Cervical cytopathy detection method based on hypergraph convolutional network

Abstract

The invention discloses a cervical cytopathy detection method based on a hypergraph convolutional network. The method comprises the following steps: performing sliding window slicing processing, unsupervised image decomposition and dyeing normalization on a cervical cytopathy image, generating normalized image input, and constructing an image input sample set; a target detection model is constructed based on YOLO11, a block-level feature extraction network (BBMM) module is embedded to enhance the perception ability, and a sparse attention module is adopted to perform key region feature enhancement; constructing a hypergraph neural network HGNN module based on a Patch-level relationship, and extracting a structural relationship between cells; integrating an uncertainty quantification mechanism, and generating a confidence thermodynamic diagram; a front-end and rear-end separated diagnosis platform is built, image uploading, detection result display, frame selection correction and interactive management are supported, and whole-process auxiliary diagnosis is achieved; the method has the advantages of high detection accuracy, high interpretability, flexible deployment and the like, and is suitable for intelligent early screening and clinical auxiliary diagnosis scenes of cervical cytopathy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for detecting cervical cell lesions based on a hypergraph convolutional network. Background Art

[0002] Detection of cervical cytological changes is an important tool for cervical cancer screening and prevention. However, the widely used Pap smear and liquid-based thin-layer cytology (LCT) techniques still rely primarily on manual interpretation by pathologists. These methods are not only time-consuming and labor-intensive, but also susceptible to subjective factors such as physician experience and fatigue, potentially leading to missed diagnoses, misdiagnoses, and high false-negative rates.

[0003] In addition, whole slide image (WSI) pathology sections face many challenges due to their ultra-high resolution, sparse cell distribution, and non-uniform staining methods. Existing AI solutions still have certain limitations, such as difficulty in balancing the microscopic characteristics of cells and the macroscopic structure of tissues, insufficient adaptability to different laboratory staining methods, and inability to effectively model the correlation information between cells. There is an urgent need for more intelligent and efficient solutions. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for detecting cervical cell lesions based on a hypergraph convolutional network, so as to solve the problems of heavy burden on pathologists in manual reading of films, high diagnostic error rate due to subjective errors, inability to fully input WSI full-slide images into the detection system, inability of existing systems to take into account both cell microscopic characteristics and tissue macroscopic structures, insufficient adaptability to different staining methods, and inability to effectively model intercellular relationships.

[0005] Technical solution: The present invention provides a method for detecting cervical cell lesions based on a hypergraph convolutional network, comprising the following steps:

[0006] (1) Collect cervical cell pathology images, perform staining normalization through unsupervised color decomposition and reconstruction, and unify the image color style;

[0007] (2) Slicing the normalized image using a sliding window strategy;

[0008] (3) Build a target detection model based on YOLO11. The block-level feature extraction network BBMM module is embedded in the backbone network. The BBMM module consists of a 1×1 standard convolution branch, a deformable convolution branch, and a C3K2 residual module branch connected in parallel. The three-way output features are channel-wise concatenated to form a fusion feature. At the same time, a sparse attention module is introduced between the backbone network and the detection head.

[0009] (4) Extract the output candidate lesion area features, build a hypergraph neural network (HGNN) based on patch-level relationships, and construct a hyperedge set through a distance threshold;

[0010] (5) The feature map output by HGNN is input into the YOLO11 detection branch, and the bounding box prediction and classification are performed through the dynamic prototype matching detection head. The differentiable non-maximum suppression mechanism dNMS is used to remove duplicate candidate boxes and the confidence heat map is generated in combination with the MCDropout method.

[0011] (6) Display the test results through the diagnostic platform, supporting image uploading, result visualization and interactive correction.

[0012] Furthermore, the staining normalization process in step (1) is specifically as follows: collecting a cervical cell pathology image stained by Papanicolaou or HE, including normal cells, inflammatory cells and pathological cell areas; mapping the color distribution of the original image to the standard template staining distribution, and outputting a color consistency image.

[0013] Furthermore, in step (2), the window size is 614×614 pixels, the sliding step is 256 pixels, and the boundary area is filled with mirror images to generate image blocks of uniform size.

