HPV typing automatic identification method based on deep learning
By combining deep learning and machine learning, the problems of low detection efficiency and insufficient accuracy in existing HPV detection technologies have been solved, achieving efficient and accurate automated HPV genotyping identification and report generation.
Patent Information
- Application Number
- CN202511626387.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-02-24
AI Technical Summary
In existing HPV detection technologies, the image analysis software is outdated, the hybridization staining differentiation algorithm is not accurate enough, and there is a lack of localized operation functions, resulting in low detection efficiency and insufficient accuracy.
An automated HPV genotyping identification method based on deep learning is adopted. It utilizes the YOLOv11 target detection framework and ViT classification structure, combined with a Logistic regression model, to achieve automated processing of membrane strip images and visualization of results. Multi-dimensional feature analysis is used to improve detection accuracy.
It improves the accuracy of HPV typing, enables automated uploading and partial automation of test reports, reduces the rate of human error, and significantly speeds up the testing process.
Smart Images
Figure CN121564708A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics detection technology, specifically to an automated HPV genotyping identification method based on deep learning. Background Technology
[0002] Human papillomavirus (HPV) is a group of non-enveloped, small, double-stranded circular DNA viruses. Its structure consists of a viral protein capsid and a core of a single copy of viral genomic DNA. Belonging to the Papillomaviridae family, its genomic DNA is approximately 8000 bp in length. Clinically, HPV is classified into low-risk and high-risk types based on its carcinogenicity. Low-risk HPV infection primarily causes genital warts, while persistent infection with high-risk HPV is a major cause of cervical cancer and cervical lesions. Notably, cervical cancer has the heaviest disease burden among HPV-related diseases, is one of the most common cancers, and is a leading cause of cancer-related deaths among women worldwide. However, cervical cancer is a preventable disease, and clinical cure can be achieved through early detection and timely intervention.
[0003] In terms of HPV detection technology, since the virus is difficult to culture in vitro and its serological characteristics are not obvious after infection, clinical diagnosis currently mainly relies on nucleic acid detection technology, especially HPV DNA detection. The existing commonly used HPV detection methods in clinical practice include: PCR-reverse dot hybridization, hybridization capture method, PCR-fluorescent probe method, etc. Among them, the chip technology based on the combination of PCR amplification and DNA reverse dot hybridization has been widely used in commercial detection kits, and the HPV genotype of the sample to be tested is determined by analyzing the hybridization signal of the detection site. Taking the "Human Papillomavirus Genotyping (23 types) Detection Kit (National Medical Device Registration Certificate 20233400811)" developed by Yaneng Bio as an example, although the product is equipped with a fully automated nucleic acid molecular hybridization instrument and an automatic imaging system, it still has the following technical limitations: (1) the image analysis software version used is outdated; (2) the hybridization staining differentiation algorithm is only based on simple RGB value judgment, which is not accurate enough; (3) the system lacks localized operation functions, which affects the detection efficiency. To address the aforementioned shortcomings of existing technologies, this invention proposes an innovative HPV genotyping method. This method employs the state-of-the-art object detection framework YOLOv11, a Transformer-based image classification structure (ViT), and a Logistic regression machine learning strategy. Through multi-dimensional feature analysis, it significantly improves the accuracy of HPV genotyping. Furthermore, the system implements automated uploading of test results, better meeting the daily testing needs of laboratories. Summary of the Invention
[0004] This invention provides an intelligent HPV genotyping detection method that overcomes the problems of light staining from hybridization instruments and minor contamination caused by untimely operation or cleaning. This invention improves detection efficiency, enabling streamlined batch result processing and automatic generation of test reports.
