Intestinal polyp boundary segmentation method based on sparse connectivity and global feature enhancement

Abstract

The present invention relates to an intestinal polyp boundary segmentation method based on sparse connection and global feature enhancement. The method uses a sparse harmonic connection network as the encoder backbone to extract features from an input intestinal polyp image. A dilated spatial convolutional pyramid pooling module is used to parallel sample the deepest features at different sampling rates. The results of the parallel sampling are input into a partial feature decoder together with the features of the last two layers of the encoder for decoding to generate a rough initial prediction map. A shallow attention module is used to extract shallow boundary information from the rough initial prediction map and shallow features for preliminary refinement of the boundaries of the initial prediction map. A global feature enhancement module is used to process deep features for guiding subsequent boundary segmentation of the initial prediction map. A reverse attention module is used to obtain complete boundaries from multi-scale features to complete intestinal polyp boundary segmentation. Based on the traditional encoding and decoding structure, the present invention greatly enhances the boundary modeling capability and the regional hole filling capability.

Description

Technical Field

[0001] The present invention relates to the field of image analysis, such as the technical field of transforming bitmap images into non-bitmap images, and in particular to a method for segmenting intestinal polyp boundaries based on sparse connections and global feature enhancement in the field of deep learning image segmentation. Background Art

[0002] Gastrointestinal diseases are among the most common human illnesses, with colorectal cancer (CRC) being the third most common malignant tumor and the second most lethal cancer worldwide. Early symptoms of CRC are subtle, but as the tumor grows, symptoms such as changes in bowel habits, blood in the stool, diarrhea, and localized abdominal pain develop. However, these clinical symptoms often appear in the late stages and are difficult to distinguish from benign conditions such as inflammatory bowel disease and hemorrhoids. Therefore, early screening during the asymptomatic phase is a fundamental measure to reduce cancer incidence.

[0003] Intestinal polyps, a high-risk precancerous lesion, are the precursor to colorectal cancer. Colonoscopy is the gold standard for diagnosing colon diseases. By inserting a colonoscope directly into the intestine to examine the intestinal wall, endoscopists can locate polyp lesions and conduct accurate assessments. Early endoscopic resection of high-risk polyps can effectively reduce the incidence and mortality of CRC.

[0004] However, intestinal polyps vary in size, shape, color, and texture. Their structures range from small, translucent, and difficult-to-identify protrusions to masses 10-20 mm in diameter, with high inter-class similarity and intra-class variation. Furthermore, intestinal polyps have low contrast with the surrounding mucosa, and the presence of interfering information such as air bubbles and food debris in the intestines often results in unclear polyp boundaries. These factors can lead to misdiagnosis or missed diagnoses during colonoscopy. Therefore, computer-aided diagnosis (CAD) algorithms are urgently needed to assist doctors in making more accurate diagnoses and treatments.

[0005] Early polyp detection and segmentation algorithms often relied on traditional handcrafted features such as color and texture to train classifiers to distinguish polyps from background. While these methods are effective in specific scenarios and can achieve good results, these features rely on domain knowledge and have limited representation capabilities. In recent years, clinical applications have also placed higher demands on pixel-level segmentation tasks. High-precision polyp segmentation algorithms can effectively reduce misdiagnosis rates and are therefore of great significance. Summary of the Invention

[0006] The present invention solves the problems existing in the prior art and provides an intestinal polyp boundary segmentation method based on sparse connection and global feature enhancement.

[0007] The technical solution adopted by the present invention is a method for intestinal polyp boundary segmentation. The method uses a sparse harmonic connection network as the encoder backbone to extract features from the input intestinal polyp image. The deepest layer features are sampled in parallel at different sampling rates using a dilated spatial convolutional pyramid pooling module. The results of the parallel sampling are input into a partial decoder together with the features of the last two layers of the encoder for decoding to generate a rough initial prediction map.

