RPMS-DSAUnet: A Segmentation Model for the Pancreas in Abdominal CT Images

1  Introduction

Pancreatic cancer is a common type of pancreatic tumor. In 2020, approximately 496,000 new cases of pancreatic cancer were diagnosed worldwide, with 466,000 deaths, reflecting a mortality rate as high as 94%. In China, 120,000 new cases and an equal number of deaths were reported that year, with the annual death-to-incidence ratio approaching 1:1 and a five-year survival rate of merely 5%–10% [1]. According to the 2024 National Cancer Center report, pancreatic cancer ranks among the top five causes of cancer-related deaths in China, accounting for 120,000 fatalities annually, with male mortality rates significantly higher than those of females (notably rising trends in prostate, colorectal, and pancreatic cancer mortality) [2]. Accurate automatic organ segmentation is a prerequisite for quantitative assessment and computer-aided diagnosis. Currently, automated detection and segmentation methods for organs such as the kidney, liver, and heart have achieved a high level of accuracy, with Dice Similarity Coefficients exceeding 90% [3]. However, segmentation of the pancreas, considered one of the most challenging tasks in abdominal organ segmentation, remains difficult due to the variability and complexity of pancreatic regions in CT images [4].

Many traditional methods including thresholding [5], edge-detection [6], clustering [7], and active contour models [8] were proposed to address the demand for efficient and accurate segmentation of target organs in medical images. However, when applied to pancreatic segmentation, these traditional methods perform poorly. More critically, traditional methods heavily rely on manual operations, lack self-learning and adaptation capabilities, and struggle to effectively deal with the interference of artifacts and air shadows, resulting in low-quality segmentation results.

Deep learning has revolutionized medical image segmentation by overcoming traditional methodological limitations. Modern architectures achieve high-precision organ delineation through automated feature extraction and adaptive learning, significantly reducing manual intervention while ensuring computational efficiency. This progress has catalyzed the development of specialized models for various anatomical structures, with performance benchmarks continually advancing across clinical imaging modalities. Despite these advancements, the aforementioned medical image segmentation models face challenges in pancreas segmentation.

•   The morphological variability and fuzzy boundaries of the pancreas in CT scans make the extraction of morphological features ineffective.

•   The pancreas occupies a minimal proportion of the image. As network depth increases, semantic and spatial information loss occurs, resulting in degraded feature map resolution.

•   Severe foreground-background class imbalance biases models toward background prediction. Such imbalance suppresses the semantic representation of the small pancreatic target. Consequently, spatial misalignment arises between reconstructed decoder feature maps and ground truth annotations. This misalignment ultimately leads to suboptimal segmentation performance.

In this paper, a novel RPMS-DSAUnet is proposed for the accuracy segmentation of pancreas. The model consists of the Residual Pyramid Squeeze Attention module, the Multi-Scale Feature Extraction module and the Dimensional Squeeze Attention module. This method demonstrated superior performance on the NIH and MSD datasets. The main contribution of this paper are:

•   A Residual Pyramid Squeeze Attention (RPSA) module that employs hierarchical multi-resolution feature extraction using four-scale convolutional kernels (3 × 3 to 9 × 9) with dynamic feature weighting is incorporated into the encoder. This design covers typical pancreatic size variations while enhancing edge localization through adaptive reinforcement of pancreatic boundary information.

•   A Multi-Scale Feature Extraction (MS) module is introduced to expand convolution receptive fields while preserving feature map resolution. By integrating multi-scale semantic contexts, it can mitigate network depth-induced feature degradation.

•   The decoder implements a Dimensional Squeeze Attention (DSA) module combining spatial compression to suppress irrelevant background regions and channel-wise recalibration to prioritize pancreatic features. This dual mechanism effectively addresses class imbalance by reducing interference from adjacent organs and enhancing small target sensitivity.

To handle foreground-background imbalance and the pancreas’ minimal spatial presence, we use a hybrid loss function combining Focal loss and Dice loss to train the proposed network. Experimental validation demonstrates this loss formulation effectively suppresses irrelevant regions while enhancing focus on target areas, achieving optimal network performance through balanced optimization of class-specific segmentation metrics.

The remainder of this paper is organized as follows: Section 2 reviews related work. Section 3 details the network architecture and module functionalities. Section 4 presents experimental results and comparative analyses with state-of-the-art methods. Section 5 concludes the study with key findings and future research directions.

2  Related Work

2.1 Classic U-Net Architecture and Its Variants

The U-Net architecture [9] establishes a foundational framework for medical image segmentation. Its encoder-decoder structure with symmetric skip connections achieves precise boundary localization through multi-resolution feature fusion. However, the symmetric design restricts dynamic interactions across resolutions, leading to inadequate performance when segmenting highly variable pancreatic morphology. To address specific limitations, multiple variants have been proposed.

AttnU-Net [10] introduces spatial attention gates to suppress irrelevant adjacent organs (e.g., liver and spleen), significantly improving pancreatic boundary segmentation. Limited by its local receptive field, it struggles to capture long-range dependencies (e.g., morphological correlations between the pancreatic head and tail). Res-UNet [11] alleviates gradient vanishing via residual connections [12], while UNet++ [13] enhances network flexibility through nested dense skip connections. Nevertheless, both architectures remain constrained by the locality of convolutional operations, lacking global context modeling capabilities (e.g., spatial relationships between the pancreas and surrounding vessels). The nnU-Net framework [14] addresses cross-dataset generalization via automated preprocessing and architecture adaptation. However, its heavy reliance on full supervision limits performance in clinical settings with scarce annotations. Zhang et al. [15] propose a pancreatic segmentation method using lightweight Deep Convolutional Neural Networks (DCNN) modules and spatial prior propagation. By decoupling the complex small-target segmentation task through a twin-network architecture, the approach significantly reduces learning difficulty. Although the current methods can achieve resonable segmentation results, they usually underutilizen 3D spatial neighborhood information, leading to suboptimal accuracy in complex scenarios (e.g., abrupt morphological changes or low-contrast regions).

