
==== Front
Heliyon
Heliyon
Heliyon
2405-8440
Elsevier

S2405-8440(24)12150-0
10.1016/j.heliyon.2024.e36119
e36119
Research Article
Application of the bicharacteristic attention residual pyramid for the treatment of brain tumors
Hang Yiliu 18862928527@163.com
⁎
Zhang Qiong
Li Xiujing
Qiu Jianlin
College of Computer and Information Engineering, Nantong Institute of Technology, Nantong, Jiangsu, 226000, China
⁎ Corresponding author. 18862928527@163.com
13 8 2024
30 8 2024
13 8 2024
10 16 e3611930 5 2024
9 8 2024
9 8 2024
© 2024 The Author(s)
2024
https://creativecommons.org/licenses/by-nc/4.0/ This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
Currently, surgery remains the primary treatment for craniocerebral tumors. Before doctors perform surgeries, they need to determine the surgical plan according to the shape, location, and size of the tumor; however, various conditions of different patients make the tumor segmentation task challenging. To improve the accuracy of determining tumor shape and realizing edge segmentation, a U-shaped network combining a residual pyramid module and a dual feature attention module is proposed. The residual pyramid module can enlarge the receptive field, extract multiscale features, and fuse original information, which solves the problem caused by the feature pyramid pooling where the local information is not related to the remote information. In addition, the dual feature attention module is proposed to replace the skip connection in the original U-Net network, enrich the features, and improve the attention of the model to space and channel features with large amounts of information to be used for more accurate brain tumor segmentation. To evaluate the performance of the proposed model, experiments were conducted on the public datasets Kaggle_3M and BraTS2021. Because the model proposed in this study is applicable to two-dimensional image segmentation, it is necessary to obtain the crosscutting images of fair class in the BraTS2021 dataset in advance. Results show that the model accuracy, Jaccard similarity coefficient, Dice similarity coefficient, and false negative rate (FNR) on the Kaggle_3M dataset are 0.9395, 0.8812, 0.8958, and 0.007, respectively. The model accuracy, Jaccard similarity coefficient, Dice similarity coefficient, and FNR on the BraTS2021 dataset were 0.9375, 0.9072, 0.8981, and 0.0087, respectively. Compared with existing algorithms, all the indicators of the proposed algorithm have been improved, but the proposed model still has certain limitations and has not been applied to actual clinical trials. For specific datasets, the generalization ability of the model needs to be further improved. In the future work, the model will be further improved to address the aforementioned limitations.

Keywords

Brain tumor segmentation
Residual pyramid
Dual feature
Attention mechanism
==== Body
pmc1 Introduction

Brain tumors can originate from various intracranial tissues, accounting for ∼2 % of all tumor cases [1]. With continuous development of medical imaging, magnetic resonance imaging (MRI) of the brain is often used in clinical diagnosis [2]. Currently, surgical treatment is still the most effective treatment for craniocerebral tumors [3]. To enhance the success rate and ensure safe surgical procedures, localizing and segmenting brain tumors are essential. If segmentation relies solely on experienced doctors, the results may be time consuming and show high levels of subjectivity [4]. A branch of medical image processing in computer vision uses deep learning to segment the shape of brain tumors automatically, efficiently, and accurately. Results show clear boundaries and can be used to locate tumors, providing strong support for surgical treatment [5].

Convolutional neural networks (CNNs) [6] are capable of extracting image features and have strong robustness in medical image applications. However, most convolutional networks treat image pixels as independent labels and lack contextual semantic connections [7]. To enhance the semantic feature connections of CNNs and improve model segmentation accuracy, Zhu [8] proposed a network composed of a semantic segmentation module, an edge detection module, and a feature fusion module. He used Swin transformer to extract semantic features and edge space attention block for feature enhancement. Swin transformer requires a large amount of computation. Liu [9] improved the Swin transformer by introducing multihead self-attention and designing a layered Swin-T network to improve computational efficiency. Xu [10] combined the Swin-T network and a dual-path feature inference module to optimize the original Swin-T network, with higher edge-segmentation ability for cranial tumors. The U-Net [11] employs a coding–decoding architecture and integrates high–low-resolution skip connections. This design allows it to preserve the original image features and achieve performance in medical image segmentation [12]. However, U-Net uses the CNN as its backbone; it cannot establish a global model with context connection, and in large lesion areas, and it is not sensitive to segmentation [13]. Residual networks [14] can solve the problem of gradient disappearance and descending while preserving deep image features. Xiao X [15], Phani N B [16], and Yin-yin W [17] et al. combined U-Net with residual connections, which could segment small and fuzzy features; however, increasing network depth would lead to numerous parameters and complicated calculations. Luo [18] proposed a dual-channel U-Net model (DC-U-Net) to replace the original encoding and decoding structure with an efficient CNN architecture. In this model, the skip connection from the U-Net network is replaced by a residual module and the multilayer convolution is replaced with a MutiRes module. The multiscale residual connection can learn additional spatial features and adapt to changes in image scale. Dense connections strengthen and reuse features by maximizing data flow, which is conducive to training deep networks. Karali [19] proposed DR-Vnet, adding dense connections into the first layer and residuals into other layers to enhance transmission between features and better segment features.

The brain MRI image segmentation includes three-dimensional (3D) image segmentation, multimodal image segmentation, and two-dimensional (2D) image segmentation. Based on the U-Net structure, Zhu [20] adopted a sparse dynamic codec module for global spatial semantic feature extraction and a multilevel edge feature fusion module for multilayer feature fusion to extract more voxels, interlayer connections, and intra-axis information contained in 3D volume. Fusion information is connected to the encoder module by a skip connection, which enhances the ability of spatial edge information transmission. Zhu [21] paid attention to the information differences in multimodal cranial images and proposed a multimodal spatial information enhancement (SIE) and boundary shape correction (BSC) method that used the modal information extraction module, the SIE module, and the BSC module to extract multimodal and spatial information. This method is advantageous for segmenting key tumor boundary locations. Huang [22] proposed FEC-NET; to make better use of image features and context information, an atrous spatial pyramid pool (ASPP) module was added to U-Net to enrich multiscale features by parallel convolution of different sampling rates, fully capture fine context information, and process lesion segmentation of different sizes to restore lesion boundaries. Zhu [23] proposed an improved U-Net network. ResNet50 was used as the main backbone network, combined with the CBMA module, and the cross-entropy loss function and Dice similarity coefficient were mixed into the loss function of the network. This method was mainly applied to category-unbalanced datasets and could improve the segmentation effect of tumor regions. Inspired by symmetric U-Net, Aboussaleh [24] replaced the bidirectional feature pyramid neural network used in the original jump connection stage with the initial feature block to more accurately locate the tumor location. Sun [25] enhanced U-Net by incorporating a multiview attention mechanism and multiscale feature interaction. This modification allowed the model to capture local features from the three perspectives of channel, content, and position, resulting in improved features within the focal area. Using convolution of different receptive fields to extract multiscale features and interact with each other, the tumor edge contour can be accurately segmented.