[0014] Furthermore, the slicing process also includes: for the labeled samples, the center of the labeled box is used as the reference, and a random offset of 104 / 208 / 312 pixels is made to the upper left as the cropping starting point; a 614×614 pixel area is expanded to the lower right to generate three local image blocks, and out-of-bounds samples are automatically discarded.

[0015] Furthermore, in step (3), the C3K2 residual module consists of three serially connected 3×3 convolutional layers, each followed by batch normalization and SiLU activation function; the deformable convolution branch uses a deformable convolution kernel size of 3×3.

[0016] Furthermore, in step (3), the sparse attention module formula is as follows:

[0017]

[0018] Where Q and K are the query and key feature matrices respectively, M is the sparse mask, and d is the scaling factor. Only the key regions are enhanced.

[0019] Furthermore, in step (4), given the feature matrix X of the given vertex and the adjacency matrix H of the hypergraph, assume that D v and D e are the diagonal matrices of vertices and hyperedges respectively, then the hypergraph convolution is expressed as:

[0020]

[0021] in, Calculate the normalized adjacency matrix, It means that vertex features are aggregated through hyperedges to capture high-order relationships between vertices. Θ is a learnable parameter matrix used to transform the aggregated features.

[0022] Furthermore, step (5) includes the following steps:

[0023] (1) The feature map enhanced by the HGNN module structure is fused with the multi-scale features output by the backbone network and sent to the decoding network for cross-scale feature integration; the decoding network is designed based on the weighted bidirectional feature pyramid network BiFPN, and its fusion process is expressed as:

[0024]

[0025] in, represents the input features of the lth layer, and Represents upsampling and downsampling operations respectively, w k is the learnable fusion weight, σ is the normalization function, For channel splicing operation, It is a depth-wise separable convolution operation.

[0026] (2) Through a hypergraph-guided cross-layer feature aggregation mechanism, the global high-order semantic features output by HGNN are fused with the local features of layers C3 to C5 in the YOLO11 backbone network. The fusion process includes linear transformation, deformable convolution operations, and residual connections to enhance the consistency and spatial adaptability of multi-level semantic expressions.

[0027] (3) An adaptive spatial reweighting module is introduced to dynamically generate a position-sensitive weight map M(x,y) based on the spatial attention mechanism, and perform pixel-by-pixel enhancement on the feature map F(x,y), which is expressed as:

[0028] F′(x,y)=M(x,y)·F(x,y)

[0029] Among them, M(x,y) is the spatial response map obtained by convolutional layer learning, which is used to enhance the detection sensitivity of key areas;

[0030] (4) The enhanced feature map is input into the dynamic prototype matching detection head module to perform bounding box regression and multi-category classification prediction of the lesion area; the detection loss function introduces the prototype learning mechanism and is defined as follows:

[0031]

[0032] in, Represents the positive and negative sample prototype libraries, f iis the feature to be classified, and Δ is the discrimination boundary threshold, which is used to improve the robustness of classification discrimination.

[0033] (5) Differentiable non-maximum suppression mechanism dNMS is used to remove duplicate candidate frames. The calculation process is as follows:

[0034]

[0035] Among them, s i is the confidence of the i-th candidate box, IoU(b i ,b j ) represents the degree of overlap with other candidate boxes, and k is an adjustment parameter; overlapping but credible prediction areas are retained through soft suppression.

[0036] (6) Combined with the MC Dropout method, the detection model is forward sampled multiple times in the inference stage, the confidence distribution of each candidate area is statistically analyzed, and an uncertainty heat map is generated to mark the confidence credibility of the model in the prediction area.

[0037] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the methods when executing the program.

[0038] The computer-readable storage medium of the present invention stores a computer program, which implements the steps of any one of the methods when executed by a processor.

[0039] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the present invention can directly input WSI whole slide images into the system, the system can process cervical cell images using different staining methods and maintain a relatively uniform presentation, the system can take into account both the microscopic characteristics of cells and the macroscopic structure of tissues for all-round interpretation, and the system can output the judgment credibility for reference. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the overall flow chart of the present invention;

[0041] Figure 2 Schematic diagram of the backbone network structure of the present invention;

[0042] Figure 3 Schematic diagram of the block-level feature extraction network (BBMM) module structure of the present invention;

[0043] Figure 4 Schematic diagram of the detection process of the present invention. DETAILED DESCRIPTION

[0044] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0045] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting cervical cell lesions based on a hypergraph convolutional network, comprising the following steps:

[0046] Step 1: Acquire cervical cell pathology images, perform staining normalization processing, and unify the image color style; including the following steps:

[0047] Step 1.1: Acquire cervical cell pathology slide images. The image sources include digital slide images obtained from clinical settings of medical institutions and public medical image datasets. The image types are cell images stained with Papanicolaou or hematoxylin and eosin, and the images contain normal cells, inflammatory cells, and different levels of pathological cells.

