Polyp segmentation method and system based on dual boundary guided attention exploration

Abstract

The present invention discloses a polyp segmentation method and system based on dual boundary-guided attention exploration, which relates to the technical field of polyp segmentation, including: a PVTv2 backbone extraction network, which is used to extract backbone features to form a feature map; a multi-scale context aggregation enhancement module MCA, which is used to fuse the context feature information of two adjacent layers of the feature map, and then use dilated convolution with different dilation rates to obtain broad features of the fused feature information, and aggregate multiple broad features to obtain a preliminary predicted segmentation result; a bottom-level detail enhancement module LDE, which is used to suppress background interference and capture detail information according to the preliminary predicted segmentation result; a dual boundary-guided attention exploration module DBE, which is used to obtain the real polyp boundary by adopting an established strategy; the present invention uses a coarse-to-fine strategy to achieve layer-by-layer approximation to the real polyp boundary.

Description

Technical Field

[0001] The present invention relates to the technical field of polyp segmentation, and more particularly to a polyp segmentation method and system based on dual boundary guided attention exploration. Background Art

[0002] Colorectal cancer is a common malignant tumor of the digestive tract that poses a serious threat to human health, with its incidence ranking third among all cancers. Polyps, one of the most important precursors to colorectal cancer, can easily transform into malignant tumors if left untreated. Colonoscopy is an effective method for detecting colonic lesions, providing physicians with precise location information for timely removal before they become cancerous. However, colorectal endoscopic examinations sometimes miss polyps. Therefore, automatic and accurate segmentation of polyps in colonoscopic images is crucial for the clinical prevention of colorectal cancer.

[0003] Traditional polyp segmentation methods usually rely on manually extracted features to identify polyps, such as texture analysis, color distribution, geometric features, and intensity distribution. Although traditional methods have made considerable progress, the polyp segmentation accuracy of these methods based on manual features is still low, the generalization ability is poor, and they cannot meet the requirements of clinical practice. In recent years, with the continuous development of deep learning, polyp segmentation methods based on deep learning have been shown to be superior to traditional methods of manually extracted features. With the introduction of U-shaped networks, this encoder-decoder network has become the mainstream network architecture in medical image segmentation. Recently, attention mechanisms have been increasingly applied to medical segmentation, especially polyp segmentation, and are often used to enhance the fuzzy boundaries of polyps, or to extract global and local features. PraNet aggregates high-level feature information through a parallel decoder to predict rough areas, and establishes the dependency relationship between polyp boundaries and internal structures through a reverse attention mechanism.

[0004] Although the segmentation accuracy and generalization capabilities of these deep learning-based methods have been greatly improved compared to traditional methods, these methods still have shortcomings when facing the challenges of polyp segmentation. The first is the blurred boundaries. When the colonoscope moves in the intestine, it will cause motion blur and reflection problems, which makes the boundaries of the polyp image blurred and increases the difficulty of polyp segmentation. The second is the problem of multi-scale adaptability of polyps. Due to the variable size and shape of polyp tissue, the current polyp segmentation methods still have certain limitations in their ability to extract multi-scale features. Finally, there is the problem of the resemblance between polyps and surrounding normal tissues. Polyps have low contrast with the background, and their texture and color are highly similar to those of surrounding tissues, making them difficult to identify.

[0005] Therefore, how to overcome the problems of blurred polyp boundaries, multi-scale adaptability of polyps, and similarity between polyps and surrounding normal tissues in the existing technology is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] In view of this, the present invention provides a polyp segmentation method and system based on dual boundary guided attention exploration to overcome the above-mentioned defects.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A polyp segmentation system based on dual boundary-guided attention exploration, including: PVTv2 backbone extraction network, multi-scale context aggregation enhancement module MCA, bottom-level detail enhancement module LDE and dual boundary-guided attention exploration module DBE;

[0009] The PVTv2 backbone extraction network is used to extract backbone features to form a feature map;

[0010] The multi-scale context aggregation enhancement module MCA is used to fuse the context feature information of two adjacent layers of the feature map, and then use dilated convolution with different dilation rates to obtain broad features on the fused feature information, and aggregate multiple broad features to obtain a preliminary predicted segmentation result;

[0011] The bottom layer detail enhancement module LDE is used to suppress background interference and capture detail information according to the preliminary predicted segmentation results;

[0012] The dual boundary guided attention exploration module DBE is used to obtain the true polyp boundary by adopting a predetermined strategy.

[0013] Optionally, the features extracted by the PVTv2 backbone extraction network include a first multi-scale pyramid feature X1, a second multi-scale pyramid feature X2, a third multi-scale pyramid feature X3 and a fourth multi-scale pyramid feature X4, wherein the first multi-scale pyramid feature X1 is a detail feature; the second multi-scale pyramid feature X2, the third multi-scale pyramid feature X3 and the fourth multi-scale pyramid feature X4 are semantic features.

[0014] Optionally, the number of dilated convolutions in the multi-scale context aggregation enhancement module MCA is four.

[0015] Optionally, the bottom layer detail enhancement module LDE includes a detail fusion submodule and a detail extraction submodule;

[0016] The detail extraction submodule is used to capture the detail information of polyps from different dimensions by connecting channel attention and spatial attention in series;

[0017] The detail fusion submodule is used to reduce the interference of background information in the underlying features based on the preliminary predicted segmentation results, and then fuse the detail information with the high-level features by downsampling.

[0018] Optionally, the dual boundary guided attention exploration module DBE includes a ternary mask submodule and a boundary mask submodule;

[0019] The ternary mask submodule is used to divide the feature map into regions and assign weights to the divided regions to obtain a weighted feature map;

[0020] The boundary mask submodule is used to convert the weight feature map into a binary mask, and then generate the final boundary mask map through the binary mask.

[0021] Optionally, the established strategy in the dual boundary guided attention exploration module DBE is a layer-by-layer approximation strategy from coarse to fine.

[0022] Optionally, the weighted feature map T output by the ternary mask submodule i The expression is:

[0023]

[0024] in, Represents element-wise multiplication, f i Represents 3×3 convolution, ternary mask D t Described as:

[0025]

[0026] Among them, α l and α h Represents the two thresholds for dividing the ternary mask, and i represents the i-th pixel in the feature map.

[0027] Optionally, the boundary mask map D output by the boundary mask submodule i The expression is:

[0028]

[0029] Among them, the boundary mask D m Described as:

[0030] D m =Dilate(D s )-Erode(D s );

[0031] Among them, Dilate and Erode represent morphological dilation and erosion operations respectively.

[0032] Optionally, the outputs of the two submodules and the preliminary prediction map of each dual boundary guided attention exploration module DBE are optimized using a deep supervision method, and the governor loss function is defined as follows:

[0033] L=L main+L aux ;

[0034] Among them, L main and L aux Represent the main loss and auxiliary loss respectively, the main loss L main and auxiliary loss L aux They are described as:

[0035]

[0036]

[0037] Among them, L wbce and L wiou Represent weighted binary cross entropy loss and weighted intersection loss respectively, the main loss L main Is to calculate the feature map D i The loss between the true value map G and the auxiliary loss L aux Is to calculate the feature map T i The loss between the ground-truth graph G.

[0038] A polyp segmentation method based on dual boundary guided attention exploration, the specific steps are as follows:

[0039] Extract the main features of the original image to form a feature map;

[0040] The context feature information of two adjacent layers of the feature map is fused, and then the fused feature information is subjected to dilated convolution with different dilation rates to obtain broad features. Multiple broad features are aggregated to obtain the preliminary predicted segmentation result.

[0041] Suppress background interference and capture detail information based on the preliminary predicted segmentation results to obtain a detail feature map;

[0042] The true polyp boundary is obtained based on the preliminary predicted segmentation results and the detailed feature map by adopting the established strategy.