On this basis, this study proposes an improved U-Net network that combines the residual space pyramid and a dual feature attention module. In the coding part, this study mainly introduces the residual cavity space pyramid module, which is used to replace the original convolution operation. This module used multiscale cavity convolution to obtain different sensitive field features and performed residual fusion, which could enhance feature extraction and context connection. Dual feature fusion was used to replace the original skip connection, and downsampled and upsampled features were complemented to solve the problem of feature detail loss. Attention modules were added to perform attention fusion in space and channels, and the model focused on tumor region and edge segmentation. Based on the aforementioned techniques, the proposed brain tumor segmentation framework combining the residual cavity pyramid module and the dual feature attention module can effectively extract detailed information from the feature map to enhance the context connections and improve the accuracy of tumor edge segmentation.

The main contributions of this study are given as follows.(1) In the encoding part, after convolution of each layer of subsamples, the original 3 × 3 convolution is replaced by the hollow pyramid pool module with the fusion of residual ideas, which can extract multiscale feature information, enrich feature details, and enhance the connection between feature information via residual fusion.

(2) In the middle clipping and splicing part, double feature fusion is used to merge maximum pooling and average pooling and enhance the background image information and the texture characteristics of tumor soft tissue. This step lays a foundation for extracting foreground features during upsampling.

(3) Bicharacteristic attention modules are added in upsampling to integrate characteristics of channel attention and spatial attention, used to improve detailed characteristics of tumor soft tissue, enhance segmentation performance, and better segment the tumor edge.

2 Related work

2.1 U-Net

The U-shaped network structure is divided into two parts: the downsampling (encoding) part and the upsampling (decoding) part. These parts are related by skip connections. (1) Downsampling: the contraction path consists of convolution, activation function, and pooling. A convolution layer can extract local features of images, and the activation function will linearly map the extracted results. Generally, the ReLU function is used to reduce dimension and compress extracted features. Maximum pooling is usually selected. (2) Upsampling: the expansion path consists of convolution and deconvolution. Pooling after downsampling will reduce the image size and lose a part of image features. A deconvolution layer is required to expand the image size and restore features. (3) Skip connection: the U-Net network adopts feature dimension splicing to mix features of high- and low-level images.

During downsampling, U-Net captures global feature information, enabling accurate localization of the foreground. Skip connection is used to solve the problem of partial spatial information loss during upsampling. However, this method has problems such as underlying information redundancy, huge calculation amounts, and ineffective calculation [26]. The basic structure of the U-Net network is shown in Fig. 1.Fig. 1 Basic structure of the U-Net network.

Fig. 1

2.2 Residual network

As deep learning models add network depth to enhance network performance, the risk of gradient dispersion or gradient explosion also grows with deepening of the network [27]. This problem can be solved by batch normalization, which increases the depth of the network to dozens of layers; however, there may be degradation problems in the training process. The accuracy of training rate saturation may decline with increasing number of network layers. To solve this problem, residual modules are added to the network. When degradation occurs in a deep network, the model becomes a shallow network and the problem will revert to how the model learns to fit the identity map during continuous training. The residual idea converts the identity mapping network into a residual function, making it easier to fit the model in subsequent training processes to solve the degradation problem [28]. The conversion steps for the residual idea are as follows.Step 1 Network potential identity mapping function: F(x)=x (1)

Step 2 Function conversion: F(x)=x→F(x)=G(x)+x (2)

Step 3 Convert residual function: G(x)=F(x)−x (3)

Step 4 Identity mapping transformation: G(x)=0 (4)

By converting the identity mapping fitting in step 1 to the mapping whose result is 0 in step 4, the output results can become more sensitive to changes. It has a greater effect on the adjustment of network weights and highlights small changes in structure to achieve better training results. The common residual network structure in U-Net is shown in Fig. 2.Fig. 2 Residual structure in U-Net.

Fig. 2

2.3 Atrous spatial pyramid pool

2.3.1 Receptive field

In CNNs, the size of the input layer determines the output result element of a certain network layer, which is called the receptive field. The original image range will expand with a larger receptive field, allowing for the extraction of more comprehensive feature information that is close to the global feature but easy to ignore. As the receptive field becomes smaller, the scope of the original image narrows, capturing feature information closely tied to local details. However, this reduction in scale may lead to the omission of large-scale information.

The U-Net is used for image segmentation. Convolution and pooling operations increase the receptive field in the downsampling phase while simultaneously reducing the size of the feature map. Consequently, during the upsampling process, the corresponding size enlargement occurs. However, this reverse operation process will reduce the accuracy of feature information.

2.3.2 Atrous spatial pyramid

The addition of the hyperparameter “dilation rate” to convolution can increase the receptive field without altering the size of the output feature map of the image. This technique is known as atrous convolution. Atrous convolution uses sparse sampling for feature input, which easily leads to a lack of correlation between feature information obtained by remote partial convolution of the original image.

Spatial pyramid pooling (SPP) leverages convolutions at different scales to extract features. These features are then consolidated into a fully connected layer, addressing the challenge posed by varying image scales. ASPP enhances the SPP approach by introducing multiple parallel convolutions with varying dilation rates. ASPP constructs convolution kernels with different sensitivity fields to capture feature information across different scales. Finally, ASPP combines this information to create a detailed feature map [29]. The atrous spatial pyramid structure is shown in Fig. 3.Fig. 3 Atrous spatial pyramid structure.

Fig. 3

In the convolution part of the first layer, dilation rates of 1, 6, 12, and 18 are employed; these correspond to receptive fields of 1 × 1, 13 × 13, 25 × 25, and 37 × 37, respectively. To retain global features of the image, pooling–convolution–upsampling operation is conducted to connect all feature maps obtained in part 5. The number of channels is increased by five times; a 1 × 1 convolution changes channel numbers as the same as input.

2.4 Attention mechanism

The attention mechanism originates from the study of human vision, which makes the model focus on key parts of input data. In the neural network, when processing input data, different weights can be assigned to different parts of the input data to better capture the relevant information of the input data and improve the accuracy and efficiency of the model [30]. A channel attention mechanism is a channel-based attention mechanism that assigns different importance to different channels in a feature graph to improve the ability of the model to perceive different features. The structure of the channel attention mechanism is shown in Fig. 4.Fig. 4 Channel attention mechanism.