[0048] Step 1.2: Perform color normalization on the original image using an unsupervised color decomposition and reconstruction method to unify the coloring based on color channel space conversion; adjust the color distribution of the original image to the color distribution of the standard template image to achieve image style normalization, and output image data with high color consistency as the standard input image set for subsequent analysis.

[0049] Step 2: Slice the normalized image using a sliding window strategy to construct an image input sample set; this includes the following steps:

[0050] Step 2.1: Input the stained and normalized cervical cell pathology image into the image slicing module, set the sliding window parameters, the window size is 614×614 pixels, the sliding step is 256 pixels, and the image is cropped starting from the upper left corner in row priority order to generate a local image block;

[0051] Step 2.2: Perform integrity processing on the boundary areas during the cropping process. When the sliding window exceeds the edge of the original image, the image edge area is expanded using a mirror filling method to ensure that all image blocks have the same size and there are no empty areas.

[0052] Step 2.3: For the image sample with bounding box annotation, use the center point of the annotation box as a reference, randomly offset 104 pixels, 208 pixels, and 312 pixels in the upper left direction as the cropping starting point, and expand the 614×614 pixel area in the lower right direction to generate three local image patches; if the cropping window exceeds the boundary, the sample is automatically discarded;

[0053] Step 2.4: Uniformly number the cropped image blocks, record their spatial position information in the original image and the corresponding annotation labels, and use them to construct the image input sample set, which serves as the standard training data for the target detection model.

[0054] Step 3: Build a target detection model based on YOLO11, embed the block-level feature extraction network (BBMM) module for multi-scale perceptual feature extraction, and use the sparse attention module to enhance the key area response; including the following steps:

[0055] Step 3.1: Build an object detection model based on the YOLO11 architecture. The model consists of a backbone feature extraction network, a feature fusion network, and a multi-scale detection head.

[0056] Step 3.2: Embed a block-level feature extraction network (BBMM) module in the backbone feature extraction network. The BBMM module consists of three parallel branches: 1×1 standard convolution, deformable convolution, and C3K2 residual module. The output features of the three are concatenated in the channel dimension to form the final feature representation. The fusion formula is:

[0057] F BBMM =Concat(F 1×1 ,\F DCN ,\F C3K2 )

[0058] This structure is used to improve the model's ability to model cervical lesion areas of different scales.

[0059] Step 3.3: Introduce a sparse attention module between the backbone and the detection head, and use the channel selection and local mask generation mechanism to construct the sparse attention matrix A, which is defined as:

[0060]

[0061] Where Q and K are the query and key feature matrices respectively, M is the sparse mask, and d is the scaling factor. Only the key regions are enhanced.

[0062] Step 4: After extracting the features of the candidate lesion area, a hypergraph neural network based on patch-level relationships is constructed to model the semantic relationship between cell structures. This includes the following steps:

[0063] Step 4.1: To model the neighborhood relationship in the hypergraph module, construct the hyperedge set E by the distance threshold λ. Specifically, for each feature point x u , find all feature points whose distance is less than ∈ and compare them with x v Together they form a hyperedge, and the hyperedge e can be expressed as:

[0064] e={u|‖x u -x v ‖2<∈,u∈V}

[0065] Where ||·||2 represents the Euclidean norm, and all such hyperedges constitute the hyperedge set E. The incidence matrix H of the hypergraph G = (V, E) is defined as:

[0066]

[0067] Step 4.2: In a hypergraph convolutional network, as the number of layers increases, information may be gradually lost during propagation. By introducing residual connections, the model can directly pass input information to the output, thus avoiding information loss. To transmit high-order information on the hypergraph structure, a spatial hypergraph convolution with residual connections is used. The calculation process is as follows:

[0068]

[0069] where N V (e) represents the vertex neighborhood of hyperedge e, Θ e is a trainable parameter. For a given vertex feature matrix X and a hypergraph adjacency matrix H, assume that D v and D e are the diagonal matrices of vertices and hyperedges respectively, and the hypergraph convolution can be expressed as:

[0070]

[0071] in, Calculate the normalized adjacency matrix, It means that the vertex features are aggregated through hyperedges to capture the high-order relationship between vertices. Θ is a learnable parameter matrix used to transform the aggregated features and enhance the expressiveness of the model. X is then introduced as a residual connection to retain the original feature information and avoid information loss.

