feature map F. T Convert to a one-dimensional spatial attention map A K Finally, F T With A K Perform pixel-level multiplication to obtain the output feature map F. R The calculation is as shown in formula (2); the channel attention feature map F obtained in the above steps is... T Spatial attention diagram A K Sequential merging yields a feature map with attention weights; （1） （2） 3.4 Head Output: The loss function is replaced with SIoU to correct the spatial orientation factors between the ground truth bounding box and the predicted detection box, making the target regression box more stable; the NMS of the predicted selection box is changed to DIOU_nns to solve the regression convergence problem when the predicted box is in different positions inside the ground truth box due to the dense stacking of image cells. Step 4: Using the target detection dataset, train the constructed YOLO-based and attention neural network model to obtain the optimal parameter model, and test the trained model using unlabeled ANA karyotype immunofluorescence images collected clinically to obtain the target detection results of the image to be tested. Step 5: Clinical deployment of trained YOLO and attention neural network models.

2. The method for detecting rare antinuclear antibody karyotypes using YOLO and attention neural networks as described in claim 1, characterized in that, Step 1 specifically involves: A dataset was created by collecting immunofluorescence images of rare ANA karyotypes from clinical testing. The characteristic regions of rare ANA karyotypes were labeled with category and location information using the image annotation tool Labelimg. Rare ANA karyotypes were categorized into nine types: NuMA, CENP-F, Kinesin-5, Midbody, PCNA, Multiple nuclear dots, Centriole, Golgi, and Nuclearenvelope. XML format annotation files were generated that correspond one-to-one with the images of each type of karyotype, and a target detection dataset was constructed.

3. The method for detecting rare antinuclear antibody karyotypes using YOLO and attention neural networks as described in claim 1, characterized in that, Step 2 specifically involves: The image data augmentation preprocessing includes random cropping, horizontal flipping, vertical flipping, random rotation, and changing image attributes, including brightness, contrast, saturation, or hue. Subsequently, the image is normalized, and the image and its corresponding annotation file are divided into training set, validation set, and test set according to a 7:2:1 ratio.

4. The method for detecting rare antinuclear antibody karyotypes using YOLO and attention neural networks as described in claim 1, characterized in that, Step 4 specifically involves: After setting the training parameters, model training begins, resulting in a curve showing the change of loss value over training time. Once the loss value converges, the model is tested. If the loss value does not converge, the model parameters are adjusted until the model accuracy is at its highest, thus obtaining a well-trained model. The model performance is evaluated using SloU. The specific calculation of SloU is shown in formula (3). in Representing distance loss and angle loss, specifically the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, and the angle formed by the line connecting the two points and the perpendicular line between their heights; This represents shape loss, specifically the predicted similarity between the bounding box and the ground truth box shape; IoU specifically refers to the IoU value between the predicted bounding box and the ground truth bounding box, while the power is determined by SIoU. Predicted / True The value represents the IoU between the predicted bounding box and the ground truth bounding box. Since the ANA karyotype image uses human laryngeal cancer epithelial cell line as the detection matrix, there is a problem of dense cell stacking. Therefore, DIOU_nns is used to optimize the area between the predicted bounding box and the ground truth bounding box for convergence, making the results obtained by NMS more reasonable and effective.

5. The method for detecting rare antinuclear antibody karyotypes using YOLO and attention neural networks as described in claim 1, characterized in that, Step 5 specifically involves: The trained model is compiled in C++ on the Windows platform, the corresponding deep learning environment is configured, and the model prediction results are visualized by compiling CMake. Then, the model is deployed on a fluorescence microscope CCD camera. The model can read clinical images to be tested and perform detection until the prediction results and prediction probabilities are obtained.

Images