Medical image segmentation method based on efficient visual state space decoding
By combining an efficient visual state space decoder with a VMamba encoder and an efficient Mamba block, the computational complexity and insufficient feature capture of convolutional neural networks and Transformer models in medical image segmentation are solved, achieving efficient and accurate medical image segmentation.
Patent Information
- Application Number
- CN202510792309.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-09-30
AI Technical Summary
Existing convolutional neural networks have difficulty capturing the global features of medical images. The Transformer model has too high computational complexity when processing large-size medical images, and the Mamba model has shortcomings in capturing local features, making it difficult to meet the needs of complex medical image segmentation.
An efficient visual state space decoder is adopted, combined with the VMamba encoder and the efficient Mamba block. Through the guided fusion module and the efficient upsampling block, the state transition process is dynamically adjusted to achieve global and local feature fusion under linear computational complexity, and a lightweight encoder is designed to reduce computational cost.
It significantly improves the accuracy and robustness of medical image segmentation, reduces computational overhead, improves the segmentation performance of the model in complex environments, and can efficiently process medical images of multiple modalities.
Smart Images
Figure CN120726322A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision technology and further relates to medical image segmentation technology. Specifically, it is a medical image segmentation method based on efficient visual state space decoding, which can be used to improve the segmentation accuracy and efficiency of medical images in clinical applications. Background Art
[0002] With the continuous development of computer-aided diagnosis (CAD), automatic medical image segmentation has become an indispensable tool in medical diagnosis and treatment. This technology enables pixel-level image classification and precise location of tumors, organs, and lesions, providing a reliable basis for clinical decision-making. However, due to the high precision required for medical image segmentation, even minor segmentation errors can lead to serious treatment risks, placing extremely high demands on segmentation accuracy and efficiency in clinical applications.
[0003] Traditional medical image segmentation methods based on convolutional neural networks (CNNs) (such as UNet++, V-Net, nnU-Net, I2UNet, and UNet) have made significant progress in local feature extraction. They have effectively improved segmentation accuracy through multi-level feature fusion and deep network structures. However, due to the limited receptive field of the convolution kernel, when processing high-resolution medical images (such as whole-slice pathology images or 3D MRI sequences), CNN has difficulty capturing long-range dependencies in the image, resulting in certain limitations in modeling global semantic information (such as organ boundaries and small lesions). To this end, some researchers have tried to improve this problem by introducing attention mechanisms and deepening the number of network layers, but this often significantly increases the number of model parameters and computational burden.
[0004] To overcome these challenges, the Visual Transformer (ViT) introduced the Transformer architecture to the field of medical image segmentation. This approach breaks the traditional paradigm of local feature extraction by segmenting the image into patches and leveraging a self-attention mechanism to model global context. However, because the Transformer must compute the relationship between every pair of pixels in the image, its computational complexity grows quadratically. This creates an extremely high computational burden when processing high-resolution medical images, which has become a major obstacle to its widespread application.
[0005] To address the tension between global modeling and computational efficiency, novel architectures based on state-space models (SSMs) have gained increasing attention in recent years. The SSM originated from the classic Kalman filter and has subsequently spawned a variety of efficient computational approaches, such as linear state-space layers, structured state-space sequence models (S4), diagonal state-space models, and S4D models. These approaches leverage convolutional or approximate linear computation to efficiently process sequential data across a variety of tasks and modalities, demonstrating particular strength in capturing long-range dependencies. Building on this strength, the Mamba model innovatively introduces time-varying parameters and a hardware-aware parallel scanning algorithm. Through a selective scanning mechanism, the state transition process is dynamically adjusted, significantly enhancing the ability to model long-range dependencies while maintaining linear computational complexity. However, the Mamba model still faces challenges in processing 2D medical images: its sequential modeling approach lacks the ability to capture local features, making it difficult to fully recover the subtle texture information required for medical image segmentation. Furthermore, some existing research has focused too heavily on the design of the Mamba module in the encoder, overlooking the crucial role of the decoder in gradually reconstructing encoder features into the final segmentation result. This deficiency may limit the overall performance of the model in complex medical image segmentation tasks. Summary of the Invention
[0006] The purpose of the present invention is to solve the problem that convolutional neural networks are difficult to capture the global features of medical images, and that the Transformer model has too high a quadratic computational complexity when processing large-scale medical images. To this end, a medical image segmentation method based on efficient visual state space decoding is proposed. In this method, the input medical image is first subjected to data enhancement operations such as random flipping and scaling; then, a pre-trained encoder is used to extract image features to obtain feature maps of different scales including four stages; finally, these feature maps are input into an efficient visual state space decoder to complete the feature decoding and image segmentation tasks. The present invention can achieve efficient segmentation of medical images of multiple modalities with a small amount of parameters and computing resources, greatly reducing the computational cost and effectively improving the segmentation performance of the model in complex environments.