[0072] Step 5: Send the structure-enhanced feature map to the YOLO11 detection branch for bounding box prediction and category classification, and generate a confidence heat map by combining the uncertainty quantification mechanism; including the following steps:

[0073] Step 5.1: Fuse the feature map enhanced by the HGNN module structure with the multi-scale features output by the backbone network and send it to the decoding network for cross-scale feature integration. The decoding structure is designed based on the weighted bidirectional feature pyramid network (BiFPN), and its fusion process is expressed as:

[0074]

[0075] in, represents the input features of the lth layer, and Represents upsampling and downsampling operations respectively, w k is the learnable fusion weight, σ is the normalization function, For channel splicing operation, It is a depth-wise separable convolution operation.

[0076] Step 5.2: Through the hypergraph-guided cross-layer feature aggregation mechanism, the global high-order semantic features output by HGNN are fused with the local features of layers C3 to C5 in the YOLO11 backbone network. The fusion process includes linear transformation, deformable convolution operation and residual connection to enhance the consistency and spatial adaptability of multi-level semantic expression.

[0077] Step 5.3: Introduce the adaptive spatial reweighting module to dynamically generate a position-sensitive weight map M(x,y) based on the spatial attention mechanism, and perform pixel-by-pixel enhancement on the feature map F(x,y), which is expressed as:

[0078] F′(x,y)=M(x,y)·F(x,y)

[0079] Among them, M(x,y) is the spatial response map obtained by convolutional layer learning, which is used to enhance the detection sensitivity of key areas.

[0080] Step 5.4: Input the enhanced feature map into the dynamic prototype matching detection head module to perform bounding box regression and multi-class classification prediction of the lesion area; the detection loss function introduces the prototype learning mechanism and is defined as follows:

[0081]

[0082] in, Represents the positive and negative sample prototype libraries, f i is the feature to be classified, and Δ is the discrimination boundary threshold, which is used to improve the robustness of classification discrimination.

[0083] Step 5.5: Use the differentiable non-maximum suppression mechanism (dNMS) to remove duplicate candidate boxes. The calculation process is as follows:

[0084]

[0085] Among them, s i is the confidence of the i-th candidate box, IoU(b i ,b j ) represents the degree of overlap with other candidate boxes, and k is an adjustment parameter; overlapping but credible prediction areas are retained through soft suppression.

[0086] Step 5.6: Combined with the MC Dropout method, the detection model is forward sampled multiple times during the inference phase to calculate the confidence distribution of each candidate region and generate an uncertainty heat map to mark the confidence level of the model in the predicted region and assist doctors in making diagnostic judgments.

[0087] Step 6: Display the test results and visualization content through the diagnostic system platform, supporting image upload, test echo, and interaction.

[0088] In order to verify the practical effect of the intelligent detection method for cervical cell lesions described in the present invention, the following experiments were designed and completed:

[0089] (1) Dataset settings

[0090] The cervical cell image data used in the experiment comes from multiple pathology image datasets, including ComparisonDetector and CRIC. The data is divided into training, validation, and test sets using a 6:2:2 ratio. All images undergo staining normalization, sliding window slicing, and data augmentation, with the input image size uniformly set to 640×640 pixels.

[0091] (2) Training configuration and system environment

[0092] The experimental system was built on a workstation with 16GB of memory, equipped with an NVIDIA GeForce RTX 4060 graphics card, and supporting CUDA11.3. The training framework used PyTorch, the total training batch size was set to 200, the batch size was 16, the optimizer used was Adam, and the initial learning rate and momentum coefficient were both 0.937.

[0093] (3) Training process and convergence curve

[0094] During the training process, the various loss function indicators (Box Loss, Cls Loss, DFL) continued to decline and eventually stabilized, while the Precision, Recall, and mAP values ​​increased rapidly and remained high.

