Boundary guided attention based encoder-decoder medical image segmentation system and method

Abstract

The application relates to a boundary attention guiding-based encoder-decoder medical image segmentation system and method, and belongs to the technical field of medical image processing.The application comprises an encoder, a decoder, a feature enhancement module and a boundary attention guiding module.Different scale features are obtained in the encoder by inputting a medical image, and the features are output to the decoder; the decoder completes medical image segmentation according to the features; the feature enhancement module is further connected after the output of the different scale features of the encoder, and is used for enhancing the semantic features of the encoder, combining semantic information of different receptive fields, and inputting the feature output of the encoder into the decoder again after the feature output of the encoder passes through the feature enhancement module; the boundary attention guiding module is arranged between adjacent decoder modules in the decoder, and is used for strengthening boundary information of the medical image.The application is excellent in generalization ability and robustness, and the obtained medical image segmentation result is clear in boundary and obvious in small target.

Description

A Medical Image Segmentation System and Method Based on Boundary-Guided Attention Encoder-Decoder Technical Field

[0001] This invention belongs to the field of medical image processing technology, and relates to a medical image segmentation system and method based on a boundary-guided attention encoder-decoder. Background Technology

[0002] Convolutional neural network (CNN) models can extract image features through deep convolutional structures, enabling them to learn more complex and detailed image data representations. Fully Convolutional Networks (FCNs) are end-to-end CNN structures designed for image segmentation. In FCNs, fully connected layers are transformed into convolutional layers, allowing them to accept input images of arbitrary size and output segmentation results of the same size as the input image, thus achieving end-to-end pixel-level image segmentation. However, FCNs also have limitations. Multiple upsampling results in the segmented image losing much detail, leading to insufficient processing of image details and a tendency for blurring or edge blurring. Furthermore, the performance of FCN networks may be limited when dealing with complex backgrounds and noisy environments.

[0003] Unet is a standard encoder-decoder architecture. The Unet network structure has a natural advantage in medical image segmentation because medical images typically have complex structures and textures, and require the simultaneous segmentation of multiple different tissues or organs. The Unet encoder-decoder structure can better handle this multi-level, multi-scale information. Therefore, in the field of medical image segmentation, Unet has become the benchmark network for many algorithms, and many innovative methods based on Unet have been widely applied and achieved excellent results. The Unet encoder-decoder structure has also been widely innovated and applied. The following section will introduce several starting points for improving the encoder-decoder architecture.

[0004] The Unet++ proposed by ZHOU et al. integrates multiple Unets together, which share a single encoder. Their decoders are also flexibly organized and stacked, allowing them to share information and learn collaboratively. Furthermore, Unet++ redesigns skip connections; the new skip structure can fuse features at different scales, achieving better feature fusion results compared to the addition of features at the same level in Unets.

[0005] Fang et al. proposed a dual-decoder network (SFA) with region and boundary constraints, where two decoders share a single encoder. Skip connections are used between the encoder and the two decoders to fuse features at the same scale and upsampled high-level features. An SKM module is added to the encoder, and a unique loss function is added to the loss function for the decoder predicting boundaries, making the entire network more sensitive to boundaries and resulting in segmentation results with clear boundaries.

[0006] However, due to the characteristics of medical images, such as blurred boundaries, low image contrast, and significant variations in the shape, size, color, and texture of lesions or organs at different stages of image acquisition, most studies have shown poor segmentation results for small targets. This will have a significant impact on the early detection of diseases. Furthermore, boundary segmentation of lesion regions in medical images is challenging because visual interference exists between benign areas and lesion areas at their boundaries, while clear boundaries are crucial for segmentation. Therefore, existing segmentation methods still have considerable room for improvement. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a medical image segmentation system and method based on a boundary-guided attention encoder-decoder, for accurate segmentation of medical images, solving the problems of small target segmentation and unclear boundaries.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Scheme 1: A medical image segmentation system based on a boundary-guided attention encoder-decoder, comprising an encoder, a decoder, a feature enhancement module, and a boundary-guided attention module. The encoder processes the input medical image, obtaining features at different scales, and outputs these features to the decoder. The decoder performs medical image segmentation based on these features. A feature enhancement module is connected after the encoder's feature outputs at different scales to enhance the encoder's semantic features and combine semantic information from different receptive fields. The encoder's feature outputs then pass through the feature enhancement module before entering the decoder. The boundary-guided attention module is positioned between adjacent decoder modules in the decoder to strengthen the boundary information of the medical image.

