SEF-UNet: advancing abdominal multi-organ segmentation with SEFormer and depthwise cascaded upsampling

Squeeze-and-excitation networks

Squeeze-and-Excitation (SE) block (Hu, Shen & Sun, 2018) is an architectural unit that focuses on channel relationships. This module selectively enhances the weight of informative features and suppresses less useful features by learning global information.

The SE block consists of two main stages: Squeeze and Excitation. For an input X, a transformation F_tr:X → U is performed, where X ∈ R^{H′×W′×C′} and U ∈ R^H×W×C. This transformation can be operations such as convolution or convolution sets. The SE module is introduced into this transformation process to recalibrate the features.

Firstly, through the squeeze operation, features are aggregated into a H × W feature map, consolidating global information to generate channel descriptors. Following is the excitation operation, where the model activates specific channels by learning sample-specific activations through a channel-dependent gating mechanism. Subsequently, the obtained activation data is used to weight the feature map U. It is noteworthy that, since the size and number of channels of the feature map remain unchanged, the output can be directly passed to subsequent layers without significant adjustments to the original model’s structure. The introduction of the SE module contributes to improved model performance, allowing for more effective capture and utilization of global contextual information.

The entire process can be represented by the following formula: (1)${\displaystyle SE\left(x\right)=\sigma \left(W_2\delta \left(W_1\text{AvgPool}\left(x\right)\right)\right)\cdot x}$

For input x, the average value of each channel is obtained through global average pooling to acquire global information for each channel. This helps reduce information within each channel, incorporating global contextual information into the description of each channel. The aforementioned part constitutes the Squeeze operation. The Excitation stage includes the first fully connected layer, ReLU activation function, the second fully connected layer, and the Sigmoid function. The first fully connected layer is a compressed fully connected layer, reducing the number of channels from C to $\frac{C}{r}$. The second fully connected layer is the excitatory fully connected layer, restoring the number of channels from $\frac{C}{r}$ to C. The goal of these two layers is to learn the weights for each channel, enhancing or diminishing the feature response of that channel.

By introducing the Squeeze and Excitation stages, SENet can adaptively learn the importance of each channel and capture key information more effectively through updated feature maps. This has led to significant performance improvements for SENet in various image tasks.

The SE block is computationally lightweight, providing noticeable improvements to the model with minimal additional computation. The introduction of the SE module helps the network capture crucial information more efficiently from input features, enhancing model performance. This module is typically integrated into various layers of CNNs to enhance the network’s representational capabilities.

MetaFormer

MetaFormer (Yu et al., 2022) is a versatile architecture that abstracts out parts of the Transformer except for the multi-head self-attention module. A MetaFormer block consists of two residual blocks, where the first residual block can be represented as: (2)${\displaystyle X^{\prime }=\text{TokenMixer}\left(\text{Norm}\left(X\right)\right)+X}$

For input X, it first passes through the Norm(⋅) layer, representing the normalization layer. Following that is TokenMixer(⋅), used for token information mixing, with various implementations such as Identity Mapping, Random Mixing, Separable Convolution, Attention, etc.

The second residual block primarily consists of a two-layer MLP and a non-linear activation layer, expressed as: (3)${\displaystyle Y=W_2\left(\sigma \left(W_1\left(\text{Norm}\left(X^{\prime }\right)\right)\right)\right)+X^{\prime }}$

Here, W₁ ∈ R^C×rC and W₂ ∈ R^rC×C are the learnable parameters for the two fully connected layers, and σ(⋅) is the non-linear activation function, typically using ReLU or its improved versions.

SEFormer block

The overall structure of the SEFormer block is illustrated in Fig. 3. We use the MetaFormer (Yu et al., 2022; Yu et al., 2023) architecture as the backbone and instantiate the token mixer with a combination of the Squeeze-and-Excitation module and depthwise separable convolution. The use of depthwise separable convolution retains good performance while significantly reducing the number of parameters and computations. Simultaneously, the SE module captures crucial information in the input features, enhancing the network’s expressive capabilities.

Figure 3 depicts the complete structure of the SEFormer block. The expression for the SEFormer block is as follows: (4)${\displaystyle X^{\prime }=SESepConv\left(Norm\left(X\right)\right)+X}$ (5)${\displaystyle SEFormerBlcok\left(X\right)=X^{\prime }+MLP\left(Norm\left(X^{\prime }\right)\right)}$

This module consists of two residual structures. First, the input X is normalized using LayerNorm (Norm(⋅)). Subsequently, after passing through the SESepConv(⋅) module, the output is added to the original input to obtain an intermediate value X′. This intermediate value is then normalized again through LayerNorm and processed through a Multilayer Perceptron (MLP) containing two fully connected layers. Finally, the output of the MLP is added to the unprocessed intermediate value X′ to obtain the final output of the module.