2.2 Transformer-Based Architectures and Global Context Modeling

Recent studies integrate Transformers to capture global dependencies, overcoming the locality constraints of convolutions. TransUNet [16] and UNETR [17] leverage Transformer self-attention mechanisms [18] to model global context, enhancing robustness to pancreatic morphological variations. However, the quadratic computational complexity of self-attention reduces efficiency in high-resolution. PIS-Unet [19] achieves local-global feature synergy through axial multi-layer perceptrons (MLPs), supporting multi-modal segmentation (e.g., CT-MRI fusion). TCU-Net [20] optimizes segmentation boundaries for peripancreatic vessels (e.g., mesenteric arteries/veins) via cross-scale modules. Despite these advances, few-shot generalization requires improvement, and remain highly sensitive to alignment errors in multi-modal inputs. To address data-hungry training and convergence difficulties in Transformer backbones, the nnTransfer framework [21] innovatively introduces a self-supervised generative model for automated 3D segmentation and quantification of pancreatic fat infiltration. This model autonomously captures shared anatomical representations, focusing on peripancreatic texture features and fat infiltration patterns. It accelerates convergence through parameter transfer initialization of segmentation network weights. DA-TransUNet [22] balances accuracy and efficiency via dual-attention mechanisms combined with Transformers, excelling in small-target structures (e.g., pancreas and polyps). However, segmentation insufficiency persists in low-contrast scenarios (e.g., fat-infiltrated pancreatic edges) or morphologically variant organs (e.g., atrophic pancreatitis). Hybrid architectures like DTU-Net [23] integrate DenseASPP for multi-scale feature extraction with Transformer attention, yet high model complexity hinders clinical deployment. Other studies incorporate anatomical priors. Zhang et al. [24] combine multi-atlas registration with 3D deformation constraints, though cross-institution stability requires further optimization.

3  Proposed Methods

3.1 Overview

This paper proposes an end-to-end semantic segmentation model named RPMS-DSAUnet, which is an acronym derived from the names of the three added modules. As shown in Fig. 1, our model processes 256 × 256 input images through five hierarchical layers in both encoder and decoder pathways. The encoder iteratively applies two 3 × 3 convolutions (padding = 1) with ReLU activation for high-dimensional feature extraction, followed by a 2 × 2 maxpooling operation with a stride of 2 for downsampling. The resultant feature maps undergo further downsampling and processing via the RPSA module, enhancing pancreatic recognition through critical feature extraction. To mitigate information loss during network deepening and downsampling, the MS module is integrated into the fourth and fifth encoder layers. This module can reconstruct spatial correlations by fusing multi-scale spatial information. During the decoding phase, upsampling is performed using bilinear interpolation (Scale-factor = 2) to double the feature map dimensions. The upsampled feature maps are concatenated with their corresponding encoder counterparts, and the concatenated features undergo channel reduction via a 3 × 3 convolutional layer. The fused features then enter the DSA module, which eliminates redundancy through spatial-channel squeeze-excitation operations while facilitating effective encoder-decoder information fusion and inter-layer interaction enhancement. Subsequent processing involves two consecutive 3 × 3 convolutional layers for feature refinement. The final layer employs a 1 × 1 convolution without padding to generate two-channel output maps corresponding to the segmentation classes. Except for the output layer, all convolutional layers use Rectified Linear Units (ReLU) as activation functions, while the output layer employs the Sigmoid function to support independent optimization of each overlapping region.

Figure 1: RPMS-DSAUnet network structure diagram

3.2 Residual Pyramid Squeeze Attention

Effective feature extraction is the foundation of precise segmentation algorithms. However, the small size, morphological variability and blurry boundaries of the pancreas make this task difficult. To obtain more complete feature information at the encoder stage, inspired by the Pyramid Squeeze Attention (PSA) module proposed [25], this paper introduces the Residual Pyramid Squeeze Attention (RPSA) module, as illustrated in Fig. 2.

Figure 2: Residual pyramid squeeze attention module

To address the issues of anatomical distortion caused by excessive background suppression during channel-wise weighting in the original PSA module and the loss of high-frequency details due to progressive downsampling in deep networks, this study constructs a cross-layer feature compensation mechanism by introducing residual identity mapping. This design achieves dynamic fusion of multi-scale local details (e.g., pancreatic duct edge gradients) and global semantics (e.g., pancreatic morphological distribution). The proposed RPSA module forces the learning of residual attention mappings, which can effectively reduce gradient decay and suppress high-frequency information loss. Additionally, this module simultaneously extracts channel attention and spatial attention from feature maps while maintaining high efficiency and cost-effectiveness.