Fig. 4

In the picture, the size of the feature graph input into the channel attention mechanism module is H × W, and the number of channels is C. After the global maximum pooling and global average pooling modules, a matrix of 1 × 1 × C is obtained, which is a scalar. Obtained results are sent into a shared multilayer perceptron, and two feature graphs of 1 × 1 × C are obtained. Finally, after the excitation function sigmoid, the parameter value of 1 × 1 × C is obtained, which is the attention of the feature graph of different channels. The original input image is multiplied by the weight value to form the feature result graph with the attention of different channels.

3 Bicharacteristic attention residual pyramid model

3.1 Residual pyramid

Combined with the idea of the atrous spatial pyramid and residual module, the residual pyramid (RASPP) module is proposed. Based on the original atrous convolution, the result obtained from each convolution is superimposed with the upper result, affording multiple level features on one layer and enhancing feature extraction by comparing with the original structure. The connection between different spatial features is improved, the structure of the original image is preserved, and the problem of easily loosing local information with no relation with remote information is solved using this method. The residual pyramid (RASPP) structure is shown in Fig. 5.Fig. 5 Residual pyramid structure.

Fig. 5

Batch normalization is performed before each layer of the residual pyramid, which can prevent gradient disappearance. This process can speed up network training and model convergence. A 1 × 1 convolution in the original structure is replaced by a 3 × 3 convolution, and the dilation rate is changed to 3. The atrous convolution is an ordinary convolution when the dilation rate is 1. To obtain more scale feature information, this part was changed. Upper and lower layers of image feature information were processed by pixel addition, and feature information was fixed to enhance connections between features. Concurrently, 2 × 2 average pooling was used to retain the global information of feature images. As the image size was reduced to half after pooling operation, it was recovered by upsampling. Bilinear interpolation was used in this part. After this step, the final result is the same size as the original input and the number of channels becomes five times the original. The specific steps are as follows:

Initially, define a feature graph RW×H×C, with C channels, whose width and height are W and H.

Step 1:(1) outputn1=BN(FFeature)

where FFeature∈RW×H×C, BN(x) is the batch normalization function, and outputn1 represents the first calculation output result of the n layer.

Step 2:(2) outputn2={AVG(outputn1),n=1∑k=1Coutput(n−1)2k(x,y)+outputn1K(x,y),n>1,x∈H,y∈W

where outputk∈RW×H represents the k channel feature of the feature graph and AVG(x) represents the average pooling operation.

Step 3:(3) outputn3=conv(outputn2),n>1

where conv(x) denotes convolution operation.

Step 4:(4) output=concat(output12,output23,output33,output43,output53)

where output indicates the final output and concat(x) indicates the channel concatenation operation.

3.2 Bicharacteristic attention mechanism

3.2.1 Bicharacteristic fusion

During the downsampling process in U-Net, the risk of reduced accuracy increases with increasing image size. To solve this problem, sampling results under the same layer are modified in size and connected with channels, but only channel thickening of the feature image is done; the relationship between features of various parts is not obtained.

In the bicharacteristic attention residual pyramid model, a bicharacteristic fusion mechanism is added based on replication connection. The feature results obtained from the n−1 layer through RASPP are fused with the results obtained from the n layer via sampling as the next processing input. In the model, average pooling and maximum pooling are used to extract features. Average pooling can retain more background feature information from the image, and maximum pooling can better extract texture information. By fusing bicharacteristic features from background and foreground textures, the segmentation performance of the model can be significantly enhanced. The bicharacteristic fusion module is shown in Fig. 6.Fig. 6 Bicharacteristic fusion module.

Fig. 6

In Fig. 6, after average pooling and maximum pooling, each pixel of the image needs to be added or multiplied; hence, the image input into the pooling operation must ensure unification of the image size and number of channels. The number of channels obtained by the residual pyramid (RASPP) operation of the n−1 layer feature map is five times higher than that of the same layer, while the number of channels obtained by the n layer is two times higher than that of the same layer after upconversion, and then, the number of channels is changed by a 1 × 1 convolution. After average pooling, pixels are added one by one to enhance the background information in the brain image, which is conducive to segmentation. After maximum pooling, results are multiplied pixel by pixel to enhance texture features. After convolution and pooling operation, the feature image is reduced to half of the original, and the image size is recovered by upconversion. The final result size is the same as the image size obtained via the original U-Net network, but the number of channels is doubled. This method can further extract background and foreground images, increase contrast, and improve the performance of the network. Specific steps are as follows:

Initially, the feature graph with C channels, W width, and H height, obtained by a convolution operation under n−1 layer sampling is defined by RW×H×C.

n−1 layer operation: the resultant feature map Rn−1W×H×5C is obtained through the RASPP module.

Step 1:(5) output(n−1)W×H×2C=conv(Ffeature),Ffeature∈Rn−1W×H×5C

where conv(x) indicates the convolution operation and output(n−1)1 indicates the first output of the n−1 layer operation.

Step 2:(6) output(n−1)2W2×H2×2C=AVG(output(n−1)1W×H×2C)output(n−1)3W2×H2×2C=MAX(output(n−1)1W×H×2C)

where AVG(x) indicates the average pooling and MAX(x) indicates the maximum pooling.

n layer operation: define the image input into the n layer for operation as T12(W×H)×2C.

Step 1:(7) outputn1W×H×2C=upconv(Mfeature),Mfeature∈T12(W×H)×2C

where upconv(x) is the upsampling operation and outputn1 represents the first output of the n layer operation.

Step 2:(8) outputn2W2×H2×2C=AVG(outputn1W×H×2C)outputn3W2×H2×2C=MAX(utputn1W×H×2C)

where AVG(x) indicates the average pooling and MAX(x) indicates the maximum pooling.

After the operation of layer n and layer n−1 above.

Step 1:(9) output1=∑k=12C(output(n−1)2kW2×H2(x,y)+outputn2kW2×H2(x,y))output2=∑k=12C(output(n−1)3kW2×H2(x,y)×outputn3kW2×H2(x,y))

where outputk represents the k channel feature of the feature graph and outputk(x,y) represents the eigenvalue corresponding to the coordinates on the k channel.

Step 2:(10) output3W2×H2×4C=output1W2×H2×2C+output2W2×H2×2C

where when input images output1W2×H2×2C and output2W2×H2×2C are superimposed on the channel, the number of obtained channels is twice that of the input image.