The SE Separable Convolution is the core unit of this module, as shown in the structure in Fig. 4. The entire structure mainly consists of pointwise convolution, depthwise convolution, and the branched Squeeze-and-Excitation structure.

The specific expression for this module is as follows: (6)${\displaystyle SEConv\left(X\right)=Con{v}_{pw2}\left(SE\left(Con{v}_{dw}\left(\sigma \left(Con{v}_{pw1}\left(X\right)\right)\right)\right)\right)}$

where Conv_dw(⋅) represents depthwise convolution, independently convolving each channel of the input. Conv_pw1 and Conv_pw2 denote pointwise convolution, merging the output of depthwise convolution through a 1 × 1 convolution.

For the input X, the calculation formula to obtain the output Y through depthwise convolution is as follows: (7)${\displaystyle Y_{i,j,c}=\sum _{m,n}{X}_{i\times s+m,j\times s+n,c}\times K_{m,n,c}}$

where i, j are the spatial coordinates of the output tensor, c is the channel index, s is the stride of depthwise convolution, m, n are the coordinates of the convolution kernel for depthwise convolution, X_i,j,c is the element of the input tensor, and K_m,n,c is the parameter of the depthwise convolution kernel.

The calculation formula for pointwise convolution is: (8)${\displaystyle Y_{i,j,c}=\sum _k{X}_{i,j,k}\times K_{1,1,k,c}}$

where K_1,1,k,c is the convolutional kernel parameter for pointwise convolution, and the other parameters are consistent with Eq. (7).

The Squeeze-and-Excitation structure mainly consists of two stages: Squeeze and Excitation. Assuming the input feature map X has dimensions H × W × C, where H and W represent the height and width of the input feature map respectively, and C represents the number of channels. During the squeezing stage, global average pooling is used to capture the global information of each channel of the input feature map X. (9)${\displaystyle z_c=\frac{1}{H\times W}\sum _{i=1}^H\sum _{j=1}^W{x}_{cij}}$

Here, z_c is the global average of the c-th channel, and x_cij is the value of input X at the c-th channel, where the row and column indices are i and j, respectively.

In the Excitation stage, two fully connected layers are introduced, and the Sigmoid activation function is used to learn the weights of each channel. These two fully connected layers are referred to as Squeeze and Excitation, respectively. The first fully connected layer is used to compress the number of channels, reducing the global average value z_c of each channel by a compression ratio r: (10)${\displaystyle s_c=\sigma \left(W_1\cdot z_c\right)}$

where W₁ is the weight matrix of the first fully connected layer, σ is the activation function, typically chosen as ReLU.

The second fully connected layer is used to restore the number of channels by exciting the original feature map with the learned weights s_c: (11)${\displaystyle f_c=\sigma \left(W_2\cdot s_c\right)}$

where W₂ is the weight matrix of the second fully connected layer, and σ is the Sigmoid activation function.

Finally, by applying the weights of each channel to the original feature map, the updated feature map is obtained: (12)${\displaystyle y_{cij}=f_{\text{c}}\cdot x_{cij}}$

where y_cij represents the value at position (c, i, j) in the updated feature map.

Therefore, the overall representation of the Squeeze and Excitation process can be expressed as: (13)${\displaystyle SE\left(X\right)=\sigma \left(W_2\sigma \left(W_1\left(AvgPool\left(X\right)\right)\right)\right)\ast X}$

dCUP as decoder

To further enhance model computational efficiency, reduce parameter count, and improve segmentation accuracy, we designed a depthwise cascaded upsampling (dCUP) structure, which utilizes a depthwise convolution layer. This structure is constructed using components shown in Fig. 5.