[0010] Optionally, the encoder uses Res2Net as the feature extractor, and the medical image is extracted by the encoder to obtain features at 5 different scales. .

[0011] Optionally, the system includes two decoders that share one encoder; of the two decoders, decoder1 is used to predict the boundary and decoder2 is used to predict the lesion area; wherein, in decoder2, a boundary-guided attention module is provided between its adjacent decoder modules, and the input of the boundary-guided attention module is the output of the two decoders in the previous layer.

[0012] Each decoder consists of four decoder modules. In decoder1, the input of each decoder module is the output of the previous layer and the output of the feature enhancement module. In decoder2, the input of each decoder module is the output of the boundary-guided attention module and the output of the feature enhancement module.

[0013] Optionally, the boundary-guided attention module includes two convolutional layers, a channel attention module, and a spatial attention module. The features input to the boundary-guided attention module are sequentially passed through the two convolutional layers, the channel attention module, and the spatial attention module before being output.

[0014] Option 2: A medical image segmentation method based on a boundary-guided attention encoder-decoder. The method segments medical images according to the aforementioned system. Specifically, it involves first inputting the medical image into an encoder to extract five features at different scales. ,in These are low-level features. High-level features are generated; then, the low-level features are concatenated and input into the feature enhancement module, while the high-level features are input into the feature enhancement module separately. The outputs of the feature enhancement modules are then fed into two decoders, decoder1 and decoder2, respectively. In decoder2, a boundary-guided attention module is added, whose input is the outputs of the two decoder modules in the previous layer. The output of the boundary-guided attention module is concatenated with the output of the feature enhancement module and then fed into the next layer decoder module of decoder2. At the same time, deep supervision is introduced to supervise each decoder module of the two decoders. Finally, the output of decoder2 is used as the final medical image segmentation map.

[0015] Furthermore, in decoder1, the input of each decoder module is the concatenation of the output of the previous decoder module and the output of the feature enhancement module.

[0016] Furthermore, in the feature enhancement module, the input features are processed through four parallel residual branches. Each branch The following processing procedure is adopted:

[0017] 1) Use Convolution reduces the number of channels;

[0018] 2) After another Convolution and a dilation rate of Dilated convolution;

[0019] 3) Finally, it passes through a channel attention module;

[0020] The features of the four branches are concatenated along the channel dimension, through... Convolution reduces the number of channels to 32; then the output features are combined with the original features and passed through a... Features with reduced channel counts from convolution are summed and then processed through a BN layer and a ReLU function. The processed features are then passed through a spatial attention module to reduce the influence of irrelevant regions.

[0021] Furthermore, the boundary-guided attention module uses the output features of the decoder layer above decoder2. and the output features of the decoder layer above decoder1 As input, the two features are concatenated and passed through two convolutional layers to reduce the number of channels to 32. The features are then input into the CBAM module, and the output is obtained from the boundary-guided attention module.

[0022] Furthermore, the lesion segmentation loss function is:

[0023]

[0024] In the formula, and Let represent the predicted value and the actual value at position (i, j), respectively. Indicates the calculated weights. This represents the weighted IoU loss based on global constraints and local pixel-level constraints. This represents the binary cross-entropy (BCE) loss.

[0025] The loss function for boundary distribution map prediction is:

[0026]

[0027] In the formula, Indicates the location of the generated boundary map pixels, Indicates the location of the predicted boundary map pixels, This represents the threshold used when calculating the loss. hour, =1, hour, =0.

[0028] The beneficial effects of this invention are as follows: This invention proposes a dual-decoder structure, with the two decoders outputting image boundaries and lesion regions respectively. Simultaneously, this invention employs a boundary-guided attention module to fuse the information stream containing boundary features from one decoder into the stream of the other decoder, making the final image segmentation result more focused on the boundary region. Furthermore, this invention uses an adaptive feature enhancement module to enhance the semantic features of the encoder, combining semantic information from different receptive fields. This invention demonstrates excellent generalization ability and robustness, obtaining medical image segmentation results with clear boundaries and obvious small targets.

[0029] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0031] Figure 1 is a system architecture diagram of the present invention;

[0032] Figure 2 is a schematic diagram of the feature enhancement module structure;

[0033] Figure 3 is a schematic diagram of the boundary-guided attention module structure;

[0034] Figure 4 is a schematic diagram of in-depth supervision;

[0035] Figure 5 shows a visual comparison of different segmentation methods on polyp images;

[0036] Figure 6 shows a visual comparison of different segmentation methods on skin images. Detailed Implementation

[0037] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0038] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0039] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0040] Figure 1 shows the system architecture of the present invention. The present invention is based on the BGAUnet neural network model. BGAUnet consists of an encoder and two parallel decoders, with the two decoders sharing one encoder.

[0041] In BGAUnet, the encoder uses Res2Net as the feature extractor, assuming the input image size is... Then, the encoder will extract the resolution as follows: Five different scale features , These are low-level features. These are high-level features. Features at all five scales will serve as input to both decoders before entering the decoders. First, it goes through the Feature Enhancement Module (FEM), where low-level features are concatenated before being fed into the FEM.

[0042] Two decoders, decoder1 and decoder2, share a single encoder, enabling them to jointly learn and utilize features extracted by the encoder to generate predictive outputs. Decoder1 predicts boundaries, while decoder2 predicts lesion regions. Each decoder contains four decoder modules. The input to each decoder comes from the output of the previous layer and the output of the skip connections (FEM). Specifically: in decoder1, each decoder module receives the output of the previous layer and the output of the feature enhancement module; in decoder2, each decoder module receives the output of the Boundary Guided Attention Model (BGA) module and the output of the feature enhancement module. The two decoders are not independent; the BGA modules act as a bridge connecting them. Decoder2 contains three BGA modules, whose inputs come from the outputs of decoder1 and decoder2. The information generated by decoder1 guides decoder2, which uses contextual information generated by the BGA module to better segment the image. In this way, the two decoders can cooperate with each other to complete a more accurate and robust medical image segmentation task.

[0043] FEM is an improvement upon RFB (Receptive Field Block). FEM adds channel attention (CA) at each scale for additional global refinement, and adds spatial attention (SA) after the aggregated features to enhance the model's focus on spatial information of the image, as shown in Figure 2. FEM contains four parallel residual branches. Each branch First use Convolution reduces the number of channels, followed by two more convolution steps, i.e., one... Convolution and a dilation rate of Dilated convolution. A CA module is added at the end of each branch, and then the features from the four branches are concatenated along the channel dimension. Convolution reduces the number of channels to 32. Then, the output features are combined with the original data and passed through a... Features with reduced channel counts from convolution are summed and then processed using a Batch Normalization (BN) layer and a ReLU function. The processed features are then augmented with an Automatic Subtraction (SA) mechanism to reduce the influence of irrelevant regions. In addition, Inception-V3 proposes: a... Convolution can be and Replacing it with two convolutions not only avoids loss of accuracy but also speeds up inference; therefore, in this invention... Convolution was replaced with and Two convolutions.

[0044] Figure 3 shows a schematic diagram of the boundary-guided attention module structure. The boundary distribution map represents the probability map in the feature map that the current pixel belongs to the boundary. It is assumed that the boundary distribution has a mean of 0 and a standard deviation of... If the distribution follows a Gaussian distribution, then the boundary distribution diagram can be represented as:

[0045]

[0046] In the formula, Represents pixels The shortest Euclidean distance to the boundary; This represents the standard deviation. BGAUnet predicts the boundary distribution map in decoder1 and uses the boundary distribution map to calculate the loss function. Each branch of decoder1 is used to predict the boundary distribution map.

[0047] In traditional attention mechanisms, models typically calculate attention based on pixel-level weights. This can easily lead to attention being scattered across regions inside the object, neglecting details at the object's boundaries. BGA, however, incorporates boundary information before the attention calculation, thus better focusing on boundary regions. The BGA module uses the output features of the decoder layer above decoder2. and the output features of the decoder layer above decoder1 As input, the two features are concatenated and then passed through two convolutional layers (Conv3×3+BN+ReLU) to reduce the number of channels to 32. Next, the features are input into the CBAM module, and the output is the result of the BGA module. The BGA module can introduce boundary information from decoder1 to decoder2, while the spatial and channel attention modules in CBAM can effectively improve the model's attention to specific spatial locations and important channels. This not only compensates for the boundary information lost due to downsampling but also enriches semantic features, resulting in better segmentation results.

[0048] The lesion segmentation loss function of this invention is:

[0049]

[0050] In the formula, and Let represent the predicted value and the actual value at position (i, j), respectively. Indicates the calculated weights. This represents the weighted IoU loss based on global constraints and local pixel-level constraints. This represents the binary cross-entropy (BCE) loss.

[0051] The loss function for boundary distribution map prediction is:

[0052]

[0053] In the formula, Indicates the location of the generated boundary map pixels, Indicates the location of the predicted boundary map The pixels. It is a specific function, when x is true, When x is false, . This represents the threshold used when calculating the loss; the probability that the current pixel belongs to the boundary is less than [a certain value]. When the expression is true, it is false and contributes nothing to the loss value; conversely, when the expression is true, it contributes to the loss value. Therefore, in the above formula, hour, =1, hour, =0.

[0054] The total loss function of this invention is shown in the following formula:

[0055]

[0056] In the formula, It is the overall loss function; It is the overall loss of the boundary segmentation decoder (decoder1); This represents the overall loss of decoder2, which segments the lesion region. Deep supervision is used in this invention for the four outputs of decoder1. All samples are upsampled to the same size as the true image G of the lesion region segmentation; for the four outputs of decoder2 All upsampled to the boundary ground truth map Same size.

[0057] Figure 4 shows a schematic diagram of the deep supervision structure used in this invention. Deep supervision is a deep learning model training technique aimed at improving the model's learning efficiency and generalization performance. In this invention, deep supervision is applied to both decoders during training. Specifically, for decoder1, boundary distribution maps are used for supervision at four scales; for decoder2, the true segmentation labels are used for supervision at four scales.

[0058] In this embodiment, the present invention is used to perform polyp segmentation on medical images, and compared with several representative polyp segmentation methods, including Unet, Unet++, SFA, Pranet, and BDG-Net. This embodiment uses an RTX A5000 graphics card and PyTorch 1.7 as the Python code framework. The model is trained for 150 epochs with a batch size of 16. During training, AdamW is used as the optimizer with a momentum of 0.9 and weight decay. The learning rate is All images were adjusted before being input into the network. The size of the dataset is the same during the testing phase. For data augmentation, a multi-scale training method with scales [0.75, 1, 1.25] was used, along with random cropping, vertical flipping, horizontal rotation, Gaussian noise, discarding pixels from a specific channel at a specific location, and contrast transformation. During training, the code saves the network parameters that perform best on the test set.