[0008] The shallow attention module is used to extract the boundary information contained in the shallow features to preliminarily refine the boundaries of the initial prediction image;

[0009] The global feature enhancement module is used to process deep features to guide the subsequent boundary segmentation of the initial prediction map;

[0010] The reverse attention module is used to obtain complete boundaries from multi-scale features and complete intestinal polyp boundary segmentation.

[0011] Preferably, the sparse harmonic connection network includes several HarD Blocks, and the number of channels in each layer of any HarD Block is k×m n , where k is the initial growth rate, m is the dimension compression factor, for the Lth layer, if L can be 2 N If n is divisible by an integer, then n takes the maximum value of N; the output of each HarD Block is the connection between the last layer and all odd layers, and the depth of HarD Block is a power of 2.

[0012] Preferably, the features extracted from the sparse harmonic connection network are {f i ,i=1,2,3,4,5}, where the shallow features are {f i ,i=1,2}, the deep features are {f i ,i=3,4,5}.

[0013] Preferably, the rough initial prediction map S c =PFD(f3,f4,f5'), where f5'=ASPP(f5).

[0014] Preferably, the shallow attention module constructs a Gaussian kernel k, uses shallow features f2 to extract boundary detail information, and applies the maximum function to increase S c The weight coefficient of the shallow attention module and the rough initial prediction map are added to obtain the initial prediction map of the preliminary refined boundary.

[0015] Preferably, the global feature enhancement module includes a global average pooling branch, a 3×3 adaptive pooling branch, a 5×5 adaptive pooling branch and a feature map Non-local branch. The deep features are input to each branch, and the obtained results are connected and output to obtain g i=GFE(f5),i=3,4,5.

[0016] Preferably, the enhanced global feature g i The local features f3, f4, and f5 extracted from the last three layers of the encoder are added pixel by pixel as the side input h for the subsequent boundary extraction part. i ,

[0017] Preferably, the reverse attention weight of the inverted attention module σ(·) is the sigmoid function, Indicates S i+1 Upsampled output; output feature R i By deep features and reverse attention weight A i Multiply them to get R i =f i ·A i ;

[0018] The inverted attention module deletes the currently predicted area from each deep feature, guiding the entire network to discover complementary target areas and details in turn. The three inverted attention modules obtain a high-precision prediction map through three inversion operations, and finally complete the intestinal polyp boundary segmentation through the sigmoid operation.

[0019] Preferably, the loss function of the network in the training of the method combines the weighted intersection-over-union loss and the binary cross entropy loss, and is in, represents the weighted IoU loss, represents the BCE loss of global constraints and local constraints.

[0020] Preferably, the method uses four up-sampled network outputs of the same size as the annotation image GT Calculate the total loss as in, S g 、S i The upsampling result is obtained and the network parameters are updated through back propagation.

[0021] The present invention relates to a method for intestinal polyp boundary segmentation based on sparse connection and global feature enhancement. The method uses a sparse harmonic connection network as the encoder backbone to extract features from an input intestinal polyp image. A dilated spatial convolutional pyramid pooling module is used to parallel sample the deepest features at different sampling rates. The results of the parallel sampling are input into a partial feature decoder together with the features of the last two layers of the encoder for decoding to generate a rough initial prediction map. A shallow attention module is used to extract shallow boundary information from the rough initial prediction map and shallow features to preliminarily refine the boundaries of the initial prediction map. A global feature enhancement module is used to process deep features to guide subsequent boundary segmentation of the initial prediction map. A reverse attention module is used to obtain complete boundaries from multi-scale features to complete intestinal polyp boundary segmentation.

