Method for detecting cervical lesion cast-off cells or masses based on YOLOv10 network model
By improving the YOLOv10 network model and combining SGPM and CGM-SA modules, the problems of scale differences and cell overlap in cervical lesion cell images were solved, improving detection accuracy and efficiency, and making it suitable for automatic detection of cervical lesion cells.
Patent Information
- Application Number
- CN202511439513.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-10
- Publication Date
- 2026-01-13
AI Technical Summary
Existing technologies struggle to simultaneously perceive the global structure of large cell clusters and capture the detailed features of small cells when processing images of cervical lesions. Furthermore, the dense overlap of cells leads to high rates of missed and false detections, limiting the efficiency and accuracy of automated detection.
The YOLOv10 network model is adopted, combined with a simplified multi-scale pooling feature extraction module (SGPM) and a channel-guided multi-scale spatial attention module (CGM-SA) to enhance the ability to perceive features of cells at different scales and distinguish overlapping cells. The training process is optimized through a composite loss function.
It significantly improves the detection capability for multi-scale targets, reduces the false negative and false positive rates, and achieves a balance between high accuracy and high efficiency, making it suitable for large-scale clinical screening.
Smart Images
Figure CN121330668A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and medical image processing technology, specifically a method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model. Background Technology
[0002] Cervical cancer is a common gynecological malignancy worldwide, and early screening is crucial to reducing mortality. Traditional Pap smears and liquid-based thin-layer cytology (LBC) tests rely heavily on pathologists' manual microscopic reading, which is time-consuming, inefficient, and susceptible to subjective influences on diagnostic results.
[0003] In recent years, deep learning-based object detection technology has been applied to the automatic identification of cervical lesions, demonstrating great potential. Currently, Zhang et al. proposed using Faster R-CNN and Retina-Net to detect cervical lesions on large-scale cervical cell datasets. Liang et al. proposed adding a global context-aware framework and a soft-scale anchor matching strategy to YOLOv3 for cervical lesion cell detection. Cao et al. proposed an attention feature pyramid network for lesion cell detection in cervical cytology images. However, existing technologies still face two major challenges when processing complex clinical cytology images:
[0004] 1. Cell polymorphism: Diseased cells vary greatly in size and morphology. The fixed receptive field of traditional convolutional networks is difficult to simultaneously take into account the global structural perception of large cell clusters and the capture of detailed features of small cells.
[0005] 2. Dense cell overlap: Cells in smears are often densely arranged or even overlapping each other, making it difficult for the model to distinguish individual cell instances, which can easily lead to missed detections and false detections.
[0006] Existing technologies either suffer from complex model structures and high computational costs, or lack targeted design, thus failing to effectively address the aforementioned challenges and limiting their application in large-scale clinical screening. Therefore, there is an urgent need for an automated detection method that balances efficiency, accuracy, and generalization ability, optimized for the specific characteristics of cervical lesion cell images.
[0007] Therefore, a new solution is needed to address the above problems. Summary of the Invention
[0008] The purpose of this invention is to provide a method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model, so as to solve the technical problems mentioned in the background art.
[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model, comprising at least the following steps:
[0010] S1: Construct a basic network model, using YOLOv10 as the detection framework and as the backbone network of the overall structure;
[0011] S2: Feature extraction, an improved and simplified multi-scale pooling feature extraction module is integrated into the backbone network to enhance feature perception of cells at different scales. The multi-scale pooling feature extraction module is SGPM.
[0012] S3: Multi-scale spatial information acquisition, a channel-guided multi-scale spatial attention module in the neck network, which improves the model's ability to distinguish overlapping cells and background interference. The channel-guided multi-scale spatial attention module is the CGM-SA module.
[0013] S4: The feature fusion neck module receives and fuses multi-scale features processed by the multi-scale pooling feature extraction module and the multi-scale spatial attention module;
[0014] S5: Connected to the output of the detection head module and the feature fusion neck module, it is used to predict the location bounding box and category of lesion cells based on the fused features;
[0015] S6: Model training, input cervical cell images into the improved YOLOv10 network for training, and save the optimal model;
[0016] S7: Model inference and detection, using the optimal model for prediction, saving the results, and comparing the results.
[0017] Furthermore, the SGPM module consists of an identity connection path used to preserve detailed information of the original feature map.
[0018] The SGPM module includes at least two max pooling paths, each containing a max pooling layer with a pooling kernel of a different fixed size.
[0019] A feature splicing unit is used to splice the output feature maps of the identity connection path and the at least two max pooling paths along the channel dimension to form an aggregated feature.
[0020] Furthermore, the application process of the SGPM module includes at least the following steps:
[0021] First, input the feature map. ,in, For batch size, Input the number of channels. The height and width of the feature map;
[0022] The number of input channels is halved by using a 1×1 convolution, reducing subsequent computation and extracting compact feature representations;
[0023]
[0024] Then, multi-scale pooling feature extraction is performed, through continuous... The submax pooling operation, while maintaining the feature map spatial size, obtains multi-scale contextual information. This represents the number of pooling operations;
[0025]
[0026]
[0027] in, For maximum pool operation, It is the kernel size for pooling operations;
[0028] Then, global pooling enhancement is performed, and global max pooling is applied to the feature map to obtain the global response of the entire image. Then, it is expanded back to the spatial size of the original feature map so that global information and local multi-scale features are fused in the same spatial dimension.
[0029]
[0030]
[0031] Then, feature concatenation and fusion are performed. The original, pooled, multi-scale and global enhancement features are concatenated in the channel dimension, and then the information of each scale is fused by 1×1 convolution to obtain a unified output feature map.
[0032]
[0033]
[0034] Finally, optional channel attention enhancement is performed, which selectively applies channel attention weighting to the output features to highlight key information and improve the performance of target cell detection.
[0035]
[0036] Summary of formula:
[0037]
[0038] in, For channel attention module, For the first Second-most pooling This is the extended global pooling feature.
[0039] Furthermore, the CGM-SA module employs a dual-cascaded weighted collaborative mechanism, namely, a serial attention application process with residual connections, to refine the channel dimension and spatial dimension of the input feature map in sequence.
[0040] The CGM-SA module consists of a channel attention module and a multi-scale spatial attention module connected in sequence.
[0041] The CGM-SA module first generates guiding weights through a channel attention module to perform preliminary screening of input features; then, the screened features are fed into a multi-scale spatial attention module to learn their spatial saliency, thereby achieving secondary precise enhancement of the feature map.
[0042] Furthermore, the CGM-SA module data processing flow includes the following steps:
[0043] First, variable conventions are defined, and the input feature map is used. Simultaneously, two parallel branches are fed in to generate Query and Key vectors, where, For batch size, For the number of channels, The height and width of the feature map.
[0044] Then space pooling is performed:
[0045] Value branch: Performs channel dimensionality reduction on the input features through convolution to reduce computation and extract a compact representation;
[0046]
[0047]
[0048] Query branch: Performs channel dimensionality reduction on the input features through convolution, reducing computation and extracting a compact representation;
[0049]
[0050]
[0051] Feature aggregation: Channel dimensionality reduction of input features is performed through convolution to reduce computation and extract compact representations; This represents the attention weight tensor obtained after convolution and normalization operations on the input features. The subscript... This indicates that the attention weight corresponds to the dimensionality-reduced channel space (C / 2 or C / r); the superscript T indicates that this is a transposed vector used for matrix / vector multiplication; its size is... ;
[0052]
[0053] Channel attention generation: The aggregated features are further convolved, normalized, and activated to generate channel attention maps, which are used to measure the importance of each channel.
[0054]
[0055] Apply channel weights: Multiply the channel attention map element-wise with the input features to achieve feature enhancement along the channel dimension;
[0056]
[0057] Then channel pooling is performed:
[0058] Query features: The input feature map is first passed through a 1×1 convolutional layer, and then reshaped to obtain the feature representation used for channel relationship modeling;
[0059]
[0060] Global average pooling: The above channel features are subjected to global average pooling to obtain a compressed channel vector describing the global semantics;
[0061]
[0062]
[0063] Value features: The input features are convolved and reshaped again to obtain the channel feature distribution used to measure the importance of each spatial location;
[0064]
[0065]
[0066] Spatial attention generation: The global pooling result and spatial distribution features are combined through matrix operations to generate a spatial attention map, which is used to highlight important regions in the image;
[0067]
[0068]
[0069] Apply spatial weights: Generate a spatial attention map by performing matrix operations on the global pooling results and spatial distribution features to highlight important regions in the image;
[0070]
[0071] Then, multi-scale spatial attention enhancement is performed. After completing channel pooling and spatial pooling, the contextual information at different scales is further integrated through multi-scale feature interaction and enhancement modules, so that the model can better capture targets of different sizes and shapes and improve the robustness of feature representation.
[0072]
[0073] The final output features can be summarized as follows: channel attention is first applied to the input features, followed by spatial attention, and multi-scale information fusion is performed through an enhancement module. The overall process can be abstracted as follows: .
[0074] Furthermore, the detection head module is trained using a composite loss function, which includes a Slide Loss classification loss function for handling sample imbalance and a hybrid bounding box regression loss function for optimizing the localization of irregular shapes.
[0075] Compared with the prior art, the beneficial effects of the present invention are:
[0076] 1. Enhanced detection capability for multi-scale targets: Through a simplified global pooling information extraction module (SGPM), this invention can simultaneously capture features of cells of different sizes, significantly improving the robustness of identifying lesion cells with different morphologies and sizes.
[0077] 2. Improved detection accuracy in densely overlapping scenes: Through the channel-guided multi-scale spatial attention module (CGM-SA), key features are dynamically enhanced and background interference is suppressed, effectively "decoupling" the features of overlapping cells and significantly reducing the false negative and false positive rates in dense areas;
[0078] 3. Achieving a balance between high accuracy and high efficiency: The core modules of the model (SGPM and CGM-SA) are both designed with lightweight features, which improves detection performance without significantly increasing computational costs and the number of parameters. It is suitable for clinical screening scenarios that require high-throughput processing and has the potential for clinical promotion.
[0079] 4. Improved training stability and targeting: By adopting a customized composite loss function, it can better address the problem of sample imbalance during training and achieve more accurate localization of irregularly shaped cells. Attached Figure Description
[0080] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0081] Figure 1 This is a flowchart of the cervical lesion cell image detection method of the present invention;
[0082] Figure 2 This is a model architecture diagram of the present invention;
[0083] Figure 3 This is a simplified structural diagram of the global pooling information extraction branch module of the present invention;
[0084] Figure 4 This is a structural diagram of the channel-guided multi-scale spatial attention module of the present invention;
[0085] Figure 5 This is a graph of the Slide Loss function of the present invention;
[0086] Figure 6 This is a schematic diagram illustrating the detection of cervical lesion cells / clumps according to the present invention, wherein... Figure 6 (a) Original image of cervical lesion cells / clumps; Figure 6 (b) is a flowchart for detecting cervical lesion cells / clumps; Figure 6 (c) is a heatmap of the baseline YOLOv10 model for detecting cervical lesion cells / clumps; Figure 6 (d) is a heatmap of the improved YOLOv10 model for detecting cervical lesion cells / clumps. Detailed Implementation
[0087] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0088] Based on YOLOv10, this invention first introduces the SGPM module into the feature extraction network to enhance the ability to perceive the features of lesion cells of different sizes.
[0089] Secondly, the input features are passed through a channel-guided multi-scale spatial attention module in the neck network to improve the model's ability to distinguish overlapping cells and background interference.
[0090] Then, considering the impact of classification loss on detection performance, we propose to combine a composite loss function for optimized training to predict the location and category of lesion cells, thereby improving model performance.
[0091] It effectively solves the problems of significant polymorphism in cervical lesion cells, as well as the problem of dense cell arrangement and even overlap, and can improve the detection accuracy of cervical lesion cells / clumps.