[0043] From the above technical solutions, it can be seen that the present invention discloses a polyp segmentation method and system based on dual boundary guided attention exploration, which has the following advantages over the existing technology:

[0044] 1. The present invention uses the PVTv2 backbone extraction network to extract more powerful backbone features from colonoscopy images, showing stronger global information extraction capabilities and better input interference robustness;

[0045] 2. The present invention aggregates and enhances the features of each stage through the multi-scale context aggregation enhancement module MCA, obtains the most extensive features from different receiving areas, and adapts to the multi-scale changes of polyps, thereby obtaining richer local and global features;

[0046] 3. The present invention uses the underlying detail enhancement module LDE to extract more underlying detail information to promote the performance of the overall model, while suppressing background interference and achieving more refined polyp segmentation results;

[0047] 4. The present invention uses a coarse-to-fine strategy to output features through a dual boundary-guided attention exploration module DBE to achieve layer-by-layer approximation to the real polyp boundary. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0049] Figure 1 This is a network structure diagram of the polyp segmentation system based on dual boundary guided attention exploration in the present invention;

[0050] Figure 2 This is a structural diagram of the multi-scale context aggregation enhancement module in the present invention;

[0051] Figure 3 This is a structural diagram of the bottom-level detail enhancement module in the present invention;

[0052] Figure 4 This is a structural diagram of the dual-boundary guided attention exploration module in the present invention;

[0053] Figure 5 This is a result diagram of qualitative comparison between the present invention and other typical model methods;

[0054] Figure 6 A visualization of the output characteristics of the dual boundary-guided attention exploration module at different stages of the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] The embodiment of the present invention discloses a polyp segmentation system based on dual boundary guided attention exploration, such as Figure 1As shown in the figure, it includes the PVTv2 backbone extraction network, the multi-scale context aggregation enhancement module MCA, the bottom-level detail enhancement module LDE and the dual boundary-guided attention exploration module DBE. The PVTv2 backbone extraction network is used to extract more powerful backbone features from colonoscopy images and provide more foreground information for the subsequent decoding stage; the multi-scale context aggregation enhancement module MCA is used to solve the problem of adaptability of multi-scale features of polyps. Through multiple dilated convolutions with different dilation rates, a wider range of features are obtained from different receptive fields; the bottom-level detail enhancement module LDE is used to solve the similarity problem between polyps and surrounding tissues. By extracting more bottom-level detail information to promote the improvement of the overall model performance, a more refined polyp segmentation result is obtained; the dual boundary-guided attention exploration module DBE is used to solve the boundary ambiguity problem. A coarse-to-fine strategy is adopted to achieve layer-by-layer approximation to the real polyp boundary.

[0057] In this embodiment, the PVTv2 backbone extraction network of the present invention outputs four multi-scale pyramid features X1, X2, X3 and X4. Compared with the traditional CNN method that pays more attention to local information, PVTv2 exhibits stronger global information extraction capabilities and better input interference robustness. The output feature X1 contains rich detail information, such as texture, boundary and color, while X2, X3 and X4 have rich semantic information.

[0058] In this embodiment, if Figure 2 As shown in the figure, the multi-scale context aggregation enhancement module MCA can help obtain more detailed local and global feature information, and supplement the position and spatial information through the interaction of adjacent high- and low-level feature information. Specifically, a 1×1 convolution is used on two feature maps from different layers to reduce the feature channels to 32 to reduce computing resources. Then, the lower-layer features are downsampled and connected with the higher-layer features and sent to four groups of dilated convolution branches with different dilation rates. The dilation rates of different branches are set to (1, 2, 4, 8) to obtain more context information. Each branch contains three groups of convolution blocks. Each group of convolution blocks decomposes the standard convolution into a 3×1 convolution, followed by a 1×3 convolution. Then, through this factorized convolution, 33% of parameters can be saved with the same number of filters. Finally, the features of the four branches are connected together as the output of the MCA module.

[0059] In this embodiment, if Figure 3As shown in the figure, the bottom-level detail enhancement module LDE is used to extract richer detail information. The appearance of polyps is usually very similar to the surrounding normal tissues. However, the low-level feature map X1 usually contains more detail information such as texture, boundary and color. Therefore, the bottom-level detail enhancement module is used to extract richer detail information. The bottom-level detail enhancement module LDE consists of two parts. The first part is a detail fusion submodule, which reduces the interference of background information in the bottom-level features through the auxiliary role of the preliminary prediction result map, and then fuses more detail information with high-level features through downsampling. The second part is a detail extraction submodule, which uses channel attention and spatial attention in series to capture the texture, boundary and color detail information of the polyps from different dimensions.