[0059] This embodiment uses five polyp segmentation datasets: Kvasir-SEG, ClinicDB, ColonDB, ETIS, and Endoscene. These five datasets have been widely used in segmentation tasks in recent years to evaluate the effectiveness of network models. The Pranet configuration is used for splitting the training and test sets for polyp segmentation. The ISIC2018 skin lesion segmentation dataset is also used; in this embodiment, all 2594 images in the ISIC2018 dataset are adjusted to... The dataset was set to a size and randomly allocated to the training and test sets in an 8:2 ratio. The training set contained 2076 images, and the test set contained 518 images. The dataset is shown in Table 1.

[0060] Table 1 Dataset

[0061]

[0062] For polyp segmentation, this embodiment uses two widely used evaluation metrics: Dice and IoU. Specifically, the average values ​​of Dice and IoU, mDice and mIoU, are used as evaluation metrics. For fair comparison, all evaluation metric code uses MATLAB code from PraNet. For skin segmentation, this embodiment uses five common skin segmentation evaluation metrics: Recall, Precision, Accuracy, Dice, and IoU. Polyp segmentation methods include Unet, Unet++, SFA, PraNet, and BDG-Net. The results of these comparison methods are directly obtained from their respective literature, provided that the training and test sets are the same, and the test code is also the same. If no direct results are available, calculations are performed using their published code.

[0063] For skin lesion segmentation, this embodiment selects several representative skin lesion segmentation methods for comparison, including Unet, Unet++, ResUnet, At-Unet, ResUnet++, CPFNet, and CA-Net. Since the dataset settings for different methods differ, for a fair comparison, this embodiment will retrain the model code of the comparison methods under the same training settings and save the parameters of the best experimental results. These comparison methods are shown in Table 2:

[0064] Table 2 Comparison Methods

[0065]

[0066] Comparison results:

[0067] The comparison results for polyp segmentation are shown in Table 3. Among the five datasets, this invention achieved the best results on three datasets. On ClinicDB, it was 1.5% better than the second-best dataset. On the ETIS dataset, which contains many small objects, the accuracy of this invention was significantly improved. mDice and mIoU were 7.1% and 8.5% higher than the second-best dataset, respectively. Meanwhile, on the other two datasets, this invention also achieved second-best performance. This demonstrates that this invention performs well in both overall segmentation and small object segmentation.

[0068] Table 3 Performance comparison of different segmentation models on the polyp dataset

[0069]

[0070] The comparison results for skin lesion segmentation are shown in Table 4. The present invention outperforms all other methods in five evaluation metrics. The scores for the five metrics are 0.913, 0.922, 0.961, 0.904, and 0.887, respectively, which are 0.3%, 0.77%, 0.1%, 0.3%, and 0.2% higher than the second-ranked method. The comparative data demonstrate the effectiveness of the present invention.

[0071] Table 4. Performance comparison of different methods on skin datasets

[0072]

[0073] This embodiment uses FLOPs and the number of parameters to evaluate the computational complexity of the model, as shown in Table 5. FLOPs are... The results were calculated at the input resolution. As can be seen from the table, the computational time complexity of this invention ranks second among all methods, but it has a high number of parameters.

[0074] Table 5 shows the average inference speed of different models. 'N / A' indicates that data is unavailable.

[0075]