[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an automated HPV genotyping identification method based on deep learning, comprising the following steps: Step 1: Use a trained deep learning image detection model to determine the staining status of the IC mark in the membrane strip, and identify and locate the handwritten content (including numbers, letters and combinations of numbers and letters) on the membrane strip. The membrane strip is an image acquired by an automatic shooting device embedded in the hybridization instrument, and the image adopts RGB color mode. Step 2: After completing the membrane strip localization, the trained deep learning image detection model is used again to locate the staining region of each HPV genotype, and the HPV genotype staining region is divided into negative and positive using a dual-channel method (ViT classification framework with self-attention mechanism + Logistic regression model). Step 3: Employing a unique data preprocessing mechanism, the positive HPV types in Step 2 are scored by calculating the different pixel scores of HPV genotyping staining, and the positive HPV types to be verified are distinguished. Step 4: Visualize the test results. By structuring the HPV genotyping scores and connecting them to the test management system platform, the test report can be visualized.
[0006] In an embodiment of the present invention, the deep learning image detection model training step in step one is as follows: Nucleic acid samples to be tested are collected from reproductive tract secretions or cervical exfoliated cells. The categories of the samples to be tested are identified (categories in the first model: total, IC, 1, 2, 3, 4, 5, 6, 7, 8, 9, 0, F, etc.), and the categories of the samples are defined. Use the LabelMe tool to select target categories for samples; Training classification is performed on the dataset obtained after selecting the box in step b, with a training set: validation set: test set ratio of 8:1:1; The YOLOv11 framework was used to train the dataset for the sample categories in step a, and the best training model was finally obtained, with an mAP (average precision across all classes) of 0.993.
[0007] In an embodiment of the present invention, the membrane strip mentioned in step one is an image acquired by an automatic imaging device embedded in the hybridization instrument, and the image adopts RGB color mode. The target membrane strip positioning includes the following steps: S101, the object detection model is YOLOv11, which includes the CSPDarknet backbone network, the lightweight residual C3K2 module, the fast spatial pyramid pooling structure SPPF module, and the feature multi-scale enhancement C2PSA module, i.e., the above-mentioned trained deep learning image detection model; the trained deep learning image detection model can predict the set of bounding boxes. , where each predicted bounding box Includes coordinate information of x, y, w, and h, and confidence level. and category information When multiple predicted bounding boxes exist, the optimal target membrane strip can be determined based on its positional centrality and confidence level. The positional centrality formula is as follows: ,in , These refer to the coordinate center position of the i-th target membrane strip, with a score ranging from [0, 1], while the confidence threshold T of the target membrane strip... 阈值 The setting is 0.7. The selection criteria for the optimal target membrane strip are: confidence level greater than 0.7 and minimum positional centrality value; if the optimal membrane strip frame cannot be detected, the system will automatically prompt for manual judgment.
[0008] S102. After determining the optimal target membrane strip, based on the intersection-to-exclusion ratio of the IC mark frame and the target membrane strip ( The optimal IC logo frame is determined using the following formula: ,in r ic This represents the rightmost coordinate value of the IC frame; r total This represents the rightmost coordinate value of the total box; l ic Indicates the leftmost coordinate value of the IC frame; l toral This represents the leftmost coordinate value of the total box; b ic This represents the bottom coordinate value of the IC frame, b. total This indicates the bottom coordinates of the total box; t ic t represents the top coordinate value of the IC frame. total This indicates the coordinate value of the topmost point of the total box; area ic The area represents the area of the IC frame. total This indicates the area of the total box.
[0009] Finish After calculation, the maximum value is the optimal IC marker frame. IC marker frame identification can be based on category information, primarily categorized as stained or unstained. It is important to note that this formula requires the IC marker frame to intersect with the target membrane strip. If the IC marker frame cannot be detected, the system will automatically prompt for manual judgment.
[0010] S103. Detect information for all location categories based on the object detection model, namely category, confidence score, and location information. Set the minimum confidence score L. 阈值 =0.5, filtering out invalid positions; using DBSCAN clustering to extract position information and screen out valid position information; the final content of the sample can be arranged according to the valid position information (from left to right or from top to bottom), which is actually determined by the writing direction in the image; finally, the optimal membrane strip is located.
[0011] In the embodiments of the present invention, the deep learning detection model used in step one can effectively determine the coloring of the IC mark frame and eliminate the phenomenon of light coloring or local irregular coloring.