[0060] In this embodiment, if Figure 1 As shown in the figure, the dual boundary-guided attention exploration module DBE adopts a coarse-to-fine layer-by-layer approximation strategy. According to the prediction results of the previous scale and the detail enhancement features, more real polyp areas are obtained from the coarse uncertain areas through the ternary mask submodule, and then the boundary mask submodule is used to explore finer polyp boundaries from the detailed boundary areas. There are four dual boundary-guided attention exploration modules constructed in a cascade manner, which explore the uncertain areas and boundary areas layer by layer to achieve layer-by-layer approximation of the real polyp boundaries.

[0061] In this embodiment, if Figure 4 As shown in the figure, the dual boundary guided attention exploration module DBE consists of two parts. The first part is a ternary mask submodule, which divides the feature map into three areas, namely foreground, background and uncertain area, and assigns different weight values ​​to different areas. In order to emphasize the uncertain area, the pixels in the uncertain area are set to the highest weight 1, in order to balance the high response area, the pixels in the foreground area are set to 0, and in order to suppress the interference of the background area, the pixels in the background area are set to the lowest weight -1; the second part is a boundary mask submodule, which first converts the feature map output by the ternary mask submodule of the first part into a binary mask, and then performs morphological dilation and erosion operations on the binary mask through morphological operations, and performs element-wise subtraction operations on the mask after dilation and erosion operations to generate the final boundary mask map.

[0062] exist Figure 4 The upper-middle ternary mask submodule upsamples the prediction result of the previous scale to obtain D u , then transformed into a ternary mask D t The output feature L corresponding to the underlying detail enhancement module LDE i The result of element-wise multiplication is then convolved with D u By element-wise addition, we can obtain the output feature map T of the ternary mask submodule.i , the specific description is as follows:

[0063]

[0064] in, Represents element-wise multiplication, f i Represents 3×3 convolution, ternary mask D t Described as:

[0065]

[0066] Among them, α l and α h Represents the two thresholds for dividing the ternary mask, and i represents the i-th pixel in the feature map.

[0067] The lower boundary mask submodule is derived from the output feature map T of the previous ternary mask submodule. i Extract boundary information and output feature map T i Converted into a binary mask D with a threshold of 0.5 s , then D s Perform morphological operations to generate boundary mask D m , boundary mask D m Compared with the input feature map L of the previous ternary mask submodule i The result of element-wise multiplication is then convolved with T i By element-wise addition, we can obtain the output feature map D of the boundary mask submodule. i , the specific description is as follows:

[0068]

[0069] Among them, the boundary mask D m Described as:

[0070] D m =Dilate(D s )-Erode(D s );

[0071] Among them, Dilate and Erode represent morphological dilation and erosion operations respectively.

[0072] In this embodiment, the loss function of the present invention is designed as a combination of the weighted IoU loss function and the weighted binary cross entropy BCE loss function. The outputs of the two submodules and the preliminary prediction map of each dual boundary guided attention exploration module adopt the deep supervision method as the optimization target. The governor loss function is defined as follows:

[0073] L=L main +L aux ;

[0074] Among them, L main and L aux Represent the main loss and auxiliary loss respectively, the main loss L main and auxiliary loss L aux They are described as:

[0075]

[0076]

[0077] Among them, the main loss L main Is to calculate the feature map D i The loss between the true value map G and the auxiliary loss L aux Is to calculate the feature map T i The loss between the ground-truth graph G.

[0078] In this embodiment, the present invention also includes using the average dice similarity coefficient mDice, the average intersection over union ratio mIoU and the mean absolute error MAE to quantitatively evaluate the segmentation performance of polyps. The specific expressions of each index are as follows:

[0079] The formula for the average dice similarity coefficient mDice is as follows:

[0080] The formula for average intersection over union (mIoU) is as follows:

[0081] The mean absolute error (MAE) formula is as follows:

[0082] D m =Dilate(D s )-Erode(D s );

[0083] Among them, TP represents true positive, FP represents false positive, FN represents false negative, and n represents the number of test images. and p i Represents the prediction and corresponding true value of the i-th pixel among the total n pixels.