The dCUP architecture comprises upsampling layers, feature fusion layers, convolution layers, activation functions, etc. The upsampling layers aim to increase the resolution of the feature map, gradually improving resolution by stacking multiple dCUP blocks to generate an output of the same size as the input image. Feature fusion involves concatenating feature maps from the corresponding stages of the encoder with the current feature map of the decoder stage. The concatenated data undergoes a convolution layer to adjust the channel count to the target channel count, followed by batch normalization and ReLU operations to enhance network training stability. Finally, through a depthwise convolution layer, features are further fused to capture local details and texture information. The expression for the entire process is as follows: (14)${\displaystyle Y=Con{v}_{dw}\left(\sigma \left(Norm\left(Conv\left(cat\left(feature,Up\left(X\right)\right)\right)\right)\right)\right)}$

where Up(⋅) represents bilinear interpolation upsampling, which is concatenated with features from the encoder using the cat method. σ(⋅) is the activation function, and Norm(⋅) is the batch normalization operation. Since there are three sets of features from the encoder, only the first three dCUPs need to receive features and perform feature fusion operations, while subsequent dCUPs do not need to combine with features but proceed directly to the subsequent steps.

The decoder structure of the entire model consists of four dCUPs. Such upsampling blocks can simultaneously achieve upsampling, feature concatenation, and feature processing. The introduction of the depthwise convolutional layer significantly enhances the ability to restore details and semantic information of the original input image with only a small increase in the number of parameters and computational complexity.

The entire model extracts high-level semantic information from the input image through the encoder, retains low-level detail information through skip connection, and gradually combines these pieces of information in the decoder to generate the final segmentation result.

Dataset

Our dataset is Multi-Atlas Labeling Beyond the Cranial Vault (BCV) (https://www.synapse.org/#!Synapse:syn3193805/wiki/217789). The dataset has a total of 30 abdomen CT scans taken during portal venous contrast phase. The pixel size of the scans is 512 × 512 and the slices ranges from 85 to 198. Field of view size is approximately 280 × 280 × 280 mm³−500 × 500 × 650 mm³, and the voxel spatial resolution of ([0.54∼0.54] × [0.98∼0.98] × [2.5∼5.0]) mm³

The BCV dataset consists of a total of 13 categories, and we have selected 8 abdominal organs for the segmentation task. These 8 organs are the spleen (Sp), right kidney (RK), left kidney (LK), gallbladder (Ga), liver (Li), stomach (St), aorta (Ao), and pancreas (Pa). We selected these eight categories because these organs are of significant importance in clinical diagnosis and treatment. Additionally, this selection aligns with the multi-organ segmentation tasks in existing literature, facilitating a fair and effective comparison with existing methods.

In all 30 samples, complete annotations for the 8 classes are provided. We randomly chose 21 samples for training, and the remaining 9 samples were used for testing. The 21 training samples were divided into 2,698 slices horizontally along the axial direction. All experimental results were averaged over the 9 test samples.

Inspired by Xie et al. (2021), for the original CT dataset, we first truncated the Hounsfield Unit (HU) values to the range [−125, 275] and then normalized them to scale the values between [0, 1]. During the training process, we applied random flips, rotations, and other operations to the input slice data to increase the diversity of the training set. Subsequently, the image size was resized to 224 × 224 to obtain fixed-size inputs for further training.

Implementation details

The SEF-UNet model is implemented in Python 3.7 with PyTorch 1.13. It utilizes the SGD optimizer with a learning rate of 0.01, momentum set to 0.9, and weight decay of 1e−4. The batch size for all experiments is set to 24, and the input resolution is 224 × 224. Each model is trained for 150 epochs on 2,698 slices, with each epoch containing 113 iterations, resulting in a total of 16,950 iterations. The experiments were conducted on a single NVIDIA GeForce RTX 4060 GPU.

Other relevant parameters during training include: the encoder consists of four blocks, with the repetition count for each block, i.e., [L₁, L₂, L₃, L₄], set as [3, 3, 9, 3], and the corresponding channel numbers for each block (downsampling module is used for adjusting image resolution and channel numbers, while the block is mainly used for feature extraction with unchanged channel numbers) set as [64, 128, 320, 512].

The backbone network in the encoder is pretrained on ImageNet21k. The decoder stage uses four dCUPs with an upsampling factor of 2, and the output channel numbers for each upsampling block are [256, 128, 64, 16]. The first three dCUPs receive fused skip features from the encoder, while the last one does not receive fusion information and is mainly used for further feature processing.

The training data consists of processed 2D slices. During the prediction process, inference is performed slice by slice, and the predicted 2D slices are stacked together to construct the 3D prediction result.

