학술

A boundary-sensitive and structure-adaptive lightweight network for medical image segmentation

PLOS ONE
CC BY
이 매체는 공공·자유 라이선스로 본문을 직접 표시합니다.

Figures
Abstract
Reliable medical image segmentation remains challenging in weak-contrast boundary regions, ambiguous anatomical transitions, and structurally heterogeneous areas. In addition, many high-performing models improve overlap-based metrics by introducing increasingly complex architectures, making it difficult to balance segmentation accuracy, boundary fidelity, and model compactness. To address these issues, this paper presents a boundary-sensitive and structure-adaptive lightweight framework for CT liver segmentation. The proposed method integrates three task-oriented components in a U-shaped architecture: Boundary-Sensitive Semantic Recalibration Attention (BSRA), Structure-Adaptive Scale Selection Fusion (SASF), and a Detail-Compensated Residual Lightweight Block (DCRB). Experiments were conducted on the CT subset of the CHAOS dataset under a patient-level five-fold cross-validation protocol. For each fold, metrics were first averaged across the corresponding test cases, and the reported values summarize the five fold-specific test means. Compared with the baseline U-Net, introducing BSRA increased Dice from 0.8443 to 0.8492 and improved Boundary F1-score from 0.7728 to 0.8081. The Boundary F1 improvement was statistically significant in the paired case-level Wilcoxon signed-rank test, whereas the Dice change was not significant, supporting the boundary-oriented role of BSRA. With the further incorporation of SASF, the performance-oriented variant achieved the best overall results, including a Dice coefficient of 0.8929, an Intersection over Union (IoU) of 0.8663, a Boundary F1-score of 0.8466, a 95% Hausdorff Distance (HD95) of 6.5729, and an Average Symmetric Surface Distance (ASSD) of 1.8947. After introducing DCRB, the lightweight variant still maintained competitive performance, attaining a Dice coefficient of 0.8710, an IoU of 0.8463, and a Boundary F1-score of 0.8392, while reducing the parameter count from 14.70M to 2.47M and the FLOPs from 31.81G to 8.44G. Additional validation on LiTS reproduced the same overall performance trends on a second liver CT dataset. These results show that the proposed framework improves region-level segmentation, boundary-level structural fidelity, and model compactness, providing a practical balance for CT liver segmentation.
Citation: Guo Y, Liu Y, Ma S (2026) A boundary-sensitive and structure-adaptive lightweight network for medical image segmentation. PLoS One 21(11): e0363127. https://doi.org/10.1371/journal.pone.0363127
Editor: Xiaowei Li, Sichuan University, CHINA
Received: May 21, 2026; Accepted: November 4, 2026; Published: November 20, 2026
Copyright: © 2026 Guo et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The datasets used in this study are publicly available and have been appropriately cited in the manuscript, including the CT subset of the CHAOS dataset and the Liver Tumor Segmentation Benchmark (LiTS) dataset. No new raw data were generated in this study. The source code and scripts required to reproduce the main results are publicly available at GitHub: https://github.com/xu-c22/imageseg.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
1 Introduction
Medical image segmentation is a fundamental task in medical image analysis and computer-aided clinical workflows, and has been widely studied for anatomical delineation, lesion analysis, treatment planning, and quantitative assessment because it provides precise structural information for downstream clinical decision-making [1,2]. Among various imaging modalities, computed tomography (CT) provides rich anatomical information and is extensively used in abdominal organ analysis and disease evaluation [3,4]. However, accurate medical image segmentation remains challenging in practice. First, organ boundaries in CT images are often ambiguous due to low contrast, partial volume effects, and intensity overlap between adjacent tissues. Second, the morphology, scale, and spatial distribution of target organs vary considerably across slices and patients, which makes it difficult for models based on fixed receptive fields to achieve robust segmentation. In addition, although many recent methods pursue higher segmentation accuracy through increasingly sophisticated architectures and stronger context modeling, such designs often introduce higher model complexity and computational cost, limiting their efficiency and practicality in real-world medical scenarios [5,6].