[0007] To achieve the above object, the technical solution of the present invention includes the following steps:
[0008] (1) Obtain medical images from open source datasets on the Internet, perform preprocessing operations on them, and use the preprocessed images to construct a data sample set;
[0009] (2) An image segmentation model based on efficient visual state space decoding is constructed using a pre-trained VMamba encoder and an efficient visual state space decoder; the VMamba encoder includes four encoding layers from top to bottom, and its outputs correspond to the four inputs of the efficient visual state space decoder respectively; the efficient visual state space decoder includes four decoding units, which are recorded as the first, second, third and fourth decoding units from top to bottom, wherein the first to third decoding units have the same structure and are all composed of a guided fusion module, an efficient upsampling block and an efficient Mamba block; the fourth decoding unit is composed of only one efficient Mamba block; the efficient Mamba block is used to obtain the decoding feature map of the decoding unit in which it is located, and the efficient upsampling block is used to upsample the decoding features of the decoding unit below the decoding unit in which it is located; the guided fusion module is used to fuse the sampling results of the efficient upsampling block in the decoding unit in which it is located with the input coding feature map;
[0010] (3) Adjust the parameters of the VMamba encoder and load the pre-trained weights to extract the encoding features of the input medical image, and obtain the encoding feature maps of four stages through four encoding layers;
[0011] (4) feature decoding is performed on the encoded feature map of the fourth stage by the efficient Mamba block in the fourth decoding unit to obtain a fourth decoded feature map, which is used as the input of the efficient upsampling block in the third decoding unit and the output of the fourth stage;
[0012] (5) In the third decoding unit, the following operations are performed: first, the input decoding feature map is upsampled using the efficient upsampling block, and then the obtained result is input into the guided fusion module, and the fusion feature map of the upsampled feature map and the encoded feature map is obtained through the guided fusion module, and finally the fusion feature map is used as the input of the efficient Mamba block to obtain the decoding feature map; the decoding feature map is recorded as the third decoding feature map, which is used as the input of the efficient upsampling block in the second decoding unit and the output of the third stage respectively;
[0013] (6) The same operation as step (5) is performed in the second and first decoding units, and the obtained feature maps are recorded as the second and first decoding feature maps respectively; wherein the second decoding feature map is used as the input of the efficient upsampling block in the first decoding unit and the output of the second stage, and the first decoding feature map is directly used as the output of the first stage;
[0014] (7) The output of each stage is upsampled multiple times and summed pixel by pixel to obtain the segmented images of the four stages;
[0015] (8) Use the loss function to correct the segmented image, achieve model optimization, and use the optimized model to obtain the final image segmentation result.
[0016] Compared with the prior art, the present invention has the following advantages:
[0017] First, the present invention designs an efficient visual state space decoder, which includes an efficient upsampling module. This module implements feature map upsampling based on channel-dimensional pixel rearrangement technology, which can significantly reduce computational overhead while ensuring upsampling quality. At the same time, the decoder also includes a guided fusion module that dynamically fuses the detailed features from the encoder with the semantic features of the decoder, effectively bridging the gap between low-level visual information and high-level semantics, thereby improving the accuracy and robustness of the segmentation results.
[0018] Second, the decoder constructed by the present invention integrates an efficient Mamba module, which adopts a two-dimensional scanning mechanism with linear computational complexity and can effectively model the global dependencies in the image. At the same time, combined with deep convolution operations, it further enhances the local feature extraction capability and realizes the efficient fusion of global and local information.
[0019] Third, for the lightweight encoder with the number of output channels of [32, 64, 160, 256], the efficient visual state space decoder of the present invention only contains about 0.73M parameters; and under the standard encoder with the number of output channels of [64, 128, 256, 512], the number of decoder parameters is only 2.79M; it greatly reduces the computational cost and effectively improves the segmentation performance of the model in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is a flowchart of the overall implementation of the present invention;
[0021] Figure 2 Schematic diagram of the efficient visual state space decoding structure constructed in the present invention;
[0022] Figure 3 This is a diagram showing the segmentation effect of the present invention; different colors represent different categories in the Synapse dataset, and the white box is the highlighted part; in the Kvasir-Seg and ISIC2018 datasets, red represents over-predicted areas and blue represents missed areas. DETAILED DESCRIPTION
[0023] The present invention will be further described below with reference to the accompanying drawings.
[0024] Example 1: Refer to the attached Figure 1-2 The present invention proposes a medical image segmentation method based on efficient visual state space decoding, which specifically includes the following steps:
[0025] Step 1) Obtain medical images from an open source dataset on the Internet and perform preprocessing operations on them. The preprocessing operations used in this embodiment include random vertical flipping, random horizontal flipping, resizing, and normalization; and construct a data sample set using the preprocessed images.
[0026] Step 2) Use the pre-trained VMamba encoder and efficient visual state space decoder to build an image segmentation model based on efficient visual state space decoding; refer to, Figure 1 The VMamba encoder includes four coding layers from top to bottom, whose outputs correspond to the four inputs of the efficient visual state space decoder; Figure 2 The efficient visual state space decoder includes four decoding units, which are recorded as the first, second, third and fourth decoding units from top to bottom. The first to third decoding units have the same structure and are all composed of a guided fusion module, an efficient upsampling block and an efficient Mamba block; the fourth decoding unit is composed of only one efficient Mamba block; the efficient Mamba block is used to obtain the decoding feature map of the decoding unit in which it is located, and the efficient upsampling block is used to upsample the decoding features of the decoding unit below the decoding unit in which it is located; the guided fusion module is used to fuse the sampling results of the efficient upsampling block in the decoding unit in which it is located with the input encoding feature map;
[0027] Step 3) Adjust the parameters of the VMamba encoder and load the pre-trained weights, perform coding feature extraction on the input medical image, and obtain coding feature maps of four stages through four coding layers; in this embodiment, the parameters of the VMamba encoder include the input image size, the number of input channels, and the number of output channels.
[0028] Step 4) Feature decoding is performed on the encoded feature map of the fourth stage by the efficient Mamba block in the fourth decoding unit to obtain a fourth decoded feature map, which is used as the input of the efficient upsampling block in the third decoding unit and the output of the fourth stage;
[0029] Step 5) In the third decoding unit, the following operations are performed: first, the input decoding feature map is upsampled using the efficient upsampling block, and then the obtained result is input into the guided fusion module, and the fused feature map of the upsampled feature map and the encoded feature map is obtained through the guided fusion module, and finally the fused feature map is used as the input of the efficient Mamba block to obtain the decoding feature map; the decoding feature map is recorded as the third decoding feature map, which is used as the input of the efficient upsampling block in the second decoding unit and the output of the third stage respectively.
[0030] In this embodiment, the efficient upsampling block is used to upsample the input decoding feature map. The specific implementation steps are as follows:
[0031] (5a1) Use the deep convolution function to extract channel features of the input image, increase the number of channels to four times the original number, and obtain the expanded channel feature map;
[0032] (5a2) Using batch normalization function and ReLU activation function to process the dilated channel feature map to obtain a normalized feature map;
[0033] (5a3) Using the pixel rearrangement method, the pixels added to the channels of the normalized feature map are rearranged. Specifically, the pixels added to the channels are assigned to the length and width dimensions, so that the length and width of the feature map are increased to twice the original length and width, thereby obtaining a pixel rearranged feature map.
[0034] (5a4) Use 1×1 convolution to adjust the number of channels of the pixel rearrangement feature map to be consistent with the number of channels of the encoder feature map of its corresponding stage to obtain the upsampled feature map.
[0035] In this embodiment, the fusion module is guided to obtain the fused feature map of the upsampled feature map and the encoded feature map. The specific implementation steps are as follows:
[0036] (5b1) Add the encoded features and the upsampled decoded features pixel by pixel;
[0037] (5b2) Use the channel attention mechanism and spatial attention mechanism to process the summed feature map respectively, and then sum the results pixel by pixel;
[0038] (5b3) Merge the summed feature maps obtained in steps (5b1) and (5b2) in the channel dimension;
[0039] (5b4) After processing the merged feature map using the channel rearrangement method and 7×7 group convolution, the Sigmoid activation function is used to scale the pixel values of the feature map to [0, 1];
[0040] (5b5) The feature map obtained in step (5b4) and its compensation value are used as weights to multiply the original input encoding features and upsampled features pixel by pixel and then add them together;
[0041] (5b6) Use 1×1 convolution to adjust its channel number to the same as the input channel number to obtain the fused feature map.
[0042] In this embodiment, the above-mentioned decoding feature map is obtained according to the following steps:
[0043] (5c1) Using the instance normalization function and the ReLU activation function to process the input feature map to obtain a normalized feature map;
[0044] (5c2) Use the depth convolution function to process the normalized feature map to obtain a depth convolution feature map;
[0045] (5c3) Use the Mamba module in Vision Mamba to perform global feature extraction on the deep convolution feature map to obtain a global feature map;
[0046] (5c4) Perform a residual connection between the input feature map and the global feature map to obtain the decoded feature map.
[0047] Step 6) Perform the same operations as step 5) in the second and first decoding units, and record the resulting feature maps as the second and first decoding feature maps, respectively; the second decoding feature map serves as the input of the efficient upsampling block in the first decoding unit and the output of the second stage, and the first decoding feature map directly serves as the output of the first stage;
[0048] Step 7) The output of each stage is multiply upsampled and summed pixel by pixel to obtain segmented images of the four stages; this embodiment specifically uses an interpolation function and a 1×1 convolution function to adjust the size of the decoded feature map of each stage to the same as the initial model input, adjusts the number of channels to the total number of categories, and then adds and sums the decoded feature maps pixel by pixel.
[0049] Step 8) Use the loss function to correct the segmented image, optimize the model, and use the optimized model to obtain the final image segmentation result. The loss function used in this step of this embodiment includes: when targeting a two-category segmentation task, using a joint loss function consisting of binary cross entropy loss and Dice loss; for a multi-category segmentation task, using a joint loss function consisting of cross entropy loss and Dice loss; the formula is as follows:
[0050]
[0051] L seg =0.5L CE / BCE +0.5L Dice
[0052] Among them, L CE 、L BCE 、L Dice 、L seg Represents cross entropy loss, binary cross entropy loss, Dice loss and total segmentation loss during training; y i is the true label, Compared with the predicted results, C is the number of categories and N is the total number of pixels; Indicates the results corresponding to different predicted categories; L CE / BCE Indicates the use of L in the two-category segmentation task CE , use L in multi-category segmentation tasks BCE .
[0053] The effects of the present invention will be further described below in conjunction with simulation experiments.
[0054] 1. Simulation conditions:
[0055] The medical image segmentation method based on efficient visual state space decoding constructed in the present invention is implemented based on Pytorch and trained on an Nvidia GeForce RTX 3090 GPU equipped with 24GB of memory. The present invention uses random horizontal flipping, random vertical flipping and size scaling for data augmentation; in addition, all input images are adjusted to 256×256 pixels. For the Synapse dataset, the initial learning rate is set to 0.01, a multi-learning rate strategy is adopted, the maximum number of iterations is set to 400, and the batch size is set to 24. At the same time, a stochastic gradient descent (SGD) optimizer with a momentum of 0.9 and a weight decay of 1e-4 is used. For the Kvasir-Seg and ISIC2018 datasets, the initial learning rate is set to 5e-4, the maximum number of iterations is set to 300, and the batch size is set to 16. In addition, the AdamW optimizer and the CosineAnnealingLR scheduler are used simultaneously.