Specially, the input feature map Ei is divided into four parts, denoted as [Ei0,Ei1,Ei2,Ei3]. The number of channels for each divided part is C′, where C′ is calculated as follows:

C′=C4.(1)

For each divided channel feature map, we employ four convolutional kernel sizes tailored to pancreatic anatomical characteristics, enabling multi-scale spatial information extraction: 1) Small 3 × 3 kernels capture high-frequency edges for local details (e.g., pancreatic ducts, microscopic lobules); 2) Medium 5 × 5 and 7 × 7 kernels model intermediate structures (e.g., pancreatico-duodenal junction zones); 3) Large 9 × 9 kernels integrate spatial context for organ-level relationships (e.g., pancreatic tail-splenic hilum interface). This hierarchical design achieves cross-scale feature enhancement, thereby improving model robustness against the pancreas’ “small size, morphological variability, and fuzzy boundaries” challenges. The formula for calculating the obtained feature maps is as follows:

Fi=fconv(ki×ki),(2)

here fconv(ki×ki) means convolution which kernel size = ki,i=3,5,7,9.

Then, the multi-scale feature maps are concatenated to obtain the feature map F. The formula for calculating F is given in Eq. (3):

F=Concat([F0,F1,F2,F3]).(3)

After extracting multi-scale features, it is necessary to extract channel attention weights for the feature maps Fi of different scales. This enhances important channels while suppressing redundant ones. The calculation process of SEWeight(⋅) involves the following three steps (Eqs. (4)–(6)).

•   Squeeze:

zc=1H×W∑i=1H∑j=1WFi(c)(i,j)(4)

•   Excitation:

sc=σ(W2⋅δ(W1⋅zc))(5)

•   Reweight:

Wi(c)=sc⋅Fi(c)(6)

Here, Fi(c) denotes the c-th channel of input feature map Fi (c=1,2,⋯,C). H and W are the spatial height and width of the feature map. zc is the Global descriptor for channel c (scalar). W1∈RC×Cr, W2∈RCr×C are weight matrices of fully-connected layers. r is the Channel reduction ratio (default = 16). δ denotes the ReLU activation function. σ is the Sigmoid activation function (constrains sc∈[0,1]). sc is the Attention weight for channel c.

The entire calculation process for the multi-scale channel attention weight vector W can be formulated by Eq. (7):

W=W0⊕W1⊕W2⊕W3,(7)

where ⊕ means entry-wise addition.

In order to achieve information interaction, softmax is further utilized to set weights for the channel attention information.

atti=Softmax(Wi)=exp⁡(Wi)∑i=0i=3exp⁡(Wi).(8)

Finally, to prevent the module from losing its original spatial structure during the learning process, a residual structure is added on the existing basis, resulting in a feature map Oi that retains more multi-dimensional information from the input feature map.

Oi=M⊕Ei.(9)

Next, the feature map Fi of the corresponding scale is multiplied channel-wise with the weighted attention vector. This process can be formulated as:

Mi=Fi⊗atti,i=0,1,2,3,(10)

here ⊗ denotes entry-wise multiplication.

3.3 Extraction of Multi-Scale Semantic Information

The RPSA module can achieve multi-scale feature extraction through channel grouping, but it lacks the ability to capture spatial multi-scale information. Therefore, network deepening may lead to spatial information loss. To compensate for the shortcomings of the RPSA module and extract richer spatial semantic information, the Multi-Scale (MS) module is designed as shown in Fig. 3. This module acquires multi-scale information based on pancreatic distribution characteristics in medical images, followed by feature fusion to enhance channel-wise feature representation. The proposed design effectively mitigates valid information loss caused by network depth, enabling deep encoders to extract more critical pancreatic region features. Initially, a 5×5 convolution kernel is used to preliminarily extract features from the input feature map Ei, resulting in feature map F1. Subsequently, multi-scale depthwise separable convolutions are employed to further extract features. Then, the results of these convolutions at different scales are element-wise summed to obtain feature map F2.

F1=fconv5×5(Ei),(11)

F2=∑i=3,5,7fconv⁡(i×1)(fconv⁡(1×i)(F1)).(12)

Figure 3: Multi-scale feature extraction module

The multi-scale depthwise separable convolutions can capture information of different scales and orientations in the image. The extracted features F2 are taken as the attention for F1, and then transformed through a 1×1 convolution kernel to obtain an attention weight W.

W=F1⊕F2.(13)

Multiply W and Ei element-wise can obtain the weighted output feature map Oi. In this way, important features are emphasized while irrelevant features are suppressed. At the same time, Oi contains richer spatial information, which prevents the loss of information caused by the network depth and further enhances model accuracy. To save computational costs in the network, the MS module is placed in the fourth and fifth layers of the encoder.

Oi=fconv(1×1)(W)⊗Ei.(14)

3.4 Multi-Dimensional Squeeze Attention Module

The input of the DSA module consists of two parts: the feature map extracted by the encoder and the upsampled feature map from the decoder. Di enters the DSA through skip connections and undergoes weighted aggregation with Ei in terms of feature maps via a soft attention mechanism, thereby combining the feature map information from both. Specifically, the feature map Ei integrates information across different channel dimensions for each pixel point through a 1×1 convolution, preserving the original planar structure of the image, and thereby achieving channel reconstruction. Ultimately, a weight map containing channel information is produced. Meanwhile, the spatial and channel information of the feature map Di is squeezed and excited to derive effective spatial and channel feature information. This process is implemented by the Spatial Compression Module (SCM) within DSA module, as illustrated in Fig. 4. This architecture allows for the extraction of both spatial-dimensional and channel-dimensional information, therefore, the critical feature information extracted by the encoder serves as attention for the decoder. With the interaction between local and global information, the decoder can recover the feature map better, leading to more precise segmentation results. The detailed structure of DSA module is shown in Fig. 5. Here, Oi represents the output feature map of the DSA module, Di denotes the feature map extracted by the i-th layer of the encoder, and Ei indicates the feature map output by the (i + 1)-th layer of the decoder.