[0022] The beneficial effects of the present invention are:

[0023] (1) A method for intestinal polyp boundary segmentation based on sparse connectivity and global feature enhancement is proposed. Based on the traditional encoding and decoding structure, the ability to model the boundary and fill regional holes is greatly enhanced;

[0024] (2) The sparse harmonic connection network HarDNet is proposed as the encoder backbone, which enables the segmentation network to achieve high segmentation accuracy while also improving the inference speed to a certain extent;

[0025] (3) A global feature enhancement module (GFE) is proposed to reduce the influence of background noise by strengthening the guiding ability of global features, so that the overall network can better cope with the influence of interference factors such as foreign body coverage, bubbles, and light spots in the polyp segmentation task;

[0026] (4) A shallow attention module SA is proposed to extract boundary information from shallow features to preliminarily refine the boundaries of the initial prediction image, and then a reverse attention module RA is used to approximate a more accurate polyp area and boundary by gradually erasing the foreground. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall network flow diagram of the present invention;

[0028] Figure 2 This is the internal structure diagram of the encoder sparse HarDNet component block HarD Block;

[0029] Figure 3 It is the decoding flow chart of decoder PD;

[0030] Figure 4 This is the GFE module structure diagram;

[0031] Figure 5 This is the internal structure diagram of the RA module;

[0032] Figure 6 This is a comparison of the segmentation results of the present invention and different networks. DETAILED DESCRIPTION

[0033] The present invention is further described in detail below with reference to the embodiments, but the protection scope of the present invention is not limited thereto.

[0034] The present invention relates to a method for intestinal polyp boundary segmentation based on sparse connections and global feature enhancement. The method uses a sparse harmonic connection network as the encoder backbone to extract features from an input intestinal polyp image. A dilated spatial convolutional pyramid pooling module is used to perform parallel sampling of the deepest layer features at different sampling rates. The results of the parallel sampling are input into a partial decoder together with the features of the last two layers of the encoder for decoding to generate a rough initial prediction map.

[0035] The shallow attention module is used to extract shallow boundary information from the rough initial prediction map and shallow features, which is used to preliminarily refine the boundary of the initial prediction map;

[0036] The global feature enhancement module is used to process deep features to guide the subsequent boundary segmentation of the initial prediction map;

[0037] The reverse attention module is used to obtain complete boundaries from multi-scale features and complete intestinal polyp boundary segmentation.

[0038] In this paper, a lightweight Harmonic Densely Connected Networks (HarDNet) is introduced as the encoder backbone;

[0039] The Atrous Spatial Pyramid Pooling (ASPP) module is applied between the encoder and decoder to sample the deepest features in parallel at different sampling rates.

[0040] In the decoder design, a partial decoder (PD) is used to generate a rough initial prediction map;

[0041] The shallow-level attention module (SA) is used to extract shallow boundary information and preliminarily refine the boundary of the initial prediction image;

[0042] A global feature enhancement module (GFE) is proposed to extract global features to guide subsequent boundary segmentation tasks.

[0043] The reverse attention module (RA) is used to gradually mine finer boundary details from multi-scale features.

[0044] In the present invention, the polyp image data obtained are taken from the public polyp datasets CVC-ClinicDB, Kvasir, and ETIS-Larib, which contain polyp images of various types, shapes, and colors.

[0045] First, the polyp image is input into the segmentation network proposed in this invention, and sparse HarDNet is used as the feature encoder to encode it, and five different levels of features {f i ,i=1,2,3,4,5}.

[0046] The sparse harmonic connection network includes several HarD Blocks, and the number of channels in each layer of any HarD Block is k×m n , where k is the initial growth rate, m is the dimension compression factor, for the Lth layer, if L can be 2 N If n is divisible by N, then n takes the maximum value of N; the output of each HarD Block is the connection between the last layer and all odd layers, and the depth of HarD Block is a power of 2.

[0047] In the present invention, memory traffic consumption should be considered when designing a neural network framework for high-resolution image application scenarios. The classic network DenseNet starts from feature reuse, fully utilizes the features of each layer and strengthens feature transfer, achieving better results and fewer parameters. The backbone network HarDNet used here can be regarded as a sparse version of DenseNet. Because its connection shape is similar to harmonics, it is named sparse harmonic connection network. Based on DenseNet, this network reduces the number of connections between layers to reduce memory overhead, thereby improving the inference speed while maintaining high performance.

[0048] In this invention, the sparse HarDNet is composed of multiple HarD Blocks, and the number of channels in each layer of HarD Block is k×m n , where k is the initial growth rate and m is the dimension compression factor. For the Lth layer, if L can be 2 N If n is divisible by 0, then n takes the maximum value of N. Layers where n is not 0 are called key layers. Key layers have more connections and therefore have a greater impact on overall network performance. To ensure the final layer has the maximum number of channels, the depth of the HardD Block is set to a power of 2. This allows all previous even-numbered layers to be released from memory upon reaching the final layer. Increasing the channel width of key layers balances the channel ratio between input and output, thus compensating for accuracy loss.

[0049] The connection scheme of the present invention makes the gradient pass through the key layers more often and skip the odd layers. In order to alleviate the degradation caused by skipping the odd layers, the output of the HarD Block is defined as the connection between the last layer and all odd layers;

[0050] In applications, a small number of 1×1 convolutions are used between HarD blocks to compress parameters. While increasing functional modules to enable the network to achieve high segmentation accuracy, it can also maintain a certain inference speed.

[0051] The features extracted from the sparse harmonic connection network are {f i ,i=1,2,3,4,5}, where the shallow features are {f i ,i=1,2}, the deep features are {f i ,i=3,4,5}.

[0052] Before decoding, the ASPP module uses multi-scale features to capture more accurate context information. The deepest feature f5 is sampled in parallel at different sampling rates to obtain f5', that is, f5' = ASPP(f5);

[0053] Compared with deep features, shallow features often lead to greater computational consumption due to their larger resolution, but have less contribution to performance. Therefore, the present invention adopts a partial feature decoder (PD) in the decoding part, which obtains a rough prediction map S by aggregating only deep features f3, f4 and f5'. c , rough initial prediction map S c =PFD(f3,f4,f5'); After discarding the shallow features, the computational complexity of the network can be effectively reduced, and the training speed can be greatly accelerated.

[0054] The shallow attention module constructs a Gaussian kernel k, uses shallow features f2 to extract boundary details, and applies the maximum function to increase S c The weight coefficient of the shallow attention module and the rough initial prediction map are added to obtain the initial prediction map of the preliminary refined boundary.

[0055] In this invention, the initial prediction graph S is preliminarily refined by the shallow-level attention (SA) module. c However, since the PD decoder discards the high-resolution shallow features, the coarse prediction map S c The resolution is relatively low, which means that the boundary information may be filtered out. In order to improve the effectiveness of the coarse prediction map as the initial guidance area for the subsequent boundary modeling, a shallow attention module SA is proposed to extract boundary information from shallow features containing more boundary detail information to refine S c; Considering the resolution limitation, from the shallow features {f i ,i=1,2}, a relatively low resolution f2 is selected to minimize the impact on the inference speed.

[0056] In this invention, SA(·) is a shallow attention module, which is essentially a Gaussian kernel k with standard deviation σ=32 and kernel size λ=4, including normalization operation. A maximum function is applied in SA(·) to increase the coarse prediction S c The weight coefficient of the overall network frame Figure 1 It can be seen that the output S of the SA module a Compared with the initial prediction graph S c More prominent areas of the border.

[0057] The global feature enhancement module includes a global average pooling branch, a 3×3 adaptive pooling branch, a 5×5 adaptive pooling branch, and a feature map non-local branch. The deep features are input to each branch, and the results are connected and output to obtain g i =GFE(f5),i=3,4,5.