[0056] 2. Simulation content:
[0057] The effectiveness of the proposed method was validated using three medical image modalities: computed tomography (CT), endoscopic images, and dermoscopic images from the Synapse, Kvasir-Seg, and ISIC2018 datasets, respectively. The Synapse dataset contains 30 abdominal CT scans and 3,779 clinical axial abdominal CT images. The dataset was randomly divided into 18 scans for training and 12 for testing. Following previous methods, the segmentation performance of abdominal organs (aorta, gallbladder, spleen, left and right kidneys, liver, pancreas, and stomach) was evaluated using the Dice similarity coefficient (DSC) and mean Hausdorff distance (HD). The Kvasir-Seg dataset is a publicly available dataset specifically designed for endoscopic image segmentation tasks; it contains 1,000 gastrointestinal polyp images and their corresponding masks. ISIC2018 is the official dataset for the "Skin Lesion Analysis and Melanoma Detection" challenge at the 2018 International Symposium on Biomedical Imaging (ISBI). It contains 2594 RGB skin lesion images and their corresponding labels, and has become the main benchmark for evaluating skin lesion segmentation algorithms.
[0058] 3. Simulation results:
[0059] Figure 3The results of the present invention in multiple medical image segmentation tasks are demonstrated. For the Synapse dataset, different colors represent different organ categories, and white boxes are used to highlight key areas. For the Kvasir-Seg and ISIC2018 datasets, red areas indicate overprediction by the model, and blue areas indicate missed detections. As can be seen from the figure, the present invention can achieve accurate segmentation of medical images, effectively assisting doctors in completing tasks that traditionally require manual annotation and segmentation, significantly improving the efficiency and accuracy of clinical auxiliary diagnosis.
[0060] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.
[0061] The above simulation analysis proves the correctness and effectiveness of the method proposed in the present invention.
[0062] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.
[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.
Claims
1. A medical image segmentation method based on efficient visual state space decoding, characterized in that: The steps include: (1) Obtain medical images from open source datasets on the Internet, perform preprocessing operations on them, and use the preprocessed images to construct a data sample set; (2) An image segmentation model based on efficient visual state space decoding is constructed using a pre-trained VMamba encoder and an efficient visual state space decoder; the VMamba encoder includes four encoding layers from top to bottom, and its outputs correspond to the four inputs of the efficient visual state space decoder respectively; the efficient visual state space decoder includes four decoding units, which are recorded as the first, second, third and fourth decoding units from top to bottom, wherein the first to third decoding units have the same structure and are all composed of a guided fusion module, an efficient upsampling block and an efficient Mamba block; the fourth decoding unit is composed of only one efficient Mamba block; the efficient Mamba block is used to obtain the decoding feature map of the decoding unit in which it is located, and the efficient upsampling block is used to upsample the decoding features of the decoding unit below the decoding unit in which it is located; the guided fusion module is used to fuse the sampling results of the efficient upsampling block in the decoding unit in which it is located with the input coding feature map; (3) Adjust the parameters of the VMamba encoder and load the pre-trained weights to extract the encoding features of the input medical image, and obtain the encoding feature maps of four stages through four encoding layers; (4) feature decoding is performed on the encoded feature map of the fourth stage by the efficient Mamba block in the fourth decoding unit to obtain a fourth decoded feature map, which is used as the input of the efficient upsampling block in the third decoding unit and the output of the fourth stage; (5) In the third decoding unit, the following operations are performed: first, the input decoding feature map is upsampled using the efficient upsampling block, and then the obtained result is input into the guided fusion module, and the fusion feature map of the upsampled feature map and the encoded feature map is obtained through the guided fusion module, and finally the fusion feature map is used as the input of the efficient Mamba block to obtain the decoding feature map; the decoding feature map is recorded as the third decoding feature map, which is used as the input of the efficient upsampling block in the second decoding unit and the output of the third stage respectively; (6) The same operation as step (5) is performed in the second and first decoding units, and the obtained feature maps are recorded as the second and first decoding feature maps respectively; The second decoding feature map serves as the input of the efficient upsampling block in the first decoding unit and the output of the second stage, and the first decoding feature map serves directly as the output of the first stage; (7) The output of each stage is upsampled multiple times and summed pixel by pixel to obtain the segmented images of the four stages; (8) Use the loss function to correct the segmented image, achieve model optimization, and use the optimized model to obtain the final image segmentation result.