The loss function used in the experiments is a combination of CrossEntropyLoss and DiceLoss. For a sample with N classes and true labels y₁, y₂, …, y_N, and model output probability distribution p₁, p₂, …, p_N, the loss function is defined as: (15)${\displaystyle L_{seg}=\left(L_{CE}+L_{Dice}\right)/2=\left(-\sum _{i=1}^N{y}_i\cdot log\left(p_i\right)+1-\frac{2\sum _{i=1}^N{y}_i\cdot p_i+ϵ}{\sum _{i=1}^N{y}_i+\sum _{i=1}^N{p}_i+ϵ}\right)/2}$

where ϵ is a small constant added for numerical stability.

CrossEntropyLoss effectively measures the matching degree between the model’s output probability distribution and the true labels. On the other hand, DiceLoss provides a smoother computation, contributing to training stability. However, DiceLoss may appear overly optimistic in certain situations as it does not consider subtle differences between true labels and predictions. Therefore, we choose to combine these two losses to comprehensively consider both the matching degree of the model’s output probability distribution with true labels and sensitivity to subtle differences, aiming for a more holistic model training approach.

Result

We conducted experiments comparing our model with several traditional medical image segmentation models and advanced segmentation models. The models include the classic UNet (Ronneberger, Fischer & Brox, 2015), UNet++ (Zhou et al., 2018) that enhances UNet with more skip connections to improve multi-scale feature modeling, UNet3+ (Huang et al., 2020) that introduces attention mechanism and dense connections, as well as two important and classic models in the medical image segmentation field that incorporate Transformer: the hybrid CNN-Transformer model TransUNet (Chen et al., 2021) and the pure Transformer model SwinUNet (Cao et al., 2022).

Table 1 shows the average dice similarity coefficient (DSC), 95% average Hausdorff distance (HD) and segmentation results for each abdominal organ. The bar chart (Fig. 6) visually presents the average DSC for each organ. From the results, it can be observed that gallbladder and pancreas have relatively poorer segmentation results compared to other organs. This is because they overlap significantly with other organs, making it challenging to distinguish them from adjacent organs and other structures. The segmentation accuracy for other organs mostly exceeds 85%, averaging 89.56%, indicating good segmentation results.

The SEF-UNet model achieved the best results in terms of average DSC (83.90%) and average HD (15.64%). Compared to basic CNN models like UNet, UNet++, and UNet3+, the SEF-UNet showed significant improvements in segmentation accuracy. It exhibited an 8.29% increase in average DSC compared to UNet and a 6.03% increase compared to the best-performing UNet++. Additionally, it demonstrated a reduction in Hausdorff distance by 13.03%–16.86%, surpassing a 40% decrease. When compared to the hybrid CNN-Transformer model TransUNet and the pure Transformer model SwinUNet, SEF-UNet not only showed improved average DSC by 2.59% and 3.41%, respectively, but also exhibited a reduction in HD by 13.15% and 3.65%, showcasing a clear advantage.

The SEF-UNet model has a parameter count of 24.65 M, placing it in the middle range when compared to other models. It has a comparable parameter count to the larger CNN model UNet3+, but it significantly reduces the parameter count by over 75% compared to the Transformer-based TransUNet. This optimization addresses the issue of excessive parameters in Transformer models, leading to improved segmentation performance without excessive computational resource consumption.

Additionally, to further validate the improvements of our experimental model in terms of real-time performance and low latency, we compared the GFLOPs and the average time taken to segment each test sample across all models, which is shown in Table 2. It can be observed that our model has the lowest GFLOPs, indicating the least number of floating-point operations required. Compared to the highest GFLOPs model, UNet3+, our model reduces the GFLOPs by 29.8 times.

Furthermore, we measured the average time taken by each model to process nine test samples from the BCV dataset. Our model takes only 1.41 min on average, making it the fastest model. This is more than three times faster than the slowest model, TransUNet. It is also important to note that a single test sample is 3D data, with its width and height resized to the target size during training, while the number of channels remains unchanged. The average number of channels for all samples is 125.9. The testing process involves processing a single 2D image and then combining them into a 3D image. Therefore, the average time taken by our experimental model to process a single 2D image is 0.67 s. These data can be found in Table 2.

In addition, we have presented the average DSC and HD of both the proposed experimental model and the comparative models for each test case (Table 3). The more intuitive comparative results are illustrated in Figs. 7 and 8. It can be observed that our experimental model exhibits the most consistent performance across all test cases, with both DSC and HD outperforming those of the other comparative models on the whole Fig. 9 displays the visual segmentation results of nine test cases. It can be observed that the segmentation boundaries are clear, indicating stable segmentation performance.