Step 3:(11) output4W×H×4C=upconv(output3W2×H2×4C)

As feature fusion step needs to pool image and reduce image size, it is necessary to conduct the upper convolution operation when final image is output. The bilinear interpolation method is used to increase the image size relative to the original.

3.2.2 Bicharacteristic attention mechanism

Channel attention mechanisms allow the model to concentrate on useful feature information, use global average pooling for feature compression, make the feature map scalar, and enable the model to learn feature information from the image. Maximum pooling can enhance texture information and allow the model to learn the final discriminant characteristics. Unlike channel attention, which focuses on feature information, the spatial attention mechanism [31] focuses the attention of the model on location points with useful feature information. Based on bicharacteristic feature fusion, channel attention and spatial attention are combined to obtain more efficient feature discrimination and location. A bicharacteristic attention mechanism model combining spatial attention and channel attention is shown in Fig. 7.Fig. 7 Dual attention mechanism module.

Fig. 7

In the picture, the spatial attention module uses 3 × 3 and 1 × 1 convolution to extract spatial features, in which padding = 2 is used to fill in the 3 × 3 convolution to ensure that the image size does not change. A 1 × 1 convolution is used to compress the feature map in the channel dimension, and finally the weight is normalized by the sigmoid function and multiplied by the original image to obtain the result with spatial attention. In the channel attention module, FC represents the full connection layer, where hyperparameter R is set to 16, which is used to transform the feature image into a scalar. This part uses two full connection and activation functions. The ReLU function can optimize parameters, improve calculation rates, reduce interdependence between parameters, and discard information during propagation. Afterward, the sigmoid function normalizes the weight between 0 and 1, scales each channel, and multiplies the final channel weight feature parameter with the original image to obtain a feature result with channel attention. Finally, feature results with spatial attention and channel attention are added channel by channel and point by point to obtain the final dual attention module. The specific steps are as follows:

The input image is the last output value in 3.2.1 by output4W×H×4C, which is redefined as imageW×H×4C.

Spatial attention:

Step 1:(12) output11W×H×8C=conv3(imageW×H×4C)

where conv3 represents convolution whose convolution kernel size is 3. To maintain the image size without alteration, a padding value of 2 is used.

Step 2:(13) output12W×H×1=Sigmoid(conv1(output11W×H×8C))

where conv1 indicates the convolution, whose kernel size is 1, which is used to compress the number of feature graph channels to 1; and Sigmoid(x) indicates the activation function sigmoid, which is used to normalize the spatial weight value between 0 and 1 and improve subsequent processing speed.

Step 3:(14) output13iW×H=output12W×H×1×imageiW×H,i∈4C

Channel attention:

Step 1:(15) output211×1×4C=ReLU(FC(imageW×H×4C))

where FC(x) represents fully connected operation, turning the feature graph into a scalar, and ReLU(x) represents the ReLU activation function, through which the part of the scalar value less than or equal to 0 can be set to 0, then the background feature is 0. If the part greater than 0 is still the original value, then the foreground remains unchanged, and the weight of the effective feature can be expanded.

Step 2:(16) output221×1×4C=Sigmoid(FC(output211×1×4C))

where Sigmoid(x) indicates the sigmoid activation function, the calculated result is normalized between 0 and 1.

Step 3:(17) output23iW×H=output22i1×1×4C×imageiW×H,i∈4C

where i represents one channel of the image, the scalar result calculated in formula 19 is multiplied by each channel of the original input to form the final feature image with channel attention.

Final output:(18) outputiW×H(x,y)=output13iW×H(x,y)+output23iW×H(x,y),i∈4C

where outputiW×H(x,y) represents the pixel value of feature point (x,y), whose coordinate of the feature map is in channel i. Adds the results obtained by space attention, channel attention on pixel by pixel to form the final feature map with dual attention.

3.3 Bicharacteristic attention residual pyramid

Residual pyramid and bicharacteristic attention are added to U-Net, and the overall structure of the model is shown in Fig. 8.Fig. 8 Bicharacteristic attention residual pyramid model.

Fig. 8

In the bicharacteristic attention residual pyramid model, a residual pyramid (RASPP) module is added to each layer, which can extract features on different receptor fields, enhance feature extraction ability, and solve the lack of connection between distant features. RASPP can extract detailed features at different scales while retaining global information, providing prospect information for subsequent brain segmentation. The process of downsampling is still retained in U-Net, which is used to extract feature map information of different sizes to enhance the feature recognition ability of the model. In the same layer image channel connection part of the original U-Net, a bicharacteristic attention mechanism is added. During upsampling, it is used to increase the image size, which then leads to the loss of some features. The original image can retain more soft tissue information about intracranial tumors by utilizing the upper and lower samples for feature fusion. Combined with channel attention and spatial attention, feature information and location information are enhanced, which can improve the segmentation accuracy of the model. The overall framework structure of the model is shown in Table 1.Table 1 Bicharacteristic attention residual pyramid model structure.

Table 1Number	Module	Layer	Image size	
1	Input	Input layer	256 × 256 × 3	
2	Coding 1	Convolution layer (3 × 3 conv kernel, activation function ReLU)
RASPP layer (batch normalized BN, 3 × 3 conv)
Maximum pooling (step size 2)
Convolution layer (1 × 1 conv kernel)	256 × 256 × 32
256 × 256 × 160
128 × 128 × 160
128 × 128 × 32	
3	Coding 2	Convolution layer (3 × 3 conv kernel, activation function ReLU)
RASPP layer (batch normalized BN, 3 × 3 conv)
Maximum pooling (step size 2)
Convolution layer (1 × 1 conv kernel)	128 × 128 × 64
128 × 128 × 320
64 × 64 × 320
64 × 64 × 64	
4	Coding 3	Convolution layer (3 × 3 conv kernel, activation function ReLU)
RASPP layer (batch normalized BN, 3 × 3 conv)
Maximum pooling (step size 2)
Convolution layer (1 × 1 conv kernel)	64 × 64 × 128
64 × 64 × 640
32 × 32 × 640
32 × 32 × 128	
5	Coding 4	Convolution layer (3 × 3 conv kernel, activation function ReLU)
RASPP layer (Section 3.1 steps)
Maximum pooling (step size 2)
Convolution layer (1 × 1 conv kernel)	32 × 32 × 256
32 × 32 × 1280
16 × 16 × 1280
16 × 16 × 256	
6	Coding 5	Convolution layer (3 × 3 conv kernel, activation function ReLU)
RASPP layer (batch normalized BN, 3 × 3 conv)
Convolution layer (1 × 1 conv kernel)	16 × 16 × 512
16 × 16 × 2560
16 × 16 × 512	
7	Decoding 1	Upsampling (2 × 2)	32 × 32 × 256	
concat	Two-feature fusion (Step in Section 3.2.1)	32 × 32 × 512
32 × 32 × 512
32 × 32 × 256	
Bicharacteristic attention mechanism (steps in Section 3.2.2)
Convolution layer (1 × 1 conv kernel)	
Convolution layer (3 × 3 conv kernel, activation function ReLU)	32 × 32 × 256	
8	Decoding 2	Upsampling (2 × 2)	64 × 64 × 128	
concat	Two-feature fusion (Step in Section 3.2.1)	64 × 64 × 256
64 × 64 × 256
64 × 64 × 128	
Bicharacteristic attention Mechanism (steps in Section 3.2.2)
Convolution layer (1 × 1 conv kernel)	
Convolution layer (3 × 3 conv kernel, activation function ReLU)	64 × 64 × 128	
9	Decoding 3	Upsampling (2 × 2)	128 × 128 × 64	
concat	Two-feature fusion (Step in Section 3.2.1)	128 × 128 × 128
128 × 128 × 128
128 × 128 × 64	
Bicharacteristic attention Mechanism (steps in Section 3.2.2)
Convolution layer (1 × 1 conv kernel)	
Convolution layer (3 × 3 convolution kernel, activation function ReLU)	128 × 128 × 64	
10	Decoding 4	Upsampling (2 × 2)	256 × 256 × 32	
concat	Two-feature fusion (Step in Section 3.2.1)	256 × 256 × 64
256 × 256 × 64
256 × 256 × 32	
Bicharacteristic attention Mechanism (steps in Section 3.2.2)
Convolution layer (1 × 1 conv kernel)	
Convolution layer (3 × 3 conv kernel, activation function ReLU)	256 × 256 × 32	
11	Output	Output layer (1 × 1 conv, activation function Softmax)	256 × 256 × 2	

Specific steps of the algorithm are as follows:

Input: Brain MRI image and mask image.① Convolution and residual pyramid modules were used to extract the soft tissue features of brain tumors.

② The feature dimension is reduced by the maximum pooling operation.

③ A bicharacteristic attention mechanism is used to calibrate obtained features more accurately.

④ Upsampling operation is used to increase the scale of the feature image.

⑤ Finally, a 1 × 1 convolution is used to change the number of output image channels to 2.

⑥ To calculate the probability of tumor soft tissue and nontumor soft tissue for each pixel, you can use the Softmax function.

⑦ You can set the threshold to 0.5. If the probability of tumor soft tissue exceeding this threshold exceeding this threshold exceeds this threshold, it is classified as a tumor; otherwise, it is considered nontumor tissue.

Output: Tissue segmentation of a craniocerebral tumor.

4 Experiments

4.1 Datasets and preprocessing

4.1.1 Datasets

The bicharacteristic attention residual pyramid model proposed in this study is tested on Kaggle_3M and BraTs2021 datasets and compared with segmentation methods in recent years.

The Kaggle_3M dataset, obtained from the Kaggle Brain Tumor Segmentation Competition, contains a grouped folder of brain tumor sequences of 110 patients; each sequence contains more than 20 unequal slice images, each image size is 256 × 256, and the sequence images are flair mode images of brain tumors.

The BraTS2021 dataset, obtained from the brain tumor segmentation challenge (BraTS Challenge), contains 8160 MRI images from 2040 patients. Each patient folder contains T1, T1cd, T2, flair four modal sequence files and a segmentation truth seg file, and each modal data size is 240 × 240 × 155, shared segmentation label.

4.1.2 Preprocessing

In the Kaggle_3M dataset, the sequence folder of each patient contains raw data and mask data; the two types of data are separated. Some images do not contain brain tumor lesion areas. The pixels of the mask image are all 0 and should be cleared. In this case, the number of original images may not be equal to the number of mask images. According to the image name, the corresponding original images are cleared, and the final obtained image is the image with the lesion area and corresponding mask image, which can be input into the neural network for model training.

In the BraTS2021 dataset, each patient contains four types of sequence files, including flair mode files containing peritumoral edema regions, focal imaging, and enhanced tumor sequence images. Owing to the limited performance of equipment and the fact that the model proposed in this study is applicable to 2D image segmentation, only the images with superior flair as well as higher contrast and resolution are used to better train and test the model. In processing, the.nii images in the flair mode folder in each patient's sequence file were traversed, the dichotomous segmentation method [32] was used to obtain a cross-sectional section image of each case, the section image without lesion was removed, and the corresponding image in the seg folder was obtained as the mask image of the case.

4.2 Experimental parameter settings

The experiment was conducted on a 64-bit Windows 10 operating system with an AMD Ryzen 7 4800H processor version with Radeon Graphics 2.90 GHz and an NVIDIA GeForce GTX 1650 GPU. The deep learning framework is MXNet 1.7.0, and CUDA 11.0 is configured. Two datasets were separately divided into a training set, a verification set, and a test set, accounting for a ratio of 6:2:2. The training step epoch was set to 100, and the batch processing size was set to 8. Images in the training set were randomly selected for training. In the model, the loss function employs cross-entropy and the Adam optimizer is used for optimization.

4.3 Evaluation index

To verify the effectiveness of the improved method in MRI brain tumor tissue segmentation, four quantitative evaluation indexes were adopted, including the accuracy (ACC) rate, Jaccard similarity coefficient, Dice similarity coefficient, and false negative rate (FNR).

4.3.1 Accuracy

Accuracy is used to evaluate the ratio of correct results to the total predicted results. A higher accuracy leads to better prediction results for the model, but this single indicator is usually not used as an evaluation standard for the model. Accuracy (ACC) is calculated as follows:(19) ACC=TP+TNTP+TN+FP+FN

4.3.2 Jaccard similarity coefficient

The Jaccard similarity coefficient is used to compare the similarity and difference between the pixel set of a brain tumor tissue image after network segmentation and the pixel set of an expertly calibrated tissue image. A larger Jaccard coefficient value leads to a higher similarity between segmentation results and the gold standard and a better model segmentation performance. The Jaccard similarity coefficient is calculated as follows:(20) Jaccard=TPTP+FN+FP

4.3.3 Dice similarity coefficient

The Dice similarity coefficient is often used in the field of image segmentation to measure the similarity between pixels of segmentation and the gold standard image pixel set. Compared with IOU (intersection ratio), the Dice coefficient can obtain a higher index score. This coefficient is calculated as follows:(21) Dice=2×TP(TP+FN)+(TP+FP)

4.3.4 False negative rate

In image segmentation, the FNR is most commonly used to indicate the proportion of different objects divided into agreed-upon parts. A smaller FNR value leads to better segmentation performance for the model. The calculation formula is as follows:(22) FNR=FNTP+FN+FP

In all the abovementioned calculation formulas, TP represents the true positive rate; pixels of segmentation results compared with the gold standard should be divided into tumor soft tissue by the model. TN represents the true negative rate; the gold standard of comparison should be divided into the background, which is actually the pixel rate of the background. FP represents the false positive rate; the gold standard should be the background, but the pixel rate is divided into tumor soft tissue by the model. FN represents the FNR, the pixel rate that should be tumor soft tissue in the gold standard but is classified by the model as background.