[0084] Specifically, to evaluate the segmentation performance of the proposed method, experiments were conducted on five polyp segmentation datasets, including Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, ETIS, and CVC-T. The description of the datasets is as follows:

[0085] Kvasir-SEG: This dataset consists of 1000 polyp images and their corresponding ground truth polyp masks, annotated by expert endoscopists. The resolution of the images included in Kvasir-SEG ranges from 332×487 to 1920×1072 pixels.

[0086] CVC-ClinicDB: This dataset, also known as CVC-612, contains 612 open-access images from 25 colonoscopy videos with a resolution of 384 × 288. Each image has its associated manually annotated ground truth covering polyps.

[0087] CVC-ColonDB: This dataset consists of 380 images from 15 colonoscopy videos. The image resolution is 574×500.

[0088] ETIS: This dataset contains 196 images captured from 34 colonoscopy videos with an image resolution of 1225×996.

[0089] CVC-T: This dataset is a subset of EndoScene and contains 60 images from 44 colonoscopic sequences of 36 patients with an image resolution of 574×500.

[0090] To ensure a fair comparison, our experiments followed the same principles as other methods, using a total of 1,450 training images in the training set, 900 of which were from Kvasir-SEG and 550 from CVC-ClinicDB. A total of 798 test images were used in the test set, 100 of which were from Kvasir-SEG and 62 from CVC-ClinicDB. Three other datasets were also used for testing: 380, 196, and 62 images from CVC-ColonBD, ETIS, and CVC-T, respectively. Detailed information about these five polyp datasets is shown in Table 1.

[0091] Table 1 Polyp dataset information

[0092] Dataset Imagesize Imagenumber Number of train samples Number of test samples Kvasir-SEG Variable 1000 900 100 CVC-ClinicDB 384×288 612 550 62 CVC-ColonDB 574×500 380 0 380 ETIS 1225×966 196 0 196 CVC-T 574×500 60 0 60

[0093] This paper uses three experiments to validate the performance of the polyp segmentation system based on dual-boundary guided attention exploration, including qualitative and quantitative experiments and ablation studies. In the experiments, the method of this embodiment is compared with six different methods, including state-of-the-art methods: UNet, UNet++, SFA, PraNet, SANet, and CaraNet. Three evaluation metrics are also analyzed and compared on five different polyp segmentation datasets. For fair comparison, the segmentation maps of these methods are directly generated from the original code provided.

[0094] from Figure 5 It can be seen from the visual comparison between the method of this embodiment and different competing methods that it can be qualitatively seen that, compared with other competing methods, the method of this embodiment can more accurately segment the polyp area and perform better in many challenging aspects.

[0095] Tables 2 and 3 list the quantitative results of statistical comparisons of the method of this embodiment with six different competing methods on five different data sets.

[0096] Table 2 Comparison of experimental results of different methods on CVC-ClinicDB and Kvasir-SEG polyp datasets

[0097]

[0098]

[0099] Considering that the training sets were selected from CVC-ClinicDB and Kvasir-SEG, we first quantitatively analyze the learning ability of this method. As shown in Table 2, the method of this embodiment is the best in various indicators on both datasets. The mDice of the method of this embodiment is 93.9% and 92.2% in CVC-ClinicDB and Kvasir-SEG, respectively.

[0100] In order to verify the generalization performance of the method proposed in this embodiment, tests were performed on three unseen datasets (CVC-ColonDB, ETIS, and CVC-T). As shown in Table 3, the model of this embodiment achieves good generalization performance compared with the six competing methods. In particular, on the challenging datasets CVC-ColonDB and ETIS, the generalization ability of the method of this embodiment is significantly improved. On CVC-ColonDB, mDice leads CaraNet and SANet by 5.1% and 7.1%, respectively. On ETIS, it exceeds CaraNet and SANet by 5.9% and 5.6%, respectively. On CVC-T, it exceeds SANet by 1.5%.