In summary, the experimental results indicate that, in medical image segmentation tasks, the SEF-UNet model exhibits significant advantages over CNN models, hybrid CNN-Transformer models, and pure Transformer models.

Ablation experiment

To comprehensively validate the effectiveness of our experimental model, we conducted multiple sets of experiments. We explored different combinations of encoders, decoders, SEFormer, and dCUP within the SEF-UNet framework, and validated the model on the BCV dataset. Additionally, we investigated the impact of varying the number of stages in the model on segmentation performance. Furthermore, we used a larger resolution of 384 × 384 as the model input to further explore the effectiveness of the SEF-UNet model. This series of experiments aims to thoroughly assess the model’s performance under different configurations, ensuring its robustness and generalizability across various conditions.

Architecture of the encoder

To validate the effectiveness of the improved encoder architecture, we conducted experiments using classical models: ResNet-50 (He et al., 2016), high-performance CNN network ConvNeXt (Liu et al., 2022), and the state-of-the-art MetaFormer (instantiated as ConvFormer model; Yu et al., 2023) as encoder structures. These encoders were combined with our decoder, and experiments were performed on the BCV dataset.

Results from Table 4 indicate that our SEF-UNet outperforms all three comparison models in the segmentation of all organs. The average dice score improved by 4.72%–9.45%, and the average Hausdorff distance decreased by 5.3%–16.86%, demonstrating a significant enhancement in segmentation performance.

Table 4:

Comparative experiments exploring the validity of model encoder and decoder separately.

Framework		DSC	HD95	Organs
Encoder	Decoder			Ao	Ga	LK	RK	Li	Pa	Sp	St
ResNet-50	dCUP	74.45	32.50	86.75	60.53	77.78	71.25	93.25	51.07	82.93	72.02
ConvNeXt	dCUP	77.11	20.94	86.41	61.31	79.56	74.94	93.00	60.13	83.78	77.76
ConvFormer	dCUP	79.18	25.55	86.94	64.19	81.19	76.32	93.88	62.15	89.39	79.34
3SE+1Trans	dCUP	82.20	23.98	84.77	67.67	89.69	86.06	94.46	63.24	87.10	84.62
2SE+2Trans	dCUP	83.69	17.14	87.02	65.77	91.17	87.58	94.66	66.16	90.67	86.45
SEFormer	None	68.18	37.49	77.21	48.46	72.82	69.06	91.96	42.90	80.20	62.85
SEFormer	CUP	82.13	23.34	87.35	66.25	90.21	88.80	94.02	60.03	88.54	81.86
Ours		83.90	15.64	86.11	69.38	91.66	89.03	94.38	64.45	91.12	85.08

DOI: 10.7717/peerjcs.2238/table-4

Notes:

The best results are shown in bold.

Additionally, we explored replacing the last one or two stages of the four SEFormer blocks with Transformer blocks. The results showed that the model using SEFormer blocks in all four stages performed the best. This demonstrates that our SEF-UNet, incorporating the SEFormer model with the squeeze-and-excitation module as the encoder, has a stronger capability to learn effective representations for medical image segmentation tasks. This experiment confirms the superiority of the SEF-UNet model in terms of encoder architecture improvements.

In Tables 5 and 6, the DSC and HD for the ablation experiments on nine test cases are listed. Figures 10 and 11 respectively present the DSC and HD for different encoder and decoder experiments.

Table 5:

Average dice score (%) and average Hausdorff distance (mm) for different encoder and decoder experiments (sample 1–5).

Framework		Sample1		Sample2		Sample3		Sample4		Sample5
Encoder	Decoder	DSC	HD	DSC	HD	DSC	HD	DSC	HD	DSC	HD
ResNet-50	dCUP	74.07	39.51	80.43	22.44	72.26	31.20	62.99	47.58	70.00	44.50
ConvNeXt	dCUP	71.41	28.22	82.39	15.12	87.71	9.58	72.48	30.75	68.47	28.80
ConvFormer	dCUP	74.55	39.97	83.96	21.88	83.47	21.02	75.73	27.12	69.82	50.27
3SE+1Trans	dCUP	74.63	43.00	85.39	17.84	87.68	32.52	77.09	28.55	81.70	25.48
2SE+2Trans	dCUP	77.86	35.53	85.44	14.67	87.53	12.21	83.21	20.24	82.13	25.15
SEFormer	None	72.84	41.02	77.39	26.69	70.22	30.73	57.13	55.17	46.03	73.24
SEFormer	CUP	76.13	35.16	85.23	21.61	87.51	9.98	74.32	23.56	79.66	31.49
Ours		78.70	27.97	85.58	23.04	87.73	8.99	84.16	20.18	82.40	24.31

DOI: 10.7717/peerjcs.2238/table-5

Notes:

The best results are shown in bold.