[0058] Enhanced global feature g i The local features f3, f4, and f5 extracted from the last three layers of the encoder are added pixel by pixel as the side input h for the subsequent boundary extraction part. i ,

[0059]

[0060] In the present invention, a global feature enhancement module (GFE) is used to extract global enhanced features to reduce the influence of background noise. In the image segmentation task, global features, i.e., the overall attributes of the image, contain more color, texture and shape features, which are intuitive and simple to represent. However, the feature representation ability will be significantly reduced in the case of image aliasing or occlusion. Moreover, in CNN, convolution operations are good at extracting local features but have difficulty in capturing global representations, which makes it difficult for global features to provide good guidance for subsequent segmentation tasks. Therefore, a global feature enhancement module (GFE) is proposed to extract global features that are gradually lost during the encoder layer refinement process to suppress the influence of background noise.

[0061] In this invention, the non-local operation is to take the weighted sum of the features of all positions to a position as the response value of the position, which can capture the long-distance dependency between non-adjacent pixels of the image while maintaining the original resolution of the image. The calculation is simple and efficient, and it is applied to various computer vision frameworks. The four different branches can extract contextual features of different scales. The extracted features are upsampled and cascaded to obtain the global enhanced feature g.i .

[0062] In the present invention, the global feature g is obtained i The local features f3, f4 and f5 extracted from the last three layers of the sparse HarDNet are added at the pixel level as the side input of the subsequent boundary extraction part. The fused features enable the overall network to better cope with complex situations such as foreign body coverage, bubbles, and light spots in the polyp segmentation task; in order to match the output of different layers of the encoder, the output of GFE g i After different convolution and upsampling operations, the number of channels and the size of the feature map are adjusted to f. i Stay consistent.

[0063] The reverse attention weights of the reverse attention module σ(·) is the sigmoid function, Indicates S i+1 Upsampled output; output feature R i By deep features and reverse attention weight A i Multiply them to get R i =f i ·A i ;

[0064] The inverted attention module deletes the currently predicted area from each deep feature, guiding the entire network to discover complementary target areas and details in sequence. After three inversions, the three inverted attention modules obtain a high-precision prediction map, and complete the intestinal polyp boundary segmentation through sigmoid operation.

[0065] In this invention, the constraint relationship between the region and the boundary is established by the reverse attention module (RA), and the segmentation result with clear and smooth boundaries and complete and dense regions is obtained; since SA preliminarily refines the rough prediction map S c However, shallow features are likely to introduce noise. Secondly, if the complementary relationship between regions and boundaries is not fully considered, it will be difficult to effectively capture image holes and missing boundaries. In order to reduce the impact of noise in pixel-level prediction, the extracted multi-scale features are fully utilized to mine dense and complete regions and construct fine and smooth boundaries. A set of reverse attention modules RA are used to guide the network to approach more accurate polyp regions and boundaries from deep to shallow by gradually erasing the foreground.

[0066] In the present invention, the reverse attention module guides the entire network to sequentially discover complementary target regions and details by deleting the currently predicted region from each deep feature. Finally, a high-precision prediction graph S3 is obtained after three inversions through three RA blocks, and the final prediction result of the invention is obtained after the Sigmoid operation.

[0067] The loss function of the network in the training of the method combines the weighted intersection-over-union loss and the binary cross entropy loss, which is in, represents the weighted IoU loss, represents the BCE loss of global constraints and local constraints.

[0068] The method uses four up-sampled network outputs with the same size as the labeled image GT Calculate the total loss as in, For S g 、S i The upsampling result is obtained and the network parameters are updated through back propagation.

[0069] In the present invention, the training of the network model includes the following steps:

[0070] The S1 training dataset is from CVC-ClinicDB and Kvasir datasets, and the test dataset is from CVC-ClinicDB and ETIS-Larib;

[0071] When inputting the S2 training images into the model, they were uniformly resized to 352 × 352 pixels, and a multi-scale strategy {0.75, 1, 1.25} was used instead of data augmentation;

[0072] The S3 model is trained in an end-to-end manner using the Adam optimization algorithm;

[0073] The loss function used by S4 is the sum of weighted intersection over union (IoU) loss and binary cross entropy (BCE) loss. in represents the weighted IoU loss, The BCE loss represents the global constraint and the local constraint (pixel level); the weighted IoU loss highlights the importance of difficult-to-classify pixels by increasing their weights. Similarly, the weighted BCE loss pays more attention to difficult-to-classify pixels instead of giving equal weight to all pixels.