In recent years, deep learning-based segmentation methods, especially U-Net and its variants, have achieved remarkable progress in medical image segmentation. By integrating encoder-decoder architectures with skip connections, these methods significantly improve dense prediction performance and have become the dominant paradigm in this field [7,8]. Nevertheless, many existing studies mainly focus on improving average overlap metrics such as Dice coefficient and Intersection over Union (IoU), while paying insufficient attention to several issues that are particularly critical from the medical perspective. On the one hand, blurred boundaries, semantically confusable transition regions, and structurally complex local areas are still difficult to segment accurately. On the other hand, many methods enhance performance by introducing increasingly complex modules, but the resulting models are often computationally expensive and less suitable for efficient deployment [9,10]. Therefore, how to design a segmentation network that not only improves the representation of ambiguous boundary regions and scale-sensitive structures, but also maintains a lightweight and efficient architecture, remains an important yet insufficiently explored problem.
To address these issues, this paper proposes a boundary-sensitive and structure-adaptive lightweight network for medical image segmentation. The proposed framework is developed around three problem-oriented designs. First, a Boundary-Sensitive Semantic Recalibration Attention (BSRA) module is introduced at the bottleneck stage to enhance deep semantic channels that are more informative for ambiguous boundary delineation and organ discrimination. Different from generic channel attention mechanisms, BSRA jointly models global semantic importance and boundary-sensitive local variation, thereby improving the network’s response to difficult regions with unclear anatomical transitions. Second, a Structure-Adaptive Scale Selection Fusion (SASF) module is designed to capture heterogeneous contextual dependencies through multiple receptive-field branches and to adaptively reweight their contributions according to structural characteristics. In this way, the model can better handle scale-sensitive regions and shape-varying structures. Third, to alleviate the detail degradation commonly caused by lightweight convolutions, a Detail-Compensated Residual Lightweight Block (DCRB) is constructed to preserve local continuity and boundary-sensitive information while reducing parameter redundancy and computational burden. Through the coordinated integration of these three components, the proposed method aims to improve segmentation quality from the perspective of boundary representation, scale adaptation, and practical efficiency, rather than merely pursuing global average metric gains.
Experiments were conducted on CT liver segmentation tasks to evaluate the proposed method from three complementary aspects: overall segmentation accuracy, boundary quality, and model efficiency. In addition to conventional overlap-based metrics such as Dice coefficient, Intersection over Union (IoU), Precision, and Recall, boundary-sensitive metrics including Boundary F1-score, 95% Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD) were adopted to provide a more comprehensive assessment of segmentation performance in clinically challenging regions. Model complexity indicators, including parameter count, floating-point operations (FLOPs), and inference time, were also reported to evaluate the lightweight advantage and practical trade-off of the proposed framework.
The main contributions of this paper can be summarized as follows:
- A Boundary-Sensitive Semantic Recalibration Attention (BSRA) module is proposed to enhance deep semantic channels that are more relevant to ambiguous boundary regions and semantically confusable anatomical structures.
- A Structure-Adaptive Scale Selection Fusion (SASF) module is designed to dynamically regulate the importance of different receptive-field branches, thereby improving the representation of scale-sensitive and structurally complex regions.
- A Detail-Compensated Residual Lightweight Block (DCRB) is developed to reduce model redundancy while preserving local continuity and boundary-sensitive detail information.
- A boundary-sensitive and structure-adaptive lightweight segmentation framework is constructed, which aims to improve not only overall segmentation accuracy but also boundary quality and practical efficiency in medical image segmentation tasks.
2 Related work
Deep learning has substantially advanced medical image segmentation in recent years, especially through the development of U-shaped encoder-decoder architectures, attention-based feature enhancement, multi-scale context modeling, and lightweight network design [1,2]. These research directions have provided effective solutions for improving segmentation accuracy under challenging conditions. However, for medical image segmentation tasks, especially organ segmentation in CT images, several practically important issues remain insufficiently addressed, including ambiguous boundaries, semantically confusable transition regions, heterogeneous scale demands, and the trade-off between segmentation performance and computational efficiency [5,6,10]. This section reviews representative studies from four aspects: U-shaped segmentation networks, boundary-aware and attention-guided segmentation, multi-scale context modeling, and lightweight medical image segmentation, thereby providing the technical foundation and research motivation for this work.