[0092] Specifically as follows:
[0093] Please see Figures 1-2 The method for detecting exfoliated cells or clumps of cervical lesions using the YOLOv10 network model includes at least the following steps:
[0094] S1: Construct a basic network model, using YOLOv10 as the detection framework and as the backbone network of the overall structure;
[0095] S2: Feature extraction. An improved and simplified multi-scale pooling feature extraction module is integrated into the backbone network to enhance the feature perception of cells at different scales. The multi-scale pooling feature extraction module is the SGPM module.
[0096] S3: Multi-scale spatial information acquisition. In the neck network, the channel-guided multi-scale spatial attention module improves the model's ability to distinguish overlapping cells and background interference. The channel-guided multi-scale spatial attention module is the CGM-SA module.
[0097] S4: The feature fusion neck module receives and fuses multi-scale features processed by the multi-scale pooling feature extraction module and the multi-scale spatial attention module;
[0098] S5: Connected to the output of the detection head module and the feature fusion neck module, it is used to predict the location bounding box and category of lesion cells based on the fused features;
[0099] S6: Model training, input cervical cell images into the improved YOLOv10 network for training, and save the optimal model;
[0100] Use the training set images for training, train the dataset using the improved YOLOv10 model, view the training results, and save the model weights.
[0101] Table 4 presents a set of comparative experiments designed to compare the effectiveness of the improved YOLOv10 model with the baseline YOLOv10 model.
[0102] Table 4. Experimental results of the baseline YOLOv10 and the final improved YOLOv10.
[0103] Methods mAP50(%) mAP95(%) YOLOv10 60.4 32.3 YOLOv10_GAFE 62.7 34.3
[0104] Experimental results show that the improved YOLOv10 improves the mAP50 by 2.3% and the mAP90 by 2.0% compared to the baseline YOLOv10.
[0105] S7: Model inference and detection, using the optimal model for prediction, saving the results, and comparing the results.
[0106] The effect is as follows Figure 6 As shown, save the result image.
[0107] See Figure 3 The SGPM module consists of an identity connection path used to preserve detailed information of the original feature map.
[0108] The SGPM module includes at least two max pooling paths, each containing a max pooling layer with a different fixed-size pooling kernel;
[0109] A feature concatenation unit is used to concatenate the output feature maps of the identity connection path and at least two max pooling paths along the channel dimension to form aggregated features;
[0110] In this embodiment, a three-way max pooling structure is used to enhance multi-scale sensing capabilities;
[0111] The first branch is the identity mapping path, which preserves the original information of the input features;
[0112] The second branch uses a 3×3 max pooling layer to extract local contextual features;
[0113] The third branch uses a max pooling layer with a kernel size of 5×5 to extract a wider range of contextual features;
[0114] All branch outputs are spliced or merged along the channel dimension to form a scale-invariant multi-scale feature map.
[0115] In different embodiments, the SGPM module may also contain one identity path and three parallel max-pooling paths with pooling kernel sizes of 3x3, 5x5, and 7x7, respectively. The outputs of the four paths are concatenated and then fused through a convolutional layer. This process ensures that the output feature map contains both detailed information about small cells and global information about large cells.
[0116] Table 1 below presents a set of comparative experiments designed to examine the performance of the multi-scale information extraction branch module in feature extraction, specifically:
[0117] Table 1 shows the results of feature extraction using a simplified global pooling information extraction branch module.
[0118] Methods SGPM mAP50(%) mAP95(%) YOLOv10 N 60.4 32.3 YOLOv10 Y 61.5 32.7
[0119] In the experimental results in Table 1, using the simplified global pooling branch for feature extraction improved mAP50 by 1.1% and mAP95 by 0.4% compared to the original model.
[0120] The application process of the SGPM module includes at least the following steps:
[0121] First, input the feature map. ,in, Where C is the batch size and C is the number of input channels. The height and width of the feature map;
[0122] The number of input channels is halved by using a 1×1 convolution, reducing subsequent computation and extracting compact feature representations;
[0123]
[0124] Then, multi-scale pooling feature extraction is performed, through continuous... The submax pooling operation, while maintaining the feature map spatial size, obtains multi-scale contextual information. This represents the number of pooling operations;
[0125]
[0126]
[0127] in, For maximum pool operation, It is the kernel size for pooling operations;
[0128] Then, global pooling enhancement is performed, and global max pooling is applied to the feature map to obtain the global response of the entire image. Then, it is expanded back to the spatial size of the original feature map so that global information and local multi-scale features are fused in the same spatial dimension.
[0129]
[0130]
[0131] Then, feature concatenation and fusion are performed. The original, pooled, multi-scale and global enhancement features are concatenated in the channel dimension, and then the information of each scale is fused by 1×1 convolution to obtain a unified output feature map.
[0132]
[0133]
[0134] Finally, optional channel attention enhancement is performed, which selectively applies channel attention weighting to the output features to highlight key information and improve the performance of target cell detection.
[0135]
[0136] Summary of formula:
[0137]
[0138] in, Channel attention module (optional) For the first Second-most pooling This is the extended global pooling feature.
[0139] See Figure 4 The CGM-SA module employs a dual-cascaded weighted collaborative mechanism, namely a serial attention application process with residual connections, to refine the channel and spatial dimensions of the input feature map in sequence.
[0140] The CGM-SA module consists of sequentially connected channel attention modules and multi-scale spatial attention modules;
[0141] The CGM-SA module first generates guiding weights through a channel attention module to perform preliminary screening of input features. Then, the screened features are fed into a multi-scale spatial attention module to learn their spatial saliency, thereby achieving secondary precise enhancement of the feature map.
[0142] CGM-SA refines the fused features through parallel channels and spatial attention units, enhancing the saliency of lesion features and suppressing noise.
[0143] In Table 2 below, this application designed a set of comparative experiments to examine the performance of the channel-guided multi-scale spatial attention module in feature extraction, specifically as follows:
[0144] Table 2. Results of feature extraction using a channel-guided multi-scale spatial attention module.
[0145] Methods SGPM mAP50(%) mAP95(%) YOLOv10 N 60.4 32.3 YOLOv10 Y 62.5 34.1
[0146] In the experimental results in Table 2, the channel-guided multi-scale spatial attention module branch for feature extraction improved the performance of the original model by 2.1% on mAP50 and 1.8% on mAP95. Among them, the channel-guided multi-scale spatial attention module performed better in mAP95 because it enabled the model to focus on key features and obtain more refined results.
[0147] The CGM-SA module data processing workflow includes the following steps:
[0148] First, variable conventions are defined, and the input feature map is used. Simultaneously, two parallel branches are fed in to generate Query and Key vectors, where, For batch size, For the number of channels, The height and width of the feature map.
[0149] Then space pooling is performed:
[0150] Value branch: Performs channel dimensionality reduction on the input features through convolution to reduce computation and extract a compact representation;
[0151]
[0152]
[0153] Query branch: Performs channel dimensionality reduction on the input features through convolution, reducing computation and extracting a compact representation;
[0154]
[0155]
[0156] Feature aggregation: Channel dimensionality reduction of input features is performed through convolution to reduce computation and extract compact representations; This represents the attention weight tensor obtained after convolution and normalization operations on the input features. The subscript... This indicates that the attention weight corresponds to the dimensionality-reduced channel space (C / 2 or C / r); the superscript T indicates that this is a transposed vector used for matrix / vector multiplication; its size is... ;
[0157]
[0158] Channel attention generation: The aggregated features are further convolved, normalized, and activated to generate channel attention maps, which are used to measure the importance of each channel.