[0101] To verify the effectiveness of each component of the model, ablation experiments were conducted on the dual boundary-guided attention exploration module (DBE), the multi-scale context aggregation enhancement module (MCA), and the bottom-level detail enhancement module (LDE). The backbone network baseline consists of PVTv2 and PD, and the standard model consists of "Baseline+MCA+LDE+DBE." The effectiveness of different modules was evaluated by removing or modifying them from the standard model. "with / without MCA," "with / without LDE," and "with / without DBE" represent the removal of MCA, LDE, or DBE, respectively, from the standard model. The experimental results are shown in Table 3 below.

[0102] Table 3 Ablation experiment results on different polyp datasets

[0103]

[0104]

[0105] Ablation study of DBE: In order to study the effectiveness of the DBE module, the output characteristics of the DBE module at different stages are visualized. Figure 6 As shown, the redder the area, the more attention the network needs to pay. It can be observed that as the stages progress, it gradually refines the polyp boundaries and explores more polyp information. This demonstrates that the DBE module's coarse-to-fine strategy effectively enhances fuzzy edge information. To quantitatively analyze the effectiveness of the DBE module, a version lacking the DBE module, "without DBE," was trained. The DBE module was completely removed and replaced with element-wise addition. The experimental results are shown in Table 3. Compared to the standard model, the model without DBE exhibits a sharp decline in performance on the five datasets. On CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB, ETIS, and CVC-T, the lack of DBE reduces mDice by 0.8%, 0.6%, 1.7%, 1.6%, and 2.1%, respectively, compared to the performance of the standard full model, DBENet.

[0106] Ablation study of MCA: Similarly, to quantitatively analyze the effectiveness of the MCA module, a version without the MCA module, "w / o MCA", was trained. The MCA module was completely removed and the contextual information of the first three high-level layers (i.e., X2, X3, and X4) was directly fused. The results in Table 3 show that compared with the standard model, the model without MCA has a 1.0% and 0.5% decrease in mDice on the CVC-ColonDB and ETIS datasets, respectively. In particular, the CVC-ColonDB and ETIS datasets are challenging unseen datasets, which indicates that MCA has enhanced the generalization ability of multi-scale.

[0107] Ablation Study of LDE: To demonstrate the capabilities of LDE, a version lacking the LDE module was trained. As shown in Table 3, removing LDE resulted in a slight decrease in performance on all five datasets compared to the standard model. On the CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB, ETIS, and CVC-T datasets, removing LDE reduced mDice by 0.5%, 0.4%, 0.6%, 0.1%, and 1.0%, respectively. This demonstrates that the LDE module significantly improves the model's segmentation capabilities.

[0108] A polyp segmentation method based on dual boundary guided attention exploration, the specific steps are as follows:

[0109] Extract the main features of the original image to form a feature map;

[0110] The context feature information of two adjacent layers of the feature map is fused, and then the fused feature information is subjected to dilated convolution with different dilation rates to obtain broad features. Multiple broad features are aggregated to obtain the preliminary predicted segmentation result.

[0111] Suppress background interference and capture detail information based on the preliminary predicted segmentation results to obtain a detail feature map;

[0112] The true polyp boundary is obtained based on the preliminary predicted segmentation results and the detailed feature map by adopting the established strategy.