2. The method according to claim 1, wherein: The preprocessing described in step (1) includes random vertical flipping, random horizontal flipping, resizing and normalization.
3. The method according to claim 1, wherein: The parameters of the VMamba encoder in step (3) include the input image size, the number of input channels, and the number of output channels.
4. The method according to claim 1, wherein: In step (5), the efficient upsampling block is used to upsample the input decoded feature map. The implementation steps are as follows: (5a1) Use the deep convolution function to extract channel features of the input image, increase the number of channels to four times the original number, and obtain the expanded channel feature map; (5a2) Using batch normalization function and ReLU activation function to process the dilated channel feature map to obtain a normalized feature map; (5a3) using a pixel rearrangement method, rearrange the pixels added to the channel of the normalized feature map to obtain a pixel rearranged feature map; (5a4) Use 1×1 convolution to adjust the number of channels of the pixel rearrangement feature map to be consistent with the number of channels of the encoder feature map of its corresponding stage to obtain the upsampled feature map.
5. The method according to claim 4, characterized in that: The rearrangement in step (5a3) specifically allocates the added pixels on the channel to the length and width dimensions, so that the length and width of the feature map are increased to twice the original ones, thereby obtaining a pixel rearranged feature map.
6. The method according to claim 1, wherein: In step (5), the fusion feature map of the up-sampled feature map and the encoded feature map is obtained by guiding the fusion module. The implementation steps are as follows: (5b1) Add the encoded features and the upsampled decoded features pixel by pixel; (5b2) Use the channel attention mechanism and spatial attention mechanism to process the summed feature map respectively, and then sum the results pixel by pixel; (5b3) Merge the summed feature maps obtained in steps (5b1) and (5b2) in the channel dimension; (5b4) After processing the merged feature map using the channel rearrangement method and 7×7 group convolution, the Sigmoid activation function is used to scale the pixel values of the feature map to [0, 1]; (5b5) The feature map obtained in step (5b4) and its compensation value are used as weights to multiply the original input encoding features and upsampled features pixel by pixel and then add them together; (5b6) Use 1×1 convolution to adjust its channel number to the same as the input channel number to obtain the fused feature map.
7. The method according to claim 1, wherein: The decoding feature map in step (5) is obtained according to the following steps: (5c1) Using the instance normalization function and the ReLU activation function to process the input feature map to obtain a normalized feature map; (5c2) Use the depth convolution function to process the normalized feature map to obtain a depth convolution feature map; (5c3) Use the Mamba module in Vision Mamba to perform global feature extraction on the deep convolution feature map to obtain a global feature map; (5c4) Perform a residual connection between the input feature map and the global feature map to obtain the decoded feature map.
8. The method according to claim 1, wherein: As described in step (7), the output of each stage is upsampled multiple times and summed pixel by pixel. Specifically, the interpolation function and 1×1 convolution function are used to adjust the size of the decoded feature map of each stage to the same as the initial model input, the number of channels is adjusted to the total number of categories, and then the decoded feature maps are summed pixel by pixel.
9. The method according to claim 1, wherein: The loss function described in step (8) uses a joint loss function consisting of binary cross entropy loss and Dice loss for two-category segmentation tasks; for multi-category segmentation tasks, a joint loss function consisting of cross entropy loss and Dice loss is used; the formula is as follows: L seg (0.5L) CE / BCE +0.5L Dice Among them, L CE , L BCE , L Dice , L seg Represents cross entropy loss, binary cross entropy loss, Dice loss and total segmentation loss during training; y i is the true label, Compared with the predicted results, C is the number of categories and N is the total number of pixels; Indicates the results corresponding to different predicted categories; L CE / BCE Indicates the use of L in the two-category segmentation task CE , use L in multi-category segmentation tasks BCE .