[0012] In the embodiments of the present invention, the HPV typing in step one includes, but is not limited to, the following 23 types, including 18 high-risk types: HPV16, 18, 26, 31, 33, 35, 39, 45, 51, 52, 53, 56, 58, 59, 66, 68, 73, 82; and 5 low-risk types: HPV6, 11, 42, 43, 81.
[0013] In the embodiment of this invention, the dual-channel method described in step two employs a ViT image classification strategy with a self-attention mechanism and a Logistic regression model based on HPV genotype pixel values. ViT utilizes a TransformerEncoder and an MLP Header module to classify whether each HPV genotype is stained, effectively extracting network structure features, spatial correlation features, and relative position features within HPV genotypes. After training, the Top-1 accuracy can reach 100%. The Logistic regression model, starting from the pixel values of the image, transforms image features into numerical features, achieving a binary classification strategy. After training, the evaluation metric AUC can reach 0.98.
[0014] In an embodiment of the present invention, the HPV genotyping and staining quantification in step three employs a unique preprocessing mechanism, the specific steps of which are as follows: S301, HPV Genotyping Position Correction: Affine transformation is used to detect lines in HPV genotyping, the line inclination is calculated, and the nuclear membrane position is corrected by the inclination. S302. HPV Genotyping Specific Location Extraction: Randomly select stained HPV genotypes, extract the central and edge regions from the corrected HPV genotyping image, and set the side length of the central region. , edge region side length Proportion of marginal areas After determining the side lengths of the central and edge regions, obtain all pixel values for the central and edge regions respectively, calculate the mean and variance, and analyze the mean and variance under different region side lengths. The optimal side length is the one with the lowest variation in mean and variance. The edge region ratio is determined by taking three small regions (top left, top right, and bottom right) from the center position. The optimal ratio is the one that minimizes the variance of pixel values while ensuring that the pixel values are the background color. The formula for calculating the proportion of the edge region is as follows: S rate = , of which S peripheral The coordinates (horizontal or vertical) of the center point of the edge region are represented by Width and Height, which represent the width and height of a single HPV parting frame, respectively.
[0015] S303, HPV Genotyping Numerical Quantization: After determining the optimal location based on the edge region ratio, the average pixel value of the edge regions of all HPV genotypes is statistically analyzed to obtain the final average pixel value of the negative HPV genotype. The quantization score for each HPV genotype is calculated by combining the average pixel value of the central region of the positive HPV genotype. The calculation formula is as follows: Where N represents the number of RGB channels in the image, i.e., N=3; Represents the pixel value of the uncolored subtype; The pixel value represents the color classification.
[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. Improve detection accuracy and reduce human error. This invention first employs deep learning object detection models (YOLOv11 and ViT) to achieve qualitative detection of HPV genotypes, with an mAP evaluation index reaching 0.993. Then, it utilizes conventional machine learning models to transform image data into numerical data, realizing a quantification mechanism for HPV genotype scores, with an AUC effect reaching 0.98. This effectively avoids judgment errors caused by differences in instrument parameters, lighting conditions, and the energy and state of personnel.
[0017] 2. Increased detection efficiency, reduced manpower, and lower costs. Regarding detection efficiency, the average processing time for each person per image to verify detection results is approximately 20 seconds, while this invention processes a single image in just 0.5 seconds, improving efficiency by nearly 40 times. This significantly accelerates HPV genotyping detection.
[0018] 3. Achieve local automation. After being detected by the hybridization instrument, the detected image is processed by this invention, and a pathological report can be directly generated, thus achieving local automation. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the overall process of automated HPV typing identification according to the present invention. Figure 2 This is a comparison diagram of the affine transformation effects during the preprocessing of this invention; Figure 3 This is an analysis diagram showing the region delineation during preprocessing in this invention; Figure 4 This is a three-dimensional spatial distribution diagram of the stained and unstained typing of this invention; Figure 5 This is the optimal membrane frame for sample detection and screening in this invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] The definitions of the main abbreviations and key terms involved in this invention are as follows: HPV: human papillomavirus; PCR: polymerase chain reaction; RGB: Red, Green, and Blue, the three primary color light model; ViT: Vision Transformer, a novel model that introduces the Transformer architecture from the field of natural language processing to computer vision, capable of processing image data; MLP Header: A multilayer perceptron head is a key component in deep learning models used for specific task processing, typically used to map features to the final classification result.