[0113] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0114] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A polyp segmentation system based on dual boundary guided attention exploration, characterized by: include: PVTv2 backbone extraction network, multi-scale context aggregation enhancement module MCA, bottom-level detail enhancement module LDE and dual boundary guided attention exploration module DBE; The PVTv2 backbone extraction network is used to extract backbone features to form a feature map; The multi-scale context aggregation enhancement module MCA is used to fuse the context feature information of two adjacent layers of the feature map, and then use dilated convolution with different dilation rates to obtain broad features on the fused feature information, and aggregate multiple broad features to obtain a preliminary predicted segmentation result; The bottom-level detail enhancement module LDE is used to suppress background interference and capture detail information based on the preliminary predicted segmentation results; the bottom-level detail enhancement module LDE includes a detail fusion submodule and a detail extraction submodule; the detail extraction submodule is used to capture the detail information of the polyp from different dimensions by using the channel attention and spatial attention in series; the detail fusion submodule is used to reduce the interference of background information in the bottom-level features based on the preliminary predicted segmentation results, and then fuse the detail information with the high-level features by downsampling; The dual boundary guided attention exploration module DBE is used to obtain the true polyp boundary by adopting a predetermined strategy; The dual boundary-guided attention exploration module DBE includes a ternary mask submodule and a boundary mask submodule; the ternary mask submodule is used to divide the feature map into regions and assign weights to the divided regions to obtain a weighted feature map; the boundary mask submodule is used to convert the weighted feature map into a binary mask, and then generate the final boundary mask map through the binary mask.

2. The polyp segmentation system based on dual boundary guided attention exploration according to claim 1, characterized in that: The features extracted by the PVTv2 backbone extraction network include a first multi-scale pyramid feature X1, a second multi-scale pyramid feature X2, a third multi-scale pyramid feature X3 and a fourth multi-scale pyramid feature X4, wherein the first multi-scale pyramid feature X1 is a detail feature; the second multi-scale pyramid feature X2, the third multi-scale pyramid feature X3 and the fourth multi-scale pyramid feature X4 are semantic features.

3. The polyp segmentation system based on dual boundary guided attention exploration according to claim 1, characterized in that: The number of dilated convolutions in the multi-scale context aggregation enhancement module MCA is four.

4. The polyp segmentation system based on dual boundary guided attention exploration according to claim 1, characterized in that: The established strategy in the dual boundary guided attention exploration module DBE is a layer-by-layer approximation strategy from coarse to fine.

5. The polyp segmentation system based on dual boundary guided attention exploration according to claim 1, characterized in that: The weighted feature map T output by the ternary mask submodule i The expression is: in, Represents element-wise multiplication, f i Represents 3×3 convolution, ternary mask D t Described as: Among them, a l and a h Represents the two thresholds for dividing the ternary mask, i represents the i-th pixel in the feature map; L i Represents the output features corresponding to the underlying detail enhancement module LDE.

6. The polyp segmentation system based on dual boundary guided attention exploration according to claim 1, characterized in that: Boundary mask map D output by the boundary mask submodule i The expression is: Among them, the boundary mask D m Described as: D m =Dilate(D s )-Erode(D s ); Among them, Dilate and Erode represent morphological dilation and erosion operations respectively; f i represents 3×3 convolution, L i Represents the output features corresponding to the underlying detail enhancement module LDE, T i represents the output feature map of the ternary mask submodule, D s Represents a binary mask with a threshold of 0.

5.

7. The polyp segmentation system based on dual boundary guided attention exploration according to claim 1, characterized in that: The outputs of the two submodules and the preliminary prediction maps of each dual boundary guided attention exploration module DBE are optimized using a deep supervision method. The governor loss function is defined as follows: L=L main +L aux ; Among them, L main and L aux Represent the main loss and auxiliary loss respectively, the main loss L main and auxiliary loss L aux They are described as: Among them, L wbce and L wiou Represent weighted binary cross entropy loss and weighted intersection loss respectively, the main loss L main Is to calculate the feature map D i The loss between the true value map G and the auxiliary loss L aux Is to calculate the feature map T i The loss between the ground-truth graph G.

8. A method for applying the polyp segmentation system based on dual boundary guided attention exploration as described in any one of claims 1 to 7, characterized in that: The specific steps are: Extract the main features of the original image to form a feature map; The context feature information of two adjacent layers of the feature map is fused, and then the fused feature information is subjected to dilated convolution with different dilation rates to obtain broad features. Multiple broad features are aggregated to obtain the preliminary predicted segmentation result. Suppress background interference and capture detail information based on the preliminary predicted segmentation results to obtain a detail feature map; The true polyp boundary is obtained based on the preliminary predicted segmentation results and the detailed feature map by adopting the established strategy.

Images