4.4 Analysis of experimental results

To verify the superiority of the proposed model in MRI cranial tumor image segmentation, multiple sets of comparison experiments were designed to compare it with the most commonly used image segmentation model U-Net and three common variant models on Kaggle_3M and BraTS2021 datasets.1) The U-Net coding–decoding structure model uses a U-shaped structure to add concatenation connections in the same layer encoding and decoding parts to solve the problem of feature loss caused by sampling.

2) Attention U-Net integrates attention gate into the U-Net model and adds an attention module at the original skip connection, which can suppress irrelevant areas of the input image and highlight local significant features.

3) SeUNet adds a channel attention mechanism to U-Net. The channel attention module comprises two components: squeeze and excitation. The purpose of the module is to extract features from each convolutional kernel channel and establish comprehensive connections between channels.

4) ResUNet adds residual structure to U-Net before each part of the original convolution used to output the feature image. This change can solve the problem of feature loss caused by convolution.

The results obtained on the Kaggle_3M dataset are shown in Table 2. It is evident that the accuracy, Jaccard similarity coefficient, Dice similarity coefficient, and FNR of the model proposed in this study are 0.9193, 0.8634, 0.9341, and 0.0070, respectively. Compared with original U-Net and common U-Net variants, the four evaluation indexes are all improved, indicating superiority of proposed method.Table 2 Comparison results of different basic network models on the Kaggle_3M dataset.

Table 2Method	ACC	Jaccard	Dice	FNR	
U-Net	0.7939	0.7531	0.8677	0.0101	
Attention U-Net	0.8995	0.8403	0.9251	0.0072	
SeUNet	0.8404	0.8056	0.9175	0.0088	
ResUNet	0.8875	0.8252	0.9223	0.0080	
Our method	0.9193	0.8634	0.9341	0.0070	

The results obtained on the BraTS2021 dataset are shown in Table 3. The accuracy, Jaccard similarity coefficient, Dice similarity coefficient, and FNR of the model proposed in this study are as follows: 0.9195, 0.8994, 0.8952, and 0.0283, respectively. These values are all improved in the four evaluation indexes, suggesting the effectiveness of the proposed method in the segmentation of brain tumors.Table 3 Comparison results of different basic network models on the BraTS2021 dataset.

Table 3Method	ACC	Jaccard	Dice	FNR	
U-Net	0.7994	0.7759	0.7923	0.1730	
Attention U-Net	0.8765	0.8069	0.8800	0.1678	
SeUNet	0.8994	0.8351	0.8782	0.1412	
ResUNet	0.9163	0.8770	0.8921	0.1016	
Our method	0.9195	0.8994	0.8952	0.0283	

The model proposed in this study, the residual pyramid module, and the bicharacteristic attention module are added to U-Net. To verify the influence of different modules on the model, ablation experiments were set up. Residual, atrous spatial pyramid, residual pyramid, and bicharacteristic attention modules were integrated into the U-Net infrastructure. Experiments were conducted on the Kaggle_3M and BraTS2021 datasets. Obtained results are shown in Table 4, Table 5.Table 4 Comparison results of ablation experiments on the Kaggle_3M dataset.

Table 4Method	ACC	Jaccard	Dice	FNR	
Basic U-Net	0.9286	0.8427	0.8874	0.0148	
Add residual module	0.9291	0.8610	0.8802	0.0233	
Add atrous spatial pyramid module	0.9393	0.8735	0.8858	0.0138	
Add residual pyramid	0.9394	0.8751	0.8894	0.0081	
Added bicharacteristic attention mechanism	0.9394	0.8769	0.8883	0.0077	
Our method	0.9395	0.8804	0.8939	0.0060	

Table 5 Comparison results of ablation experiments in the BraTS2021 dataset.

Table 5Method	ACC	Jaccard	Dice	FNR	
Basic U-Net	0.8916	0.7137	0.8369	0.1576	
Add residual module	0.8933	0.7517	0.8441	0.1385	
Add atrous spatial pyramid module	0.9148	0.8094	0.8109	0.2291	
Add residual pyramid	0.9250	0.8122	0.9014	0.1096	
Added bicharacteristic attention mechanism	0.9353	0.8205	0.9163	0.0919	
Our method	0.9371	0.8849	0.9157	0.0789	

In the Kaggle_3M dataset, when comparing it with the original U-Net network, the addition of each module results in an improvement in segmentation accuracy to some extent. However, introducing a separate residual module leads to a decrease in the Dice coefficient and FNR. The residual pyramid module formed after combining the void space pyramid module can improve all evaluation indicators of the model. Although the dual attention mechanism does not improve the accuracy of model segmentation, it improves the Jaccard similarity coefficient and FNR. The network proposed in this study is formed by synthesizing various modules. The experiment proves that all performance evaluation indexes are improved, verifying that the model has higher edge-segmentation accuracy in tumor segmentation tasks. An image in the Kaggle_3M dataset was randomly selected for ablation module segmentation, and the obtained results are shown in Fig. 9a-h.Fig. 9 Results of ablation segmentation in the Kaggle_3M dataset.

Fig. 9

Fig. 9a shows the original tumor image, and Fig. 9b shows the superimposed image of the expert segmentation gold standard and the original image. It can be seen that the segmented tumor part can cover the tumor detected in the original image. Fig. 9c is the superimposed image of the original U-Net segmentation image and the original image. Fig. 9d displays the segmented superimposed image after incorporating the residual module. Fig. 9e is the segmented superimposed image after adding the hollow space pyramid. Fig. 9f displays the segmented superposition image after incorporating the residual pyramid. Fig. 9g shows the segmented superposition image after adding the dual attention mechanism. Fig. 9h shows the segmented superposition image after the segmentation method in this study. From the image segmentation results, it is evident that there is a gap between tumor edge segmentation and expert segmentation in Fig. 9c–e, and the edges tend to be covered in Fig. 9f–h. Thus, it is evident that the module added to this method has certain effectiveness, which can segment tumor shapes more accurately and have more accurate edge positioning.