[0022] CSPDarknet: Cross Stage Partial Darknet, is an improved backbone network structure introduced in YOLO v4, designed to improve model performance and efficiency through computational redundancy; SPPF module: Spatial Pyramid Pooling-Fast, is an improved spatial pyramid pooling structure designed to enhance the accuracy and efficiency of target detection. The C2PSA module is a high-level module in YOLO 11 used to enhance feature extraction. It combines the CSP (CrossStage Partial) structure and the PSA (Pyramid Squeeze Attention) attention mechanism to improve multi-scale feature extraction capabilities. IC Marker Frame: Internal control, an internal control design for the membrane strip. If the IC marker frame is not colored, the experimental results of the membrane strip cannot be used. IoU: intersection over union, which is the ratio of the area of intersection of two bounding boxes to the area of union of the two bounding boxes; DBSCAN Clustering: Density-Based Spatial Clustering and Application with Noise, a clustering algorithm used to cluster noisy spatial data; AUC: Area Under the Curve; mAP: mean Average Precision, is an important metric for evaluating object detection models and is usually used to compare the performance of different models. Example
[0023] A deep learning-based HPV genotyping identification method, such as Figure 1 As shown, it includes the following steps: Step 1: Use the trained deep learning image detection model to determine the staining status of the IC mark, and identify and locate the handwritten content (including letters, numbers and combinations of letters and numbers) by the experimental operator. The membrane strip is an image acquired by the automatic shooting device embedded in the hybridization instrument, and the image adopts RGB color mode. Step 2: After completing the membrane strip localization, the trained deep learning image detection model is used again to locate the staining region of each HPV genotype, and the HPV genotype staining region is divided into negative and positive using a dual-channel method (ViT classification framework with self-attention mechanism + Logistic regression model). Step 3: Employing a unique data preprocessing mechanism, the positive HPV types in Step 2 are scored by calculating the different pixel scores of HPV genotyping staining, and the positive HPV types to be verified are distinguished. Step 4: Visualize the test results. By structuring the HPV genotyping scores and connecting them to the test management system platform, the test report can be visualized.
[0024] In an embodiment of the present invention, the deep learning image detection model training step in step one is as follows: Nucleic acid samples to be tested are collected from reproductive tract secretions or cervical exfoliated cells. The categories of the samples to be tested are identified (categories in the first model: total, IC, 1, 2, 3, 4, 5, 6, 7, 8, 9, 0, F, etc.), and the categories of the samples are defined. Use the LabelMe tool to select target categories for samples; Training classification is performed on the dataset obtained after selecting the box in step b, with a training set: validation set: test set ratio of 8:1:1; The YOLOv11 framework was used to train the dataset for the sample categories in step a, and the best training model was finally obtained, with an mAP (average precision across all classes) of 0.993.
[0025] In an embodiment of the present invention, the membrane strip mentioned in step one is an image acquired by an automatic imaging device embedded in the hybridization instrument, and the image adopts RGB color mode. The specific steps for positioning the target membrane strip include: S101: The object detection model is YOLOv11, which includes the CSPDarknet backbone network, the lightweight residual C3K2 module, the fast spatial pyramid pooling structure SPPF module, and the feature multi-scale enhancement C2PSA module, i.e., the above-mentioned trained deep learning image detection model; a set of predicted bounding boxes can be obtained through prediction. , where each predicted membrane frame Includes coordinate information (x, y, w, h) and confidence level. and category information When multiple predicted bounding boxes exist, the optimal target membrane strip can be determined based on its positional centrality and confidence level. The positional centrality formula is as follows: ,in , These refer to the coordinate center positions of the i-th target membrane strip, with scores ranging from [0,1]. The confidence threshold T of the target membrane strip is... 阈值 The optimal screening criteria for the target membrane strip are set to 0.7: a confidence level greater than 0.7 and the lowest possible centrality value. If the optimal membrane strip frame cannot be detected, the system will automatically prompt for manual judgment. Please refer to [link to relevant documentation]. Figure 5 .