[0095] (4) Evaluation index results

[0096] The improved model achieved the following performance indicators on the test set:

[0097] mAP@0.5≈0.600

[0098] mAP@0.5:0.95≈0.350

[0099] Precision and Recall are stable at a high level, indicating that the model has good positioning accuracy and recall capabilities.

[0100] (5) PR curve and category performance overview

[0101] Precision-Recall curves for multiple abnormal cell types based on the improved model were plotted. The results show that the model demonstrates high recognition performance for cell types such as candida, herps, and flora, with mAP@0.5 exceeding 0.75. Overall, the model achieved an average precision of 0.600, demonstrating its applicability in multi-type lesion detection tasks.

[0102] Precision-Recall curves for a model trained on another subset of tasks. Overall performance is excellent, with mAP@0.5 exceeding 0.88 for almost all categories. This demonstrates the model's strong generalization and robustness, making it particularly well-suited for analyzing large, clearly stained samples. This experimental result further demonstrates the design's adaptability to changes in data distribution.

[0103] (6) Cross-method performance comparison

[0104] On three different datasets, the model proposed in the present invention is comprehensively compared with existing mainstream detection algorithms, including classic two-stage detectors (Faster R-CNN, Cascade R-CNN), single-stage detectors (YOLO X, RetinaNet), Sparse R-CNN based on sparse representation, and the latest Transformer architecture RT-DETR. The evaluation indicators include AP, AP@0.5 and AR. As shown in Table 1, in all sub-indicators (especially AP@0.5 and AR), the present invention maintains a leading position on all three datasets. Among them, AP@0.5 can reach 97.5, indicating that the present invention has a very strong lesion target positioning capability; and AR reaches 91.4 (set A), indicating that the model has good recall capability, which is particularly suitable for high recall demand scenarios in lesion detection.

[0105] Table 1 Comparison of performance indicators of this model and other models tested on different datasets

[0106]

[0107] (7) Ablation experiment analysis

[0108] To verify the contribution of each module to the overall detection performance, a systematic ablation experiment was conducted on three independent datasets (A / B / C). The results are shown in Table 2:

[0109] Table 2 Ablation test results of this model

[0110]

[0111] After the introduction of the BBMM module, the model improved in all indicators, indicating that the parallel multi-convolutional structure can enhance the model's ability to model complex cell structures; after the introduction of the HGNN module, the model achieved significant improvements in both AP and AR, especially in dataset C, where the AP increased to 38.9, verifying the importance of modeling intercellular structural relationships for the identification of highly heteromorphic lesions; when BBMM and HGNN were introduced at the same time, the system performance reached the optimal level, with the average precision (AP) on the three datasets reaching 79.2, 29.8, and 44.6, respectively, indicating that multi-scale perception and contextual semantic modeling have a good synergistic gain effect.

Claims

1. A method for detecting cervical cell lesions based on a hypergraph convolutional network, characterized in that: The following steps are involved: (1) Collect cervical cell pathology images, perform staining normalization through unsupervised color decomposition and reconstruction, and unify the image color style; (2) Slicing the normalized image using a sliding window strategy; (3) Build a target detection model based on YOLO11. The block-level feature extraction network BBMM module is embedded in the backbone network. The BBMM module consists of a 1×1 standard convolution branch, a deformable convolution branch, and a C3K2 residual module branch connected in parallel. The three-way output features are channel-wise concatenated to form a fusion feature. At the same time, a sparse attention module is introduced between the backbone network and the detection head. (4) Extract the output candidate lesion area features, construct a hypergraph neural network (HGNN) based on patch-level relationships, and construct a hyperedge set through a distance threshold; (5) The feature map output by HGNN is input into the YOLO11 detection branch, and the bounding box prediction and classification are performed through the dynamic prototype matching detection head. The differentiable non-maximum suppression mechanism dNMS is used to remove duplicate candidate boxes and the confidence heat map is generated in combination with the MC Dropout method. (6) Display the test results through the diagnostic platform, supporting image uploading, result visualization and interactive correction.