On the BraTS2021 dataset, compared with the original U-Net network model, the accuracy of adding any module is improved. After adding the void space pyramid module, the Dice similarity coefficient and FNR are relatively reduced, indicating that the segmentation effect of the model is decreased. However, after adding the residual pyramid module, more detailed features are obtained. All performance indicators increased. The Dice similarity coefficient of the network model proposed in this study decreased slightly compared with the addition of dual attention mechanism, but other performance indicators were improved. Therefore, it can be concluded that the model in this study has advantages over the single module in tumor segmentation tasks. An image in the BraTS2021 dataset was randomly selected, and the image extracted after slice segmentation was used for ablation experiment segmentation. Obtained results are shown in Fig. 10a-h.Fig. 10 Results of ablation experiment segmentation in the BraTS2021 dataset.

Fig. 10

Fig. 10a displays the original brain image, Fig. 10b displays the image after segmentation of tumor morphology by superimposing experts, and Fig. 10c displays the superimposed image obtained by segmentation using a basic U-Net network. It is clearly evident that edge segmentation is not accurate. Fig. 10d displays the segmented image after adding a residual module, and tumor edge segmentation has been improved to a certain extent. However, some edge segments crossed the boundary (the other color areas of the edge part in the figure). Fig. 10e is the segmentation map after adding the void space pyramid module, and the edge and shape gradually fit together to the expert segmentation image. Fig. 10f is the segmentation results obtained by adding the residual pyramid module, the dual attention module and the method proposed in this paper, respectively, and the tumor edge segmentation becomes more and more accurate. The advantages of adding various modules to the proposed method can be obtained.

An epoch of 100 was used to conduct comparative experiments with existing advanced methods [[22], [23], [24], [25]] on Kaggle_3M and BraTS2021 datasets. Obtained results are shown in Table 6, Table 7. In the Kaggle_3M dataset, the proposed method is superior to existing algorithms in accuracy, Jaccard coefficient, Dice coefficient, and FNR. On the BraTS2021 dataset, the accuracy, Jaccard coefficient, and FNR of the proposed method are higher than those of the comparison method, but the Dice similarity coefficient is lower than the algorithm [25]. Based on the evaluation index of algorithm 25, the Dice similarity coefficient appears excessively high, which could potentially indicate overfitting in the experiment. Through comparison, it can be shown that the method proposed in this study has certain superiority in the comparison method.Table 6 Experimental results of different algorithms in the Kaggle_3M dataset.

Table 6Methods	ACC	Jaccard	Dice	FNR	
Method [22]	0.9192	0.8689	0.8747	0.313	
Method [23]	0.9393	0.8698	0.8893	0.112	
Method [24]	09293	0.8723	0.8797	0.0024	
Method [25]	09294	0.8765	0.8518	0.0259	
Our method	0.9395	0.8812	0.8958	0.007	

Table 7 Experimental results of different algorithms in the BraTS2021 dataset.

Table 7Methods	ACC	Jaccard	Dice	FNR	
Method [22]	0.9148	0.8094	0.8109	0.2291	
Method [23]	0.9254	0.8219	0.8841	0.0269	
Method [24]	0.9173	0.8989	0.8941	0.1137	
Method [25]	0.926	0.8538	0.9305	0.0929	
Our method	0.9375	0.9072	0.8981	0.0087	

4.5 Results of segmentation

Among brain tumor images of each mode, the flair image has the strongest contrast; hence, it is used as the original contrast image for the experiment, and three pathological images are randomly selected in the processed Kaggle_3M and BraTS2021 datasets for segmentation. Obtained results by the model and the comparison algorithm in this study are shown in Fig. 11(a–u) and 12(a-u).Fig. 11 Image segmentation effect of the Kaggle_3M dataset.

Fig. 11

Through comparison of experimental results, it is evident that the segmentation boundary effect of the proposed method is superior to the comparison algorithm in cranial tumor segmentation, and the shape and position of the tumor can be obtained by comparing existing methods, no matter whether on the Kaggle_3M dataset or the BraTs2021 dataset. Better segmentation results can be obtained for tumors with close boundaries (as shown in the box lines in Fig. 11(a–u) and 12(a-u)), and especially for tumors with complex boundaries, more accurate boundary segmentation results can be obtained (as shown in Fig. 12d, g, j, m, p, s).Fig. 12 Image segmentation effect of the BraTS2021 dataset.

Fig. 12

5 Conclusions

The segmentation of craniocerebral tumors holds clinical significance for brain tumor surgical treatment. In the current study, a network model combining the residual pyramid and bicharacteristic attention mechanisms is proposed for segmentation of craniocerebral tumors. Experiments show that the coding part combined with the residual pyramid module can effectively expand the receptive field while retaining more image features and obtaining spatial correlation information. The bicharacteristic attention mechanism is used in the decoding part of the skip connection, which helps recover all features of the image and reduces the probability of detail feature loss. By using this method, the accuracy of BraTS can be improved, the exact location and boundary information of tumors can be obtained, and it can improve the success rate of craniocerebral surgery. However, the actual clinical data are complex and the cases are changeable; hence, the method proposed in this study has not been really used in actual clinical trials, and the segmentation accuracy of specific datasets may be decreased. The next step will be to improve the model and further improve its generalization ability.

Ethical statement

All experiments performed in the study involving human participants were in accordance with the ethical standards of the institutional research committee and the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. This study was performed using two public human brain subject datasets. Ethical approval was confirmed by the license attached with their open access data.

Data availability statement

The data that support the findings of this study are openly available at the following URLs: https://kaggle.com/mateuszbuda/lgg-mri-segmentation and https://www.kaggle.com/data sets/dschettler8845/brats-2021-task1.