[0026] S102. After determining the optimal membrane strip, based on the intersection-to-exclusion ratio of the IC mark frame and the target membrane strip ( The optimal IC logo frame is determined using the following formula: ,in r ic This represents the rightmost coordinate value of the IC frame; r total This represents the rightmost coordinate value of the total box; l ic Indicates the leftmost coordinate value of the IC frame; l toral This represents the leftmost coordinate value of the total box; b ic This represents the bottom coordinate value of the IC frame, b. total This indicates the bottom coordinates of the total box; t ic t represents the top coordinate value of the IC frame. total This indicates the coordinate value of the topmost point of the total box; area ic The area represents the area of the IC frame. total This indicates the area of the total box.
[0027] Finish After calculation, the maximum value is the optimal IC marker frame. IC marker frame identification can be based on category information, primarily categorized as stained or unstained. It's important to note that this formula requires the IC marker frame to intersect with the target membrane strip. If the IC marker cannot be detected, the system will automatically prompt for manual judgment.
[0028] In the embodiments of the present invention, the deep learning detection model used in step one can effectively determine the coloring of the IC mark frame and eliminate the phenomenon of light coloring or local irregular coloring.
[0029] S103. Detect information for all location categories based on the object detection model, namely category, confidence score, and location information. Set a minimum confidence score L. 阈值 =0.5, filtering out invalid positions; DBSCAN clustering is used to extract position information and screen out valid position information; the final content of the sample can be arranged according to the valid position information (from left to right or from top to bottom), which is actually determined according to the writing direction in the image, and finally the optimal membrane strip is located; in the embodiments of the present invention, the position information and confidence score of the detected sample are detailed in Table 1.
[0030] In Embodiment 1 of the present invention, the HPV genotyping in step one includes, but is not limited to, the following 23 HPV genotypes, including 18 high-risk types: HPV16, 18, 26, 31, 33, 35, 39, 45, 51, 52, 53, 56, 58, 59, 66, 68, 73, 82; and 5 low-risk types: HPV 6, 11, 42, 43, 81.
[0031] The HPV genotyping test in this embodiment 1 uses the "Human Papillomavirus Genotyping (Type 23) Detection Kit (PCR-Reverse Dot Hybridization Method)" (National Medical Device Registration Certificate 20233400811) developed by Yaneng Biotechnology (Shenzhen) Co., Ltd. In the embodiments of the present invention, different testing laboratories can use the HPV genotyping kit as needed.
[0032] In the embodiment of this invention, the dual-channel method described in step two employs a ViT image classification strategy with an attention mechanism and a Logistic regression model based on HPV genotype pixel values. ViT utilizes a TransformerEncoder and an MLP Header module to classify whether each HPV genotype is stained, effectively extracting network structure features, spatial correlation features, and relative position features within HPV genotypes. After training, the Top-1 accuracy can reach 100%. The Logistic regression model, starting from the pixel values of the image, transforms image features into numerical features, achieving a binary classification strategy. After training, the evaluation metric AUC can reach 0.98.
[0033] In an embodiment of the present invention, the HPV genotyping and staining quantification in step three employs a unique preprocessing mechanism, the specific steps of which are as follows: S301, HPV Genotyping Position Correction: Because the nuclear membrane position in the input image is relatively random, this step uses affine transformation to detect lines in the HPV genotyping sequence, calculates the line inclination, and corrects the random nuclear membrane position using the inclination. See [link to relevant documentation]. Figure 2 .
[0034] S302. HPV Genotyping Specific Location Extraction: Randomly select stained HPV genotypes, extract the central and edge regions from the corrected HPV genotyping image, and set the side length of the central region. , edge region side length Proportion of marginal areas After determining the region's side length, obtain all pixel values for the central and edge regions, calculate the mean and variance, and analyze the mean and variance under different region side lengths. The optimal side length is the one with the lowest variation in mean and variance. The edge region ratio is determined by taking three small regions (top left, top right, and bottom right) from the center position. The optimal ratio minimizes the pixel value variance while ensuring that the pixel value is the background color.
[0035] The formula for calculating the proportion of the edge region is as follows: S rate = , of which S peripheral The coordinates (horizontal or vertical) of the center point of the edge region are represented by Width and Height, which represent the width and height of a single HPV parting frame, respectively.
[0036] In an embodiment of the present invention, the optimal threshold is =20, =10, =0.15, see Figure 3 .
[0037] S303, HPV genotyping numerical quantification: After determining the optimal position based on the proportion of the edge region, the average pixel value of the edge region (unstained) is combined with the average pixel value of the positive HPV genotyping center region (stained). Figure 4 The distribution maps of stained and unstained HPV genotypes in 3D space are generated, and the average pixel value of unstained HPV is obtained. An HPV genotyping quantification formula is constructed using the pixel values of stained and unstained HPV types. The quantification score for each genotype is obtained based on the results of HPV genotyping at different locations. The calculation formula is as follows: Where N represents the number of RGB channels in the image, i.e., N=3; Represents the pixel value of the uncolored subtype; The pixel values represent the colorimetric classification; see Table 1 for specific calculation metrics.
[0038] Table 1. Data of the test samples (n=20)
[0039] In an embodiment of the present invention, if the relative value of the detected HPV genotype is less than 20 (calculated based on the sample detection data of the unit's laboratory, and there may be deviations between the results of different laboratories), it will be automatically included in the column that needs to be verified, that is, the HPV genotype needs to be further verified to ensure the accuracy of the detection.
[0040] This invention first employs deep learning object detection models (YOLOv11 and ViT) to achieve qualitative detection of HPV genotypes, with an mAP evaluation index reaching 0.993. Then, conventional machine learning models are used to transform image data into numerical data, realizing a quantification mechanism for HPV genotype scores, with an AUC effect reaching 0.98. This effectively avoids judgment errors caused by differences in instrument parameters, lighting conditions, and the energy and condition of the experimenters.
[0041] Furthermore, in terms of detection efficiency, the average processing time for each person per image to verify the detection results is about 20 seconds, while the technical solution of this invention only takes 0.5 seconds to process a single image, which is nearly 40 times more efficient.
[0042] In the embodiments of the present invention, the images detected by the hybridization instrument are processed by the present invention to directly determine HPV typing and generate pathological reports directly, which can achieve local automation, significantly shorten the report generation time, and improve work efficiency.
[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0044] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.
Claims
1. An automated HPV genotyping identification method based on deep learning, characterized in that, Includes the following steps: Step 1: Use the trained deep learning image detection model to determine the staining status of the IC mark in the membrane strip, and identify and locate the handwritten content on the membrane strip. The membrane strip is an image acquired by the automatic shooting device embedded in the hybridization instrument, and the image adopts RGB color mode. Step 2: After completing the membrane strip localization, the trained deep learning image detection model is used again to locate the staining region of each HPV type, and the staining region of the HPV type is divided into negative and positive using a dual-channel method of ViT classification framework with self-attention mechanism and Logistic regression model. Step 3: Using a data preprocessing mechanism, the positive HPV types in Step 2 are scored by calculating the different pixel scores of HPV typing and distinguishing the positive HPV types to be verified. Step 4: Visualize the test results. By structuring the HPV genotyping scores and connecting them to the test management system platform, the test report can be visualized.
2. The automated HPV genotyping identification method based on deep learning according to claim 1, characterized in that, The training steps for the deep learning image detection model in step one are as follows: a. Collect nucleic acid samples to be tested from reproductive tract secretions or cervical exfoliated cells, identify the category of the sample to be tested, and define the category of the sample; b. Use the LabelMe tool to select target categories for samples; c. Train classification on the dataset obtained after selecting the box in step b, with a training set: validation set: test set ratio of 8:1:1; d. Apply the YOLOv11 framework to the sample categories in step a, and train the dataset to obtain the best training model.