Table 6:

Average dice score (%) and average Hausdorff distance (mm) for different encoder and decoder experiments (sample 6–9).

Framework		Sample6		Sample7		Sample8		Sample9
Encoder	Decoder	DSC	HD	DSC	HD	DSC	HD	DSC	HD
ResNet-50	dCUP	81.01	18.93	82.15	15.18	75.68	43.44	71.47	29.75
ConvNeXt	dCUP	73.92	21.70	82.53	17.54	81.05	15.19	74.06	21.60
ConvFormer	dCUP	82.70	11.48	84.00	14.74	81.55	19.65	76.81	23.83
3SE+1Trans	dCUP	87.43	23.89	87.30	10.08	81.93	21.84	76.65	12.64
2SE+2Trans	dCUP	87.43	7.44	88.42	3.12	83.28	18.38	77.92	17.53
SEFormer	None	80.19	17.86	64.26	28.78	73.95	29.27	71.62	34.68
SEFormer	CUP	87.34	7.24	88.25	3.23	82.34	15.93	78.42	61.85
Ours		87.49	6.33	86.87	3.08	83.42	14.84	78.76	12.05

DOI: 10.7717/peerjcs.2238/table-6

Notes:

The best results are shown in bold.

Compared to several models such as ResNet, ConvNeXt, and ConvFormer serving as encoders, our experimental model demonstrates significant improvements in segmentation performance both overall and on individual cases. When investigating the impact of the number of SEFormer modules on segmentation results, it is evident that our model, consisting of four SEFormer modules, exhibits more stable DSC and HD across all cases, achieving the overall best performance.

Number of blocks

Our final model adopts a four-stage SEFormer architecture. In such a four-stage architecture, the first stage is mainly responsible for extracting low-level features, typically associated with the original information of the input data, such as edges, textures, etc., in the image. The second stage builds upon the first stage to further extract mid-level features, such as more complex structures like the shape of objects. The third and fourth stages can extract higher-level abstract features, including object categories, scene semantics, and other advanced concepts.

To validate the effectiveness of the four-stage architecture, experiments were conducted using three-stage and two-stage architectures, as shown in Fig. 12. The model with four blocks significantly outperformed the other two models, achieving a difference of 11.63% in overall DSC. There were also noticeable differences in the segmentation results of various specific organs. Upon observation, it was found that the segmentation results for organs like Aorta and Liver had smaller differences, and satisfactory results could be obtained with smaller models. However, organs like Gallbladder, Kidney, Pancreas, Stomach, Spleen, etc., with unclear boundaries and prone to confusion with other abdominal contents, required more complex models for clear segmentation. This further validates the effectiveness of our four-stage architecture.

Preprocessing

We have added a comparative experiment between pre-processed and non-pre-processed data. We will train our model using data with random flip and random rotation operations, as well as data without these operations, and then make predictions on the same test set. We have selected some slices from the prediction results of both cases for visual comparison, as shown in Fig. 13. It can be seen that, in terms of segmentation accuracy and detail precision, the results obtained by training with pre-processed data are significantly better.

Qualitative visualization

Next, we conducted a qualitative comparative analysis of the experimental results, as shown in Fig. 14. The observations include:

(1) Pure CNN methods like UNet and ResNet-50 tend to exhibit over-segmentation or under-segmentation. For instance, in the third row, UNet produces multiple segments outside the boundary between the lungs and stomach, and ResNet-50 in the second row fails to segment a significant portion of the stomach. In comparison, our SEF-UNet method demonstrates stronger segmentation capabilities, both globally and locally.

(2) SEF-UNet achieves more precise contour segmentation. In the third row, SEF-UNet nearly approaches the ground truth for the outer edge of the lungs, while the other three models extend the outer edge into some actual organ regions, leading to increased false-positive areas. This indicates that SEF-UNet has an advantage in handling organ edge segmentation issues.

Overall, our SEF-UNet model outperforms other models in precise edge segmentation, noise suppression, and accurate shape recognition.