[0076] Figure 5 shows a visual comparison of different segmentation methods on polyp images, where GT represents the ground truth label and Image represents the input image. The figure shows that for rows 2, 3, and 4, BGAUnet can segment small objects clearly compared to other methods; for rows 1 and 5, BGAUnet can obtain clear boundaries, proving that the boundary prediction branch of BGAUnet can provide boundary information for lesion region segmentation and obtain clear segmentation results. Figure 6 shows the segmentation results of BGAUnet on the ISIC2018 dataset, with Unet and Unet++ selected for comparison. The figure shows that the boundaries of this invention are obvious, the difference between the foreground and background is large, and there are no blurred boundaries.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A medical image segmentation system based on boundary-guided attention encoder-decoder, characterized in that: The system includes an encoder, a decoder, a feature enhancement module, and a boundary-guided attention module. The encoder processes the input medical image, obtains features at different scales, and outputs them to the decoder. The decoder performs medical image segmentation based on the features. The feature enhancement module is connected after the encoder's feature outputs at different scales to enhance the encoder's semantic features and combine semantic information from different receptive fields. The encoder's feature outputs then enter the decoder. The boundary-guided attention module is positioned between adjacent decoder modules in the decoder to strengthen the boundary information of the medical image. The system includes two decoders sharing one encoder. Of the two decoders, decoder1 predicts boundaries, and decoder2 predicts lesion regions. Decoder1 and decoder2 each include four decoder modules. In decoder1, the input of each decoder module is the output of the previous layer and the output of the feature enhancement module. In decoder2, a boundary-guided attention module is provided between adjacent decoder modules. The input of each decoder module is the output of the boundary-guided attention module and the output of the feature enhancement module. The input of the boundary-guided attention module is the output of the two decoders in the previous layer. The boundary-guided attention module includes two convolutional layers, a channel attention module, and a spatial attention module. The features input to the boundary-guided attention module are sequentially passed through two convolutional layers, a channel attention module, and a spatial attention module before being output.

2. The medical image segmentation system according to claim 1, characterized in that: The encoder uses Res2Net as the feature extractor, and the medical image is extracted by the encoder to obtain features at five different scales. 。 3. A medical image segmentation method based on boundary-guided attention encoder-decoder for the system of claim 1 or 2, characterized in that: The method specifically involves: firstly, extracting five features at different scales from the medical image input encoder. ,in These are low-level features. The low-level features are then concatenated and input into the feature enhancement module, while the high-level features are input into the feature enhancement module separately. The outputs of the feature enhancement modules are then fed into two decoders, decoder1 and decoder2, respectively. In decoder2, a boundary-guided attention module is added, whose input is the outputs of the two decoder modules in the previous layer. The output of the boundary-guided attention module is concatenated with the output of the feature enhancement module and then fed into the next layer decoder module of decoder2. Simultaneously, deep supervision is introduced to supervise each decoder module of the two decoders. Finally, the output of decoder2 serves as the final medical image segmentation map.

4. The medical image segmentation method according to claim 3, characterized in that: In decoder1, the input of each decoder module is the concatenation of the output of the previous decoder module and the output of the feature enhancement module.

5. The medical image segmentation method according to claim 3, characterized in that: In the feature enhancement module, the input features are processed through four parallel residual branches. Each branch The following processing procedure is adopted: 1) Use Convolution reduces the number of channels; 2) then through another... Convolution and a dilation rate of Dilated convolution; 3) Finally, it passes through a channel attention module; the features from the four branches are concatenated along the channel dimension, and then... Convolution reduces the number of channels to 32; then the output features are combined with the original features and passed through a... Features with reduced channel counts from convolution are summed and then processed through a BN layer and a ReLU function. The processed features are then passed through a spatial attention module to reduce the influence of irrelevant regions.

6. The medical image segmentation method according to claim 3, characterized in that: The boundary-guided attention module uses the output features of the decoder layer above decoder2. and the output features of the decoder layer above decoder1 As input, the two features are concatenated and passed through two convolutional layers to reduce the number of channels to 32. The features are then input into the CBAM module, and the output is obtained from the boundary-guided attention module.

7. The medical image segmentation method according to claim 3, characterized in that: The lesion segmentation loss function is: In the formula, and Let represent the predicted value and the actual value at position (i, j), respectively. Indicates the calculated weights. This represents the weighted IoU loss based on global constraints and local pixel-level constraints. The binary cross-entropy (BCE) loss function is used; the boundary distribution plot prediction loss function is: In the formula, Indicates the location of the generated boundary map pixels, Indicates the location of the predicted boundary map. pixels, This represents the threshold used when calculating the loss. hour, =1， hour, =0。

Images