3. The automated HPV typing identification method based on deep learning according to claim 2, characterized in that, Step one also includes the following steps: S101. The predicted bounding box set can be obtained by using the trained deep learning image detection model in claim 2. , where each predicted bounding box Includes coordinate information of x, y, w, and h, and confidence level. and category information When multiple predicted bounding boxes exist, the optimal target membrane strip can be determined based on its positional centrality and confidence level. The positional centrality formula is as follows: ,in , These refer to the coordinate center positions of the i-th target membrane strip, with scores ranging from [0, 1]. The confidence threshold T of the target membrane strip is then used. 阈值 The set value is 0.7; the selection criteria for the optimal target membrane strip are: confidence level greater than 0.7 and minimum positional centrality value; if the optimal membrane strip frame cannot be detected, the system will automatically prompt for manual judgment. S102. After determining the optimal target membrane strip, based on the intersection-to-exclusion ratio of the IC mark frame and the target membrane strip ( The optimal IC logo frame is determined using the following formula: ,in r ic This represents the rightmost coordinate value of the IC frame; r total This represents the rightmost coordinate value of the total box; l ic Indicates the leftmost coordinate value of the IC frame; l toral This represents the leftmost coordinate value of the total box; b ic This represents the bottom coordinate value of the IC frame, b. total This indicates the bottom coordinates of the total box; t ic t represents the top coordinate value of the IC frame. total This indicates the coordinate value of the topmost point of the total box; area ic The area represents the area of the IC frame. total This indicates the area of the total box; Finish After calculation, its maximum value is the optimal IC marker frame. The identification of the IC marker frame can be determined based on category information, mainly in two states: stained and unstained. The premise of using this formula is that the IC marker frame intersects with the target membrane strip. If the system cannot detect the IC marker frame, the system will automatically prompt for manual judgment. S103. Based on the information of all location categories detected by the target detection model, namely category, confidence score, and location information, set a minimum confidence score L. 阈值 =0.5, filtering out invalid positions; using DBSCAN clustering to extract position information and screen out valid position information; the final content of the sample can be arranged according to the valid position information, which is actually determined by the writing direction in the image; finally, the optimal membrane strip is located.
4. The automated HPV typing identification method based on deep learning according to claim 1, characterized in that, In step three, the data preprocessing mechanism includes the following steps: S301, HPV Genotyping Position Correction: Affine transformation is used to detect lines in HPV genotyping, the line inclination is calculated, and the membrane strip position is corrected by the inclination. S302. HPV Genotyping Specific Location Extraction: Randomly select stained HPV genotypes, extract the central and edge regions from the corrected HPV genotyping image, and set the side length of the central region. , edge region side length ratio of marginal areas After determining the side lengths of the central and edge regions, obtain all pixel values for both regions, calculate the mean and variance, and analyze the mean and variance for different region side lengths. The optimal side length is the one with the lowest variation in mean and variance. The edge region ratio is determined by taking three square regions outwards from the center position. The optimal ratio minimizes the pixel value variance while ensuring the pixel value is the background color. The formula for calculating the edge region ratio is as follows: S rate = , among which, S peripheral The coordinates of the center point of the edge region are represented by Width and Height, which represent the width and height of a single HPV parting frame, respectively. S303, HPV Genotyping Numerical Quantization: The optimal location is determined based on the proportion of the edge region. The average pixel value of the edge regions of all HPV genotypes is then statistically analyzed to obtain the final average pixel value for negative HPV genotypes. The quantization score for each HPV genotype is calculated by combining the average pixel value of the central region of positive HPV genotypes. The calculation formula is as follows: Where N represents the number of RGB channels in the image, i.e., N=3; Represents the pixel value of the uncolored subtype; The pixel value represents the color classification.