[0074] During the training process, the experiment selects four network outputs with the same size as the labeled image (Ground Truth, GT) after upsampling. To calculate the total loss, in, For S g 、S i The upsampling result is obtained and the network parameters are updated through back propagation.

[0075] In order to achieve the above content, the present invention also stores the intestinal polyp boundary segmentation program on the medium, adopts the intestinal polyp boundary segmentation method when the device is running, and then inputs the intestinal polyp image to be segmented into the device, and outputs the segmented result through the neural network.

[0076] The effects of the present invention can be further illustrated by experiments.

[0077] 1) Experimental conditions

[0078] The model was implemented using the PyTorch framework. All experiments were conducted on a Microsoft Windows Server 2019 Datacenter system running version 10.0.17763. The CPU was an Intel(R) Xeon(R) Gold 6161 CPU @ 2.20 GHz, with 64 GB of memory. Training was performed using an RTX 3080 GPU. Training images were collected from the CVC-ClinicDB and Kvasir datasets, containing 160 polyp images. Training images were uniformly resized to 352 × 352 pixels before input to the model. The model was trained end-to-end using the Adam optimization algorithm with a learning rate of 1e-4 and a batch size of 4. The HarDNet68 network parameters were used as pre-training parameters. Convergence was achieved after 60 epochs.

[0079] 2) Experimental results

[0080] This paper uses the average Dice coefficient (Dice) and mean Intersection Over Union (mIoU) as the main evaluation indicators:

[0081]

[0082]

[0083] The proposed model is compared with the classic medical image segmentation methods U-Net, U-Net++, and ResUNet++, as well as two newer methods proposed for polyp segmentation, HarD-MSEG and PraNet. Experiments are conducted on two experimental datasets, CVC-ClinicDB and ETIS-Larib. CVC-ClinicDB is the test set corresponding to the training dataset, which is used to verify the learning ability of the model. ETIS-Larib is a dataset that was not seen during training and is used to verify the generalization ability of the model. The experimental results are shown in Table 1.

[0084] Table 1

[0085]

[0086] As shown in Table 1, our model outperforms existing models by a significant margin on both metrics on the CVC-ClinicDB dataset. Dice improves by 2.4% and mIoU by 2.5% over the second-best performing PraNet, demonstrating excellent learning capabilities. The ETIS-Larib dataset is a challenging dataset, where all other models perform poorly. Comparative experiments on the ETIS-Larib dataset show that our model achieves a Dice of nearly 0.7, surpassing the second-best performing HarD-MSEG by 11.8%. This demonstrates the model's excellent generalization capabilities.

[0087] In addition, we conducted comparative experiments on the inference speed of the models. The table shows that U-Net++ is the fastest model, followed by HarD-MSEG. The model proposed in this paper is almost identical to HarD-MSEG in terms of inference speed. U-Net++'s high inference speed is due to its proposed pruning scheme, while the model proposed in this paper improves inference speed by reducing the number of connections and promptly releasing memory.

[0088] Figure 6 This is a qualitative comparison of the results of our method and other classic segmentation methods. From left to right, we see the original polyp image, the GT (Ground Truth), the segmentation result of our method, and the segmentation results of five comparison models. Examples include polyps of varying sizes, multiple polyps, and scenes with blurred boundaries such as light spots, dark areas, and debris. It can be seen that our model can effectively segment polyps in these situations, and the segmentation results show a higher similarity to the GT, demonstrating our model's superior boundary modeling capabilities and the ability to accurately locate and segment polyps in challenging situations.