[0159]
[0160] Apply channel weights: Multiply the channel attention map element-wise with the input features to achieve feature enhancement along the channel dimension;
[0161]
[0162] Then channel pooling is performed:
[0163] Query features: The input feature map is first passed through a 1×1 convolutional layer, and then reshaped to obtain the feature representation used for channel relationship modeling;
[0164]
[0165] Global average pooling: The above channel features are subjected to global average pooling to obtain a compressed channel vector describing the global semantics;
[0166]
[0167]
[0168] Value features: The input features are convolved and reshaped again to obtain the channel feature distribution used to measure the importance of each spatial location;
[0169]
[0170]
[0171] Spatial attention generation: The global pooling result and spatial distribution features are combined through matrix operations to generate a spatial attention map, which is used to highlight important regions in the image;
[0172]
[0173]
[0174] Apply spatial weights: Generate a spatial attention map by performing matrix operations on the global pooling results and spatial distribution features to highlight important regions in the image;
[0175]
[0176] Then, multi-scale spatial attention enhancement is performed. After completing channel pooling and spatial pooling, the contextual information at different scales is further integrated through multi-scale feature interaction and enhancement modules, so that the model can better capture targets of different sizes and shapes and improve the robustness of feature representation.
[0177]
[0178] The final output features can be summarized as follows: channel attention is first applied to the input features, followed by spatial attention, and multi-scale information fusion is performed through an enhancement module. The overall process can be abstracted as follows: .
[0179] The detection head module is trained using a composite loss function, which includes a Slide Loss classification loss function to handle sample imbalance and a hybrid bounding box regression loss function to optimize the localization of irregular shapes.
[0180] The Slide Loss classification loss function employs a non-linear sliding weight mechanism based on predicted probabilities, and its loss weight function is as follows:
[0181] W(p) = exp(-b × |p - μ|)
[0182] in:
[0183] p represents the predicted probability.
[0184] μ is the classification threshold for dynamic learning.
[0185] b is a hyperparameter controlling the sliding curvature.
[0186] By assigning higher weights to predicted values near μ, attention is increased to hard-to-classify samples.
[0187] The detection head module analyzes these feature maps and ultimately outputs the class confidence score and bounding box coordinates for each detected lesion cell. During the training phase, this module employs methods such as... Figure 5 As shown, the Slide Loss function and the Hybrid Bounding Box Loss are supervised to optimize the final detection performance.
[0188] In Table 3, this application designs a set of comparative experiments to verify the effectiveness of the improved classification loss function Slide Loss in the model.
[0189] Table 3 shows the results of using slide loss for model testing.
[0190] Methods CLS mAP50(%) mAP95(%) YOLOv10 BCE loss 60.4 32.3 YOLOv10 Slide loss 61.9 33.5
[0191] In the experimental results, using the Slide loss function for model detection improved the performance by 1.5% on mAP50 and 1.2% on mAP95 compared to the original model.
[0192] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model, characterized in that: At least the following steps are included: S1: Construct a basic network model, using YOLOv10 as the detection framework and as the backbone network of the overall structure; S2: Feature extraction, an improved and simplified multi-scale pooling feature extraction module is integrated into the backbone network to enhance feature perception of cells at different scales. The multi-scale pooling feature extraction module is the SGPM module. S3: Multi-scale spatial information acquisition, a channel-guided multi-scale spatial attention module in the neck network, which improves the model's ability to distinguish overlapping cells and background interference. The channel-guided multi-scale spatial attention module is the CGM-SA module. S4: The feature fusion neck module receives and fuses multi-scale features processed by the multi-scale pooling feature extraction module and the multi-scale spatial attention module; S5: Connected to the output of the detection head module and the feature fusion neck module, it is used to predict the location bounding box and category of lesion cells based on the fused features; S6: Model training, input cervical cell images into the improved YOLOv10 network for training, and save the optimal model; S7: Model inference and detection, using the optimal model for prediction, saving the results, and comparing the results.
2. The method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model according to claim 1, characterized in that: The SGPM module consists of an identity connection path used to preserve detailed information of the original feature map. The SGPM module includes at least two max pooling paths, each containing a max pooling layer with a pooling kernel of a different fixed size. A feature splicing unit is used to splice the output feature maps of the identity connection path and the at least two max pooling paths along the channel dimension to form an aggregated feature.
3. The method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model according to claim 2, characterized in that: The application process of the SGPM module includes at least the following steps: First, input the feature map. ,in, For batch size, Input the number of channels. The height and width of the feature map; The number of input channels is halved by using a 1×1 convolution, reducing subsequent computation and extracting compact feature representations; ； Then, multi-scale pooling feature extraction is performed, through continuous... The submax pooling operation obtains multi-scale contextual information while keeping the feature map spatial size unchanged, where t is the number of pooling operations; ； ； in, For maximum pool operation, It is the kernel size for pooling operations; Then, global pooling enhancement is performed, and global max pooling is applied to the feature map to obtain the global response of the entire image. Then, it is expanded back to the spatial size of the original feature map so that global information and local multi-scale features are fused in the same spatial dimension. ； ； Then, feature concatenation and fusion are performed. The original, pooled, multi-scale and global enhancement features are concatenated in the channel dimension, and then the information of each scale is fused by 1×1 convolution to obtain a unified output feature map. ； ； Finally, optional channel attention enhancement is performed, which selectively applies channel attention weighting to the output features to highlight key information and improve the performance of target cell detection. ； Summary of formula: ； in, For channel attention module, For the first Second-most pooling This is the extended global pooling feature.
4. The method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model according to claim 1, characterized in that: The CGM-SA module employs a dual-cascaded weighted collaborative mechanism, namely a serial attention application process with residual connections, to refine the channel dimension and spatial dimension of the input feature map in sequence. The CGM-SA module consists of a channel attention module and a multi-scale spatial attention module connected in sequence. The CGM-SA module first generates guiding weights through a channel attention module to perform preliminary screening of input features; then, the screened features are fed into a multi-scale spatial attention module to learn their spatial saliency, thereby achieving secondary precise enhancement of the feature map.
5. The method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model according to claim 4, characterized in that: The data processing flow of the CGM-SA module includes the following steps: First, variable conventions are defined, and the input feature map is used. Simultaneously, two parallel branches are fed in to generate Query and Key vectors, where, For batch size, For the number of channels, The height and width of the feature map; Then space pooling is performed: Value branch: Performs channel dimensionality reduction on the input features through convolution to reduce computation and extract a compact representation; ； ； Query branch: Performs channel dimensionality reduction on the input features through convolution, reducing computation and extracting a compact representation; ； ； Feature aggregation: Channel dimensionality reduction of input features is performed through convolution to reduce computation and extract compact representations; This represents the attention weight tensor obtained after convolution and normalization operations on the input features. The subscript... This indicates that the attention weight corresponds to the dimension-reduced channel space; the superscript T indicates that this is a transposed vector used for matrix / vector multiplication; its size is... ; ； Channel attention generation: The aggregated features are further convolved, normalized, and activated to generate channel attention maps, which are used to measure the importance of each channel. ； Apply channel weights: Multiply the channel attention map element-wise with the input features to achieve feature enhancement along the channel dimension; ； Then channel pooling is performed: Query features: The input feature map is first passed through a 1×1 convolutional layer, and then reshaped to obtain the feature representation used for channel relationship modeling; ； Global average pooling: The above channel features are subjected to global average pooling to obtain a compressed channel vector describing the global semantics; ； ； Value features: The input features are convolved and reshaped again to obtain the channel feature distribution used to measure the importance of each spatial location; ； ； Spatial attention generation: The global pooling result and spatial distribution features are combined through matrix operations to generate a spatial attention map, which is used to highlight important regions in the image; ； ； Apply spatial weights: Generate a spatial attention map by performing matrix operations on the global pooling results and spatial distribution features to highlight important regions in the image; ； Then, multi-scale spatial attention enhancement is performed. After completing channel pooling and spatial pooling, the contextual information at different scales is further integrated through multi-scale feature interaction and enhancement modules, so that the model can better capture targets of different sizes and shapes and improve the robustness of feature representation. ； The final output features can be summarized as follows: channel attention is first applied to the input features, followed by spatial attention, and multi-scale information fusion is performed through an enhancement module. The overall process can be abstracted as follows: .
6. The method for detecting exfoliated cells or clumps of cervical lesions based on the YOLOv10 network model according to claim 1, characterized in that: The detection head module is trained using a composite loss function, which includes a Slide Loss classification loss function for handling imbalanced samples and a hybrid bounding box regression loss function for optimizing the localization of irregular shapes.