CRediT authorship contribution statement

Yiliu Hang: Writing – review & editing, Writing – original draft, Validation, Software, Resources, Methodology, Investigation, Conceptualization. Qiong Zhang: Validation, Methodology, Data curation. Xiujing Li: Visualization, Software, Data curation. Jianlin Qiu: Resources, Project administration, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This study was funded by the 10.13039/501100004608 Natural Science Foundation of Jiangsu Province in 2023 under Grant No. BK20231337 , the program of the Natural Science Research of Jiangsu Province Higher Education Institutions in 2023 under Grant No.23KJD520011 , the General Program of Nantong Science and Technology Bureau in 2020 under Grant Nos.JC22022108 and MSZ202072 , and the Program of the Nantong Institute of Technology Young and Middle-aged Backbone Teacher in 2022 under Grant Nos. ZQNGGJS202237 and ZQNGGJS202234 . We would like to thank everyone who has contributed to this article. We would also like to thank the anonymous reviewers and editors for their helpful suggestions and comments.
==== Refs
References

1 Li Jiaqi. Application and accuracy of MRI in diagnosis of intracranial tumors [J]. Imaging Research and Medical Application,202,6 (09):109-111.
2 Gao J.H. Lei H. Chen Q. Magnetic resonance imageing: progresses and perspective Scientia Sinica Vitae 50 11 2020 1285 1295
3 Yunfei Shu Chen Wei Zhang Yonghong To study the effects of acute isovolemic hemodilution on cerebral oxygen supply and oxygen consumption during craniocerebral tumor surgery Journal of Modern Medicine and Health Research 7 8 2019 36 39
4 Hengyi Tian Xiao Hongbing Ji Yarong Automatic segmentation algorithm of brain tumor based on three-dimensional UNet and mixed focus loss function Chinese Journal of Medical Physics 40 9 2019 1114 1120
5 Liu Z.H. Tong L. Chen L. Deep learning based brain tumor segmentation: a survey Complex & Intelligent Systems 9 1 2023 1001 1026
6 Fukushima K. Neocognitron: a self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position Biol. Cybern. 36 4 1980 193 202 7370364
7 Li Y. Zhao B. Wen L. Multi-purposed diagnostic system for ovarian endometrioma using CNN and transformer networks in ultrasound Biomed. Signal Process Control 2024 91105923
8 Zhu Z He X Qi G Brain tumor segmentation based on the fusion of deep semantics and edge information in multimodal MRI[J] Inf. Fusion 91 2023 376 387
9 Liu Z. Lin Y. Cao Y. Hu H. Wei Y. Zhang Z. Lin S. Guo B. Swin transformer: hierarchical vision transformer using shifted windows Proceedings of the IEEE/CVF International Conference on Computer Vision 2021 IEEE Piscataway, NJ 10012 10022
10 Xu Y , Yu K , Qi G ,et al. Brain tumour segmentation framework with deep nuanced reasoning and Swin‐T[J].IET Image Process. 18:1550 [2024-06-27].DOI:10.1049/ipr2.13048.
11 Ronneberger O. Fischer P. Brox T. U-Net: convolutional networks for biomedical image segmentation[C]// Proceedings of Medical Image Computing and Computer-Assisted Intervention 2015 Springer Berlin 234 241
12 Qurri A.A. Almekkawy M. Improved UNet with attention for medical image segmentation Sensors 23 20 2023
13 Zhang Weizhi Qian Yu Su Jinshan From U-Net to Transformer: a review of the application of depth models in medical image segmentation [J/OL] Computer application 1–23 2024 04-26] http://kns.cnki.net/kcms/detail/51.1307.tp.20231026.1648.002.html
14 Nisa Q.S. Ismail R.A. Dual U-net with resnet encoder for segmentation of medical images Int. J. Adv. Comput. Sci. Appl. 13 12 2022
15 Xiao X. Chen L. Luo Z. Weighted Res-UNet for high-quality retinal vessel segmentation[C] Proceeding of 2018 9th International Conference on Information Technology in Medicine and Education(ITME) 2018 IEEE Piscataway 327 331
16 Phani N.B. Narahari P.S. Automated brain tumor detection and segmentation using modified UNet and ResNet model Soft Comput. 27 13 2023 9179 9189
17 Yinyin W. Aijun Z. Xiangyang Z. RAG‐Net: ResNet‐50 attention gate network for accurate iris segmentation IET Image Process. 16 11 2022 3057 3066
18 Lou A. Guan S. Loew M. DC-UNet: rethinking the U-Net architecture with dual channel efficient CNN for medical image segmentation [C]//Medical Imaging 2021: image Processing Bellingham: SPIE 11596 2021 758 768
19 Karaali A. Dahyot R. Sexton D.J. DR-vnet: retinal vessel segmentation via dense residual UNet [C]// Proceedings of Internatinal Conference on Pattern Recognition and Artificial Intelligence 2022 Springer Berlin 198 210
20 Zhu Z. Sun M. Qi G. Sparse Dynamic Volume TransUNet with multi-level edge fusion for brain tumor segmentation Comput. Biol. Med. 172 2024 10.1016/j.compbiomed.2024.108284
21 Zhu Z. Wang Z. Qi G. Brain tumor segmentation in MRI with multi-modality spatial information enhancement and boundary shape correction Pattern Recogn. 153 2024 10.1016/j.patcog.2024.110553
22 Huang Z. Zhang X. Song Y. FECC-net: a novel feature enhancement and context capture network based on brain MRI images for lesion segmentation Brain Sci. 12 6 2022 765 35741650
23 Zhu Jiajun Zhang Rui Zhang Haifei An MRI brain tumor segmentation method based on improved U-Net Math. Biosci. Eng.: MBE 21 Issue 1 2024 778 791 38303443
24 Aboussaleh I. Riffi J. Mahraz A.M. Inception-UDet: an improved U-net architecture for brain tumor segmentation Annals of Data Science 11 3 2024 831 853 10.1007/s40745-023-00480-6
25 Sun J. Hu M. Wu X. MVSI-Net: multi-view attention and multi-scale feature interaction for brain tumor segmentation Biomed. Signal Process Control 95 2024 10.1016/j.bspc.2024.106484
26 Zhang Weizhi, Yu Qian, Su Jinshan, et al. From U-Net to Transformer: a review of the application of depth models in medical image segmentation [J/OL]. Comput. Appl.,1–23 [2024-04-25].
27 Mayuri G. Ashish M. A systematic review of deep learning based image segmentation to detect polyp Artif. Intell. Rev. 57 1 2024
28 Shehab H.L. Fahmy M.O. Gasser M.S. An efficient brain tumor image segmentation based on deep residual networks (ResNets) Journal of King Saud University-Engineering Sciences 33 6 2020
29 Wen Y. Dongming Z. Teng F. Image segmentation of skin lesions based on dense atrous spatial pyramid pooling and attention mechanism J. Biomed. Eng. 39 6 2022 1108 1116
30 VatanpourM HaddadniaJ. Brain tumour segmentation of MR images based on custom attention mechanism with transfer learning IET Image Process. 18 4 2023 886 896
31 Luo J. Wang Q. Zou R. A heart image segmentation method based on position attention mechanism and inverted pyramid Sensors 23 23 2023
32 Jingming X.I.A. Tan Ling Ying Liang Application of LMD-UNet network in multimodal MRI image segmentation of brain tumors Chin. J. Biomed. Eng. 42 4 2019 431 441