[0089] Comprehensive experimental results show that the segmentation model proposed in this paper has good learning and generalization capabilities; in terms of quantitative indicators and qualitative segmentation results, compared with classic and popular segmentation methods in the field, it has higher precision, more accurate segmentation effects and faster inference speed, and can effectively and efficiently cope with various complex clinical segmentation scenarios.

[0090] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0092] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0093] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0094] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0095] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for intestinal polyp boundary segmentation based on sparse connectivity and global feature enhancement, characterized by: The method uses a sparse harmonic connection network as the encoder backbone to extract features from the input intestinal polyp image. The deepest layer features are sampled in parallel at different sampling rates using a dilated spatial convolutional pyramid pooling module. The results of the parallel sampling are then input into a partial decoder for decoding along with the features of the last two layers of the encoder to generate a rough initial prediction map. The sparse harmonic connection network includes several HarD Blocks, and the number of channels in each layer of any HarD Block is , where k is the initial growth rate, m is the dimension compression factor, for the layer, if Can be If n is divisible by N, then n takes the maximum value of N; the output of each HarD Block is the connection between the last layer and all odd layers, and the depth of HarD Block is a power of 2; The shallow attention module is used to extract the boundary information contained in the shallow features to preliminarily refine the boundaries of the initial prediction image; A global feature enhancement module is used to process deep features to guide subsequent boundary segmentation of the initial prediction image. The global feature enhancement module includes a global average pooling branch, a 3×3 adaptive pooling branch, a 5×5 adaptive pooling branch, and a feature map non-local branch. Deep features are input to each branch, and the obtained results are concatenated and output. The reverse attention module is used to obtain complete boundaries from multi-scale features and complete intestinal polyp boundary segmentation.

2. The intestinal polyp boundary segmentation method based on sparse connectivity and global feature enhancement according to claim 1, characterized in that: The features extracted from the sparse harmonic connectivity network are , where the shallow features are , the deep features are .

3. The intestinal polyp boundary segmentation method based on sparse connectivity and global feature enhancement according to claim 2, characterized in that: Get a rough initial prediction map through the partial decoder ,in, .

4. The intestinal polyp boundary segmentation method based on sparse connectivity and global feature enhancement according to claim 3, characterized in that: The shallow attention module constructs a Gaussian kernel , using shallow features Extract boundary details and apply the maximum function to increase The weight coefficient of the shallow attention module and the rough initial prediction map are added to obtain the initial prediction map of the preliminary refined boundary. .

5. The intestinal polyp boundary segmentation method based on sparse connectivity and global feature enhancement according to claim 4, characterized in that: The global feature enhancement module obtains .

6. The method for intestinal polyp boundary segmentation based on sparse connectivity and global feature enhancement according to claim 5, characterized in that: Enhanced global features The local features extracted by the last three layers of the encoder are 、 and Perform pixel-level addition as side input for subsequent boundary extraction , .

7. The method for intestinal polyp boundary segmentation based on sparse connectivity and global feature enhancement according to claim 6, characterized in that: The reverse attention weight of the reverse attention module , is the sigmoid function, express The upsampled output of Output feature Composed of deep features and reverse attention weights Multiplying together, we get ; The reverse attention module deletes the currently predicted area from each deep feature, guiding the entire network to discover complementary target areas and details in turn. The three reverse attention modules obtain a high-precision prediction map through three reverse operations, and finally complete the intestinal polyp boundary segmentation through the sigmoid operation.

8. The intestinal polyp boundary segmentation method based on sparse connectivity and global feature enhancement according to claim 1, characterized in that: The loss function of the network in the training of the method combines the weighted intersection-over-union loss and the binary cross entropy loss, which is ,in, represents the weighted IoU loss, represents the BCE loss of global constraints and local constraints.

9. The intestinal polyp boundary segmentation method based on sparse connectivity and global feature enhancement according to claim 1, characterized in that: The method uses four up-sampled network outputs with the same size as the labeled image GT 、 、 、 Calculate the total loss as ,in, 、 for 、 The upsampling result is obtained and the network parameters are updated through back propagation.

Images