2. The method for detecting cervical cell lesions based on a hypergraph convolutional network according to claim 1, wherein: The staining normalization process in step (1) is specifically as follows: collecting a Papanicolaou-stained or HE-stained cervical cell pathology image, which includes normal cells, inflammatory cells, and pathological cell areas; mapping the original image color distribution to the standard template staining distribution, and outputting a color consistency image.

3. The method for detecting cervical cell lesions based on a hypergraph convolutional network according to claim 1, characterized in that: In step (2), the window size is 614×614 pixels, the sliding step is 256 pixels, and the boundary area is filled with mirror images to generate image blocks of the same size.

4. The method for detecting cervical cell lesions based on a hypergraph convolutional network according to claim 1, wherein: In step (2), the slicing process also includes: for the labeled samples, randomly offset 104 / 208 / 312 pixels to the upper left with the center of the labeled box as the reference as the starting point for cropping; expand the 614×614 pixel area to the lower right to generate three local image blocks, and automatically discard out-of-bounds samples.

5. The method for detecting cervical cell lesions based on a hypergraph convolutional network according to claim 1, characterized in that: In step (3), the C3K2 residual module consists of three serially connected 3×3 convolutional layers, each followed by batch normalization and SiLU activation function; the deformable convolution branch uses a deformable convolution kernel size of 3×3.

6. The method for detecting cervical cell lesions based on a hypergraph convolutional network according to claim 1, characterized in that: In step (3), the sparse attention module formula is as follows: Where Q and K are the query and key feature matrices respectively, M is the sparse mask, and d is the scaling factor. Only the key regions are enhanced.

7. The method for detecting cervical cell lesions based on a hypergraph convolutional network according to claim 1, characterized in that: In step (4), let the feature matrix X of the given vertex and the adjacency matrix H of the hypergraph, and assume that D v and D e are the diagonal matrices of vertices and hyperedges respectively, then the hypergraph convolution is expressed as: in, Calculate the normalized adjacency matrix, It means that vertex features are aggregated through hyperedges to capture high-order relationships between vertices. Θ is a learnable parameter matrix used to transform the aggregated features.

8. The method for detecting cervical cell lesions based on a hypergraph convolutional network according to claim 1, characterized in that: Step (5) comprises the following steps: (1) The feature map enhanced by the HGNN module structure is fused with the multi-scale features output by the backbone network and sent to the decoding network for cross-scale feature integration; the decoding network is designed based on the weighted bidirectional feature pyramid network BiFPN, and its fusion process is expressed as: in, represents the input features of the lth layer, and Represents upsampling and downsampling operations respectively, w k is the learnable fusion weight, σ is the normalization function, For channel splicing operation, It is a depth-wise separable convolution operation. (2) Through a hypergraph-guided cross-layer feature aggregation mechanism, the global high-order semantic features output by HGNN are fused with the local features of layers C3 to C5 in the YOLO11 backbone network. The fusion process includes linear transformation, deformable convolution operations, and residual connections to enhance the consistency and spatial adaptability of multi-level semantic expressions. (3) An adaptive spatial reweighting module is introduced to dynamically generate a position-sensitive weight map M(x,y) based on the spatial attention mechanism, and perform pixel-by-pixel enhancement on the feature map F(x,y), which is expressed as: F′(x,y)=M(x,y)·F(x,y) Among them, M(x,y) is the spatial response map obtained by convolutional layer learning, which is used to enhance the detection sensitivity of key areas; (4) The enhanced feature map is input into the dynamic prototype matching detection head module to perform bounding box regression and multi-category classification prediction of the lesion area; the detection loss function introduces the prototype learning mechanism and is defined as follows: in, Represents the positive and negative sample prototype libraries, f i is the feature to be classified, and Δ is the discrimination boundary threshold, which is used to improve the robustness of classification discrimination. (5) Differentiable non-maximum suppression mechanism dNMS is used to remove duplicate candidate frames. The calculation process is as follows: Among them, s i is the confidence of the i-th candidate box, IoU(b i ,b j ) represents the degree of overlap with other candidate boxes, and k is an adjustment parameter; overlapping but credible prediction areas are retained through soft suppression. (6) Combined with the MC Dropout method, the detection model is forward sampled multiple times in the inference stage, the confidence distribution of each candidate area is statistically analyzed, and an uncertainty heat map is generated to mark the confidence credibility of the model in the prediction area.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that A computer program is stored, and when the program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images