2.1 U-shaped medical image segmentation networks
U-shaped encoder-decoder networks have remained one of the dominant architectural paradigms in medical image segmentation [2]. Following the success of early fully convolutional designs, subsequent studies have improved this framework from different perspectives, including volumetric representation, recursive residual learning, stacked encoder-decoder design, channel-wise interaction, and hybrid CNN-transformer modeling. For example, V-Net extended the encoder-decoder formulation to volumetric medical image segmentation and demonstrated the effectiveness of end-to-end dense prediction in 3D settings [11]. R2U-Net incorporated recurrent residual units into the U-shaped framework to enhance feature reuse and representation capacity [12], while DoubleU-Net adopted a cascaded dual U-Net design to progressively refine segmentation results [13]. ResUNet++ introduced residual connections and multi-scale feature enhancement into the U-shaped backbone, showing that structural refinement of convolutional blocks can improve segmentation robustness [14].
More recently, research on U-shaped architectures has moved beyond conventional convolution-only designs and increasingly incorporated transformer-based or hybrid modeling strategies. Swin-U-Net, UCTransNet, UTNet, nnFormer, MISSFormer, and TransBTS demonstrated the value of long-range dependency modeling and stronger cross-scale interaction in medical segmentation [15–20]. However, these general-purpose or volumetric improvements do not directly resolve the low-contrast boundary ambiguity and efficiency constraints encountered in CT liver segmentation. This motivates a more problem-driven design that explicitly aligns attention, scale selection, and lightweight feature extraction with the practical bottlenecks of CT organ delineation.
2.2 Boundary-aware and attention-guided segmentation
Although many medical image segmentation models report strong overlap-based performance, boundary ambiguity remains a persistent challenge in practical applications. Recent studies have increasingly incorporated explicit boundary modeling into segmentation frameworks, including boundary-aware fusion, boundary supervision, gradient-based contour modeling, boundary filtering, uncertainty-aware perception, and joint boundary-global context learning [21–28]. These methods show that boundary-sensitive representation is important beyond conventional region-overlap optimization. A representative example is the end-to-end boundary-aware network of Hatamizadeh et al. [10], which introduces an explicit edge branch and edge-aware objectives; this method is therefore included in our main experimental comparison as a boundary-aware baseline. However, many boundary-aware methods rely on extra boundary supervision, explicit gradient maps, or task-specific auxiliary branches, which may increase training complexity or sensitivity to noisy contours. The proposed BSRA instead performs boundary-sensitive semantic recalibration at the bottleneck stage by jointly modeling global semantic responses and local structural variation, without requiring an additional boundary label.
2.3 Multi-scale context modeling
Another important research direction is multi-scale context modeling. This is particularly relevant in medical image segmentation because anatomical structures may vary substantially in shape, scale, and local topology across patients and image slices. Early studies on contextual representation mainly originated from general semantic segmentation. For example, DeepLabv3 + combined atrous spatial pyramid pooling with an encoder-decoder design to refine segmentation results [29], DenseASPP further enhanced receptive-field diversity through densely connected atrous convolutions [30], and FPN introduced a top-down feature pyramid fusion strategy that has strongly influenced later dense prediction models [31]. These studies established the general importance of multi-scale context aggregation for pixel-wise prediction.
In medical image segmentation, multi-scale context modeling has been further adapted to better capture anatomical structures with highly variable appearance. KiU-Net proposed an overcomplete branch to preserve fine-grained structures while complementing the conventional U-shaped pathway [32]. CPFNet further designed a context pyramid fusion strategy to combine global and multi-scale contextual information more effectively [33]. More recent studies continued to strengthen this line. For instance, a UNet-based multi-scale context fusion method explicitly assigned different weights to semantic features from different encoding stages [34], while a multi-scale cross-fusion medical image segmentation network based on a dual-attention transformer showed that multi-scale context fusion can also be enhanced by attention-guided interaction mechanisms [35].