Figure 4: Spatial compression module

Figure 5: Multi-dimensional squeeze attention module

After inputting the feature map Ei into the DSA module, it first gets upsampled through deconvolution. Then, a 1×1 convolution is used to redefine its channels, enhancing the prominence of feature information in each channel. Meanwhile Di undergoes a 3×3 convolution operation to reshape the spatial features before entering SCM. The SCM obtains a feature weight M that contains the spatial information of the feature map. Subsequently, this feature weight M interacts and integrates with the upsampled feature map. After passing through the Rectified Linear Unit (ReLU) activation function, Batch Normalization, and the sigmoid function, a probability map Xp with bilateral information is generated. This allows the final attention matrix to focus not only on the information from the encoder itself but also on the feature map information generated by the decoder. The calculation process for the output of the SCM and the probability map Xp obtained from the encoder feature map is shown in Eq. (15).

XP=σ(fconv1(gr(fconv1(fCTconv(Ei))+M))).(15)

Here fconv1(⋅), fCTconv(⋅), gr(⋅) and σ(⋅) are 1×1 convolution, transposed convolution, the Relu activation function, and the sigmoid function, respectively.

The feature map Ei contains richer original spatial information. By decoupling the channel and spatial dimensions, we can obtain weight maps for both of the two dimensions in Ei, which can avoid the impact caused by differences in dimensions. By performing element-wise multiplication between the probability map Xp and M, we can obtain a feature map F that highlights semantic information. The feature map F is computed as follows:

F=XP⊗M.(16)

To aggregate information between different layers of the network using soft attention while mitigating the gradient vanishing issue during model training, a residual structure is applied in this module. Firstly, the feature map F undergoes channel reshaping through a 1×1 convolution. Then, it is concatenated with the upsampled feature map of Ei in the channel dimension. Finally, a feature map Oi with collaborative attention in both spatial and channel dimensions is obtained. The calculation process for this is shown in Eq. (17).

Oi=fcat (fCTconv (Ei),fconv1(S)).(17)

In this equation, the fcat(⋅) indicates channel concatenation.

To make the network focus more on the pancreas area and exclude the influence of surrounding organs, it is necessary to eliminate redundant information and extract relationships between adjacent channel pixel. To address this issue, this paper proposes the Spatial Compression Module (SCM), which is inspired by the Squeeze-and-Excitation (SE) module proposed by Hu et al. [26]. The Spatial Compression Module (SCM) processes the input feature map Di∈RB×C×H×W through channel-spatial decoupling, generating two compressed branches:

1.   Height-compressed branch (B×C×1×W): Global pooling along the height dimension (Eqs. (18) and (19)) collapses vertical spatial redundancy while preserving channel-width correlations. Subsequently, fully connected layers (FC) with Sigmoid activation can produce the channel attention weights Mchannel, focusing on pancreas-related channels by aggregating global width-wise statistics.

2.   Channel-compressed branch (B×1×H×W): Pooling along the channel dimension eliminates irrelevant channel noise. Convolutional operations on this branch generate the spatial attention Mspatial with a full-image receptive field, thereby capturing continuous spatial patterns (e.g., pancreatic head–portal vein spatial constraints).

gs=1H×W∑i=1H∑j=1WDs(i,j),(18)

ms=maxDs(i,j),(i,j)∈H×W.(19)

After being squeezed, the feature map outputs a channel feature vector and a spatial feature vector with a global receptive field. Then, as described in Fig. 4, information aggregation is performed on the output feature map Dsahc, Dsacw, Dsmhc, Dsmcw  to obtain a feature map that contains information in both spatial and channel dimensions.

The calculation of Dsahc, Dsacw, Dsmhc, Dsmcw can be derived from Eqs. (20)–(23).

Dsahc=favg(favg(Ds,H),C),(20)

Dsacw=favg(favg(Ds,C),W),(21)

Dshmc=fmax(fmax(Ds,H),C),(22)

Dsmcw=fmax(fmax(Ds,C),W).(23)

favg(D,C) and fmax(D,C) represent the average pooling and max pooling operations, respectively, while C denotes the dimension of the operation.

Subsequently, the fully connected layers are used to reshape the channel and spatial information separately, mapping important information into a new space. Using two fully connected layers with a dimensionality reduction followed by an expansion helps to reduce computational complexity. During the processing of Dsam, the transpose function permute is utilized to change the dimension of Dsam from 1×H×1 to 1×1×H before entering the fully connected layer, and then the dimension format is restored. The attention mechanisms for channels and spaces use Sigmoid as the activation function for the last convolutional layer, mapping values between 0 and 1 to obtain two weight matrices S1 and S2. The attention weights of both are summed to produce the matrix Ds, which allows the final output feature map to incorporate dual attention. Finally, to prevent the loss of important information due to nonlinear mapping during training, the network employs a residual structure to obtain the final output feature map Dsout. The calculation formula for Dsam, Dsma, S1, S2 and Dsout can be derived from Eqs. (24)–(28).

Dsam=Dsahc ⊕Dshmc ,(24)

Dsma=Dsacw ⊕Dsmcw .(25)

S1(Dsam)=11+e−Dsam,(26)

S2(Dsma)=11+e−Dsma ,(27)

Dsout=DS⊗(σ(fFC(ReLU(fFC(Dsam))))⊕σ(P(fFC(ReLU(fFC(P(Dsma)))))))⊕DS.(28)

fFC(⋅) represents the fully connected layer, ReLU(⋅) denotes the ReLU activation function, and P(⋅) stands for the permute function. The fusion of dual attention weights (Eq. (28)) enables:

•   Spatial downscaling (reducing H and C dimensions)

•   Global context capture (integrating cross-organ topology).

It can effectively suppress interference from adjacent organs (e.g., duodenum, spleen) and enhance pancreatic boundary localization. This design leverages joint optimization of spatial-channel redundancy reduction, which is critical for segmenting small, low-contrast anatomical targets (e.g., the pancreas) in medical images.

3.5 Loss Function

In our experiments, we used the sum of Focal loss [27] and Dice loss [28] as the final loss function. The expression of Focal loss is shown as follows:

LF=−∑[yi(1−σ(y^i))γlog⁡σ(y^i)+(1−yi)σ(y^i)γlog⁡(1−σ(y^i))],(29)

and the Dice loss, which is effective for categories imbalance, can be define as:

LDice=1−∑2yiσ(y^i)+ε∑(yi+σ(y^i)+ε),(30)

so the loss function consists of the following:

Lcom=λ1LF+λ2LDice.(31)

yi represents the true label value of a pixel, where pixels in the foreground are labeled as 1 and pixels in the background are labeled as 0. y^i denotes the final predicted value. σ(y^i) is the predicted probability value that maps a pixel to a target, representing the likelihood of the pixel belonging to the foreground or background. The modulation parameter for hard and easy examples is denoted as γ. ε is referred to as the smoothing coefficient, which is typically set to a very small positive number to prevent the denominator in the above formula from being zero.

4  Experiments and Results

4.1 Datasets

We evaluate the performance of the proposed network on two public datasets: 1) 80 abdominal contrast-enhanced CT scans from the National Institutes of Health (NIH) Clinical Center pancreas segmentation dataset [29], which is the most widely used publicly dataset for the pancreas segmentation task. The CT images and label images in the original data were converted to PNG format. The CT image slices have a thickness of 1.5 to 2.5 mm and were acquired using Philips and Siemens Multidetector CT (MDCT) scanners, saved in DICOM format (.dcm). The label images in this dataset are divided into background and pancreatic regions. The pancreas was manually segmented slice-by-slice by a medical student and then reviewed and revised by an experienced radiologist. The label images are saved in NifTi format (.nii). The CT images have a resolution of 512 × 512, and the original dataset contains 18,942 CT slices. After excluding the all-black samples, 6882 slices remained, which were divided into training and testing sets based on cases, with 68 and 12 cases, respectively. Among them, 5,894 slices were used for training, and 988 slices were used for testing. Considering that the pancreas occupies only a small proportion of the CT images, to improve computational efficiency, all CT scans were cropped to 256 × 256, which still fully covers the pancreas within the CT scans. During training, data augmentation techniques such as random scaling, aspect ratio distortion, image flipping, and adjustments to saturation and brightness were applied to the images to enhance the network’s generalization ability, making it more effective and reliable in real-world scenarios. 2) 281 abdominal contrast-enhanced CT scans with labeled pancreas and pancreatic tumor from the Medical Segmentation Decathlon (MSD) challenge pancreas segmentation dataset [30], where each CT volume is 512×512×D, and D∈[37,751] is the number of slices in the CT scan. Following previous studies [31] on the MSD dataset, we combine the pancreas and pancreatic tumor into a single entity as the segmentation target. The dataset is divide into training and testing sets based on cases, with 223 and 58 cases, respectively. The rest of the operations are consistent with the NIH dataset.

4.2 Optimizer

The network in this chapter utilizes PyTorch and is trained in an end-to-end manner on an NVIDIA A6000 GPU with CUDA 12.4 support. The network optimizer is SGD with the optimizer momentum set to 0.99 and (weight-decay = 3e–4) to suppress overfitting. The batch size for training is set to 2, and the initial learning rate is set to 0.005. The learning rate is updated according to “ExponentialLR”, which can be expressed by Formula (32).

lr=lr0×ηt,(32)

where t represents the iteration index during training, and η is the learning rate decay factor, which is set to 0.985 during the training of the network in this paper. The network is trained for a total of 350 epochs.

4.3 Evaluation Criteria

To assess the performance of pancreas segmentation, this paper uses the Dice Similarity Coefficients(Dice), Intersection over Union (IoU), and Precision (Pr) to quantitatively evaluate the segmentation results. The values of both Dice and IoU range from [0, 1], where a value of 0 indicates a completely failed segmentation, and a value of 1 indicates a perfect segmentation. Their calculation formulas are shown in (33)–(35).

Dice/%=2TpFN+2Tp+Fp×100,(33)

IoU/%=TpFN+Tp+FN×100,(34)

Pr/%=TpFN+Tp×100.(35)

TP denotes a sample that is predicted as positive (belonging to a certain category) and whose true label is also positive. FN represents a sample that is predicted as negative (not belonging to a certain category) but whose true label is positive. FP indicates a sample that is predicted as positive but whose true label is negative.

4.4 Experimental Results and Analysis

4.4.1 Ablation Experiments and Results Analysis

To validate the effectiveness of each added module in the model, we conduct the following experiments. Based on the original backbone network, the DSA module, MS module, and RPSA module were sequentially integrated, using the loss function tailored for pancreas segmentation. The experimental results are presented in Table 1. After adding the relevant modules to the U-shaped backbone network, the final Dice and IoU improved by 5.22% and 7.63%, respectively. According to the definitions of the Dice and IoU, the overlap between the prediction results and the ground truth increased, and the similarity between the two sets improved, resulting in optimized segmentation performance. The experimental results indicate that the proposed model can clearly locate the pancreas and effectively identify its boundary structures.

The Loss curve of the model presented in this paper is shown in Fig. 6. Fig. 6 demonstrates that the convergence speed and convergence results of the model in this paper are both relatively excellent.

Figure 6: Training curve of the proposed model

Experimental observations indicated that Dice and IoU metric enhancements were accompanied by marginal precision declines. This phenomenon originated from the original U-shaped backbone network progressively misclassifying background regions as segmentation targets, thereby inflating baseline precision measurements. Furthermore, mathematical formulations were derived to establish the inverse correlation between Dice/IoU increments and pixel accuracy (Pr) decrements. The complete derivation process was formalized in Eqs. (36)–(38).

GT=TP+FN,PD=TP+FP,(36)

PR=TPTP+FP,(37)

Dice=2TPFN+2TP+FP=2TPGT+PD=2GTTP+PDTP=2GTTP+1PR.(38)

As shown in the aforementioned formula, GT represents the groundtruth, and PD represents the prediction result. Since the total number of pixels in the groundtruth remains unchanged, when the positive impact of the increase in True Positives (TP) on the Dice outweighs the negative impact of the decrease in Precision on Dice, the value of Dice will also increase accordingly. Therefore, the improvement in the Dice and the decrease in pixel accuracy (Pr) are not contradictory. This further demonstrates that with the improvement of the network model, the model performance has been enhanced, avoiding predicting more background as the segmentation target.

Fig. 7 displays the segmentation results of the ablation experiments. From it, we can observe that after adding the DSA module, some erroneous segmentation foregrounds are suppressed. By further incorporating the MS module, the model can learn more accurate pancreatic structures. Finally, by incorporating the RPSA module, the model not only improves its ability to extract boundary information but also focuses on the overall shape features of the pancreas. Additionally, employing the Lcom loss function during training can reduce the impact of sample imbalance, thereby improving the segmentation accuracy. These results further validate the effectiveness of each module designed in this paper.

Figure 7: Diagram of ablation experiment results. The red patches represent the segmentation results of the model, and the yellow lines outline the groundtruth

4.4.2 Experimental Results on NIH dataset

To further validate the effectiveness of the proposed model, comparative experiments were conducted with current mainstream models on the NIH dataset. These models include U-Net [9], AttnU-Net [10], RefineNet [11], U-Net++ [13], UNETR [17], TransUnet [16] and DA-TransUnet [22]. Table 2 presents the Dice and IoU obtained by the above-mentioned models.

The experimental results presented in Table 2 demonstrate that the IoU and Dice obtained by the proposed model are higher than those of the aforementioned classic network models. This indicates that the segmentation results produced by our network have a higher degree of overlap with the groundtruth labels than the compared models, showing that the proposed model for the pancreas is effective.

In Fig. 8, comparisons reveal the qualitative segmentation results of the proposed model and compared models. From Fig. 8, it can be observed that the proposed model performs better in challenging scenarios where local and overall structures are ambiguous. When U-Net is used for pancreas segmentation, its single-scale convolutional kernels and simple skip connections result in less accurate segmentation and poorer structure representation. AttenU-Net, which introduces an attention mechanism based on U-Net, can further extract information and pay more attention to the pancreas region. ReFineNet effectively combines features at different scales through long residual connections and multi-resolution fusion to generate high-resolution semantic feature maps. However, redundancy information may lead to misclassification. UNETR employs pure Transformers as encoder to learn sequential representations and capture global multi-scale information, but may lose local details in pancreas segmentation tasks. UNet++ adopts the structure of U-Net and incorporates DenseNet’s [32] dense connections. This dense connection approach preserves and reconstructs both global and local information. Generally, UNet++ can recognize the overall structure of the pancreas, but may also introduce redundant information. DA-TransUNet still struggles with edge segmentation in low-contrast scenarios (e.g., fat-infiltrated pancreatic boundaries) and morphologically variant organs (e.g., atrophic pancreatitis). This failure stems from its limited adaptive feature extraction, which compromises boundary delineation. However, in more complex cases, it may face issues of under-segmentation or over-segmentation.

Figure 8: Segmentation results of different models. The red patches represent the segmentation results of the model, and the yellow lines outline the groundtruth

Compared to the other networks, the proposed model exhibits superior segmentation performance. The DSA module effectively combines information from the encoder and decoder, screening feature information in both spatial and channel dimensions to better restore the feature map. Furthermore, the MS module enlarges the network’s receptive field. This allows it to focus on the overall structure of the pancreas, preventing inadequate segmentation. At the same time, the MS module integrates features of different scales, enriching the features that enter the decoder. Finally, the RPSA module not only extracts useful information from the encoder part but also strengthens the pancreas feature information and suppresses redundant information, making the model pay more attention to the pancreas area.

4.4.3 Experimental Results on MSD Dataset

Table 3 shows the comparison of the proposed network with the seven aforementioned models.

As shown in Table 3, the proposed network achieves a Dice scores of 66.57% and an IoU scores of 79.93% on the MSD dataset, demonstrating significant advantages compared to state-of-the-art methods.

Fig. 9 exhibits the qualitative segmentation results of the proposed model and compared models on the MSD dataset, which reflects our network’s ability to achieve accurate segmentation of the pancreas across different datasets.

Figure 9: Segmentation results of different models. The red patches represent the segmentation results of the model, and the yellow lines outline the groundtruth

5  Conclusion

Pancreas segmentation in abdominal CT images faces significant challenges due to the organ’s morphological variability, blurred boundaries, semantic information loss from its small anatomical footprint, and severe foreground-background class imbalance. To systematically address these limitations, we propose the RPMS-DSAUnet. The approach has three pivotal strengths: firstly, the RPSA module extracts critical feature information to enhance target identification accuracy; secondly, the MS module expands the convolutional kernels’ receptive field while integrating multi-scale semantic information to mitigate feature map degradation caused by network depth; finally, the DSA module combines spatial dimension reduction with global context capture through spatial squeeze operations and channel attention mechanisms, effectively suppressing irrelevant regions while highlighting critical areas for feature map restoration.

In summary, the combination of these modules establishes a cohesive framework that solves boundary ambiguity through adaptive spatial weighting, preserves fine-grained details via multi-scale context aggregation, and optimizes feature alignment using cross-dimensional attention. The proposed RPMS-DSAUnet delivers a robust solution for precise pancreatic segmentation in abdominal CT imaging, demonstrating superior capability in handling anatomical variability and class imbalance.

Our future work will focus on: (1) Optimizing computational efficiency of the DSA module for real-time clinical deployment; (2) Extending validation to multi-center datasets to enhance model generalizability.

Acknowledgement: Not applicable.

Funding Statement: This research was supported by the National Natural and Science Foundation of China under Grant No. 12301662, and Zhejiang Provincial Natural Science Foundation of China under Grant No. LQ21F030019.

Author Contributions: Study conception and design: Tiren Huang, Chong Luo, Xu Li; Data collection: Xu Li; Analysis and interpretation of results: Tiren Huang, Xu Li; Draft manuscript preparation: Chong Luo, Xu Li. All authors reviewed the results and approved the final version of the manuscript.

Availability of Data and Materials: The NIH dataset is openly available in the Cancer Imaging Archive (TCIA) at https://www.cancerimagingarchive.net/collection/pancreas-ct/ (accessed on 24 July 2025). The MSD dataset is available from http://medicaldecathlon.com/ (accessed on 24 July 2025).

Ethics Approval: Not applicable.

Conflicts of Interest: The authors declare no conflicts of interest to report regarding the present study.

References

1. Sung H, Ferlay J, Siegel RL. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA A Cancer J Clin. 2021;71(3):209–49. doi:10.3322/caac.21660. [Google Scholar] [PubMed] [CrossRef]

2. Han B, Zheng R, Zeng H, Wang S, Sun K, Chen R, et al. Cancer incidence and mortality in China, 2022. J Natl Cancer Cent. 2024;4(1):47–53. doi:10.1016/j.jncc.2024.01.006. [Google Scholar] [PubMed] [CrossRef]

3. Zhou SK, Greenspan H, Davatzikos C, Duncan JS, van Ginneken B, Madabhushi A, et al. A review of deep learning in medical imaging: imaging traits, technology trends, case studies with progress highlights, and future promises. Proc IEEE Inst Electr Electron Eng. 2021;109(5):820–38. doi:10.1109/JPROC.2021.3054390. [Google Scholar] [PubMed] [CrossRef]

4. Huang M, Huang C, Yuan J. Fixed-point deformable U-Net for pancreas CT segmentation. In: Proceedings of the Third International Symposium on Image Computing and Digital Medicine; 2019 Aug 24–26; Xi’an, China. p. 283–7. doi:10.1145/3364836.3364894. [Google Scholar] [CrossRef]

5. Wang C, Yang J, Lv H. Otsu multi-threshold image segmentation algorithm based on improved particle swarm optimization. In: 2019 IEEE 2nd International Conference on Information Communication and Signal Processing (ICICSP); 2019 Sep 28–30; Weihai, China. p. 440–3. doi:10.1109/ICICSP48821.2019.8958573. [Google Scholar] [CrossRef]

6. Kozegar E, Soryani M, Behnam H. Mass segmentation in automated 3-D breast ultrasound using adaptive region growing and supervised edge-based deformable model. IEEE Trans Med Imaging. 2018;37(4):918–28. doi:10.1109/TMI.2017.2787685. [Google Scholar] [PubMed] [CrossRef]

7. Ng HP, Ong SH, Foong KWC, Goh PS, Nowinski WL. Medical image segmentation using K-means clustering and improved watershed algorithm. In: IEEE Southwest Symposium on Image Analysis and Interpretation; 2006 Mar 26–28; Denver, CO, USA. p. 61–5. doi:10.1109/ssiai.2006.1633722. [Google Scholar] [CrossRef]

8. Kim J, Çetin M, Willsky AS. Nonparametric shape priors for active contour-based image segmentation. Signal Process. 2007;87(12):3021–44. doi:10.1016/j.sigpro.2007.05.026. [Google Scholar] [CrossRef]

9. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. Med Image Comput Comput-Assist Interv. 2015;9351:234–41. [Google Scholar]

10. Oktay O, Schlemper J, Folgoc LL, Lee M, Heinrich M, Misawa K, et al. Attention U-Net: learning where to look for the pancreas. arXiv:1804.03999. 2018. [Google Scholar]

11. Diakogiannis FI, Waldner F, Caccetta P, Wu C. ResUNet-a: a deep learning framework for semantic segmentation of remotely sensed data. ISPRS J Photogramm Remote Sens. 2020;162:94–114. doi:10.1016/j.isprsjprs.2020.01.013. [Google Scholar] [CrossRef]

12. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2016 Jun 27–30; Las Vegas, NV, USA. p. 770–8. [Google Scholar]

13. Zhou Z, Siddiquee MMR, Tajbakhsh N. UNet++: redesigning skip connections to exploit multiscale features in image segmentation. IEEE Trans Med Imaging. 2020;39(6):1856–67. doi:10.1109/tmi.2019.2959609. [Google Scholar] [PubMed] [CrossRef]

14. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2020;18(2):203–11. doi:10.1038/s41592-020-01008-z. [Google Scholar] [PubMed] [CrossRef]

15. Zhang D, Zhang J, Zhang Q, Han J, Han J. Automatic pancreas segmentation based on lightweight DCNN modules and spatial prior propagation. Pattern Recognit. 2020;114(6):107762. doi:10.1016/j.patcog.2020.107762. [Google Scholar] [CrossRef]

16. Chen J, Lu Y, Yu Q, Luo X, Adeli E, Wang Y, et al. TransUNet: transformers make strong encoders for medical image segmentation. arXiv:2102.04306. 2021. [Google Scholar]

17. Hatamizadeh A, Yang D, Tang Y, Nath V, Myronenko A, Landman B, et al. UNETR: transformers for 3D medical image segmentation. In: 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV); 2022 Jan 3–8; Waikoloa, HI, USA. p. 1748–58. [Google Scholar]

18. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. In: Advances in neural information processing systems (NeurIPS). Red Hook, NY, USA: Curran Associates Inc.; 2017. p. 5998–6008. [Google Scholar]

19. Zhang C, Wang J. PIS-Net: efficient medical image segmentation network with multivariate downsampling for point-of-care. Entropy. 2024;26(4):284. doi:10.3390/e26040284. [Google Scholar] [PubMed] [CrossRef]

20. Shi Z, Li Y, Zou H. TCU-Net: transformer embedded in convolutional U-shaped network for retinal vessel segmentation. Sensors. 2023;23(10):4897. doi:10.3390/s23104897. [Google Scholar] [PubMed] [CrossRef]

21. Zhang G, Zhan Q, Gao Q, Mao K, Yang P, Gao Y, et al. Automated pancreatic segmentation and fat fraction evaluation based on a self-supervised transfer learning network. Comput Biol Med. 2024;170(Suppl 1):107989. doi:10.1016/j.compbiomed.2024.107989. [Google Scholar] [PubMed] [CrossRef]

22. Sun G, Pan Y, Kong W, Xu Z, Ma J, Racharak T, et al. DA-TransUNet: integrating spatial and channel dual attention with transformer U-net for medical image segmentation. Front Bioeng Biotechnol. 2024;12:1398237. doi:10.3389/fbioe.2024.1398237. [Google Scholar] [PubMed] [CrossRef]

23. Fei C, Luo J. DTUnet: a transformer-based UNet combined with DenseASPP module for pancreas segmentation. In: 2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP–BMEI); 2022 Nov 5–7; Beijing, China. p. 1–7. doi:10.1109/CISP-BMEI56279.2022.9980137. [Google Scholar] [CrossRef]

24. Zhang Y, Wu J, Liu Y, Chen Y, Chen W, Wu EX, et al. A deep learning framework for pancreas segmentation with multi-atlas registration and 3D level-set. Med Image Anal. 2021;68(2):101884. doi:10.1016/j.media.2020.101884. [Google Scholar] [PubMed] [CrossRef]

25. Zhang H, Zu K, Lu J, Zou Y, Meng D. EPSANet: an efficient pyramid split attention block on convolutional neural network. arXiv:2105.14447. 2021. [Google Scholar]

26. Hu J, Shen L, Sun G. Squeeze-and-excitation networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2018 Jun 18–23; Salt Lake City, UT, USA. p. 7132–41. doi:10.1109/CVPR.2018.00745. [Google Scholar] [CrossRef]

27. Lin TY, Goyal P, Girshick R, He K, Dollár P. Focal loss for dense object detection. In: 2017 IEEE International Conference on Computer Vision (ICCV); 2017 Oct 22–29; Venice, Italy. p. 2999–3007. doi:10.1109/ICCV.2017.324. [Google Scholar] [CrossRef]

28. Milletari F, Navab N, Ahmadi SA. V-Net: fully convolutional neural networks for volumetric medical image segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV); 2016 Oct 25–28; Stanford, CA, USA. p. 565–71. doi:10.1109/3DV.2016.79. [Google Scholar] [CrossRef]

29. Roth HR, Lu L, Farag A, Shin HC, Liu J, Turkbey E, et al. DeepOrgan: multi-level deep convolutional networks for automated pancreas segmentation. arXiv:1506.06448. 2015. [Google Scholar]

30. Simpson AL, Antonelli M, Bakas S, Bilello M, Farahani K, van Ginneken B, et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms. arXiv:1902.09063. 2019. [Google Scholar]

31. Li J, Lin X, Che H, Li H, Qian X. Pancreas segmentation with probabilistic map guided bi-directional recurrent UNet. Phys Med Biol. 2021;66(11):115010. doi:10.1088/1361–6560/abfce3. [Google Scholar] [PubMed] [CrossRef]

32. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017 Jul 21–26; Honolulu, HI, USA. p. 2261–9. doi:10.1109/CVPR.2017.243. [Google Scholar] [CrossRef]