2.4 Lightweight medical image segmentation
Besides segmentation accuracy, model efficiency has become increasingly important in medical image analysis. Many high-performing segmentation networks rely on deeper backbones, heavier feature fusion modules, or more complicated attention mechanisms, which may improve benchmark performance but also increase computational burden and deployment difficulty [36–40]. To alleviate this issue, lightweight medical image segmentation has gradually become an active research topic. For example, UNeXt reformulated the segmentation backbone with an MLP-based rapid architecture and showed that efficient design can still achieve competitive medical segmentation performance [36]. LCMUNet further combined CNN and MLP modules for real-time medical image segmentation, emphasizing the balance among accuracy, parameter size, and computational efficiency [37]. GA-UNet introduced a lightweight Ghost-and-Attention U-Net framework to improve feature extraction efficiency under a compact design [38], while LV-UNet explored a lightweight and vanilla segmentation architecture under a simplified design philosophy [39].
More recent studies have continued to strengthen this line. LUCF-Net proposed a lightweight U-shaped cascade fusion architecture for medical image segmentation [40]. Meanwhile, lightweight U-Net with depthwise separable convolution demonstrated that replacing standard convolutions with more efficient depthwise separable operations can substantially reduce computational redundancy while maintaining segmentation capability [41]. In addition, LIU-NET introduced a lightweight Inception U-Net framework for efficient brain tumor segmentation from multimodal 3D MRI images [42]. Although these methods are developed under different application scenarios, they consistently indicate that lightweight segmentation has evolved from simple model compression into a more systematic research direction, where architectural redesign is explicitly used to balance efficiency and segmentation quality.
In summary, prior studies have made substantial progress from three major directions: attention-guided feature enhancement, multi-scale context representation, and lightweight architectural optimization. Nevertheless, these directions are often explored independently, and their designs are not always explicitly aligned with three closely coupled issues that are especially important in CT organ segmentation: insufficient emphasis on ambiguous boundary regions, inadequate adaptation to heterogeneous scale demands, and local detail degradation caused by lightweight modeling. Motivated by these observations, this work aims to construct a problem-driven framework in which boundary-sensitive semantic recalibration, structure-adaptive scale selection, and detail-compensated lightweight feature extraction are integrated into a unified segmentation network.
3 Method
3.1 Overall architecture
Medical image segmentation, especially organ segmentation in CT images, faces three tightly coupled challenges: ambiguous anatomical boundaries, heterogeneous scale demands across different regions, and the degradation of local structural details under lightweight modeling. To address these issues in a unified framework, this paper proposes a Boundary-Sensitive and Structure-Adaptive Lightweight Segmentation Network, which is built upon a U-shaped encoder-decoder architecture and enhanced by three problem-driven modules, namely Boundary-Sensitive Semantic Recalibration Attention (BSRA), Structure-Adaptive Scale Selection Fusion (SASF), and Detail-Compensated Residual Lightweight Block (DCRB).
Different from conventional improvement strategies that mainly stack generic encoder-decoder refinements, recurrent residual units, or hybrid CNN-transformer modules for higher average overlap scores [12,13,17], the proposed framework is designed from the perspective of the actual bottlenecks in CT liver segmentation. Specifically, the initial image-level stem employs DCRB to extract shallow features, and the subsequent encoder and decoder stages continue to use DCRB as the primary lightweight feature extraction unit. DCRB is paired with MaxPool operations in the encoder and Upsample operations in the decoder to realize the resolution changes of the U-shaped architecture. At the bottleneck, BSRA recalibrates deep semantic channels for ambiguous boundary delineation, and SASF adaptively fuses multiple receptive-field responses according to structural characteristics.
The overall architecture of the proposed network is illustrated in Fig 1.
The model adopts a U-shaped encoder-decoder framework. The initial stem and the subsequent encoder and decoder stages are all implemented with DCRB as the primary lightweight feature extraction unit. In the encoder, DCRB is paired with MaxPool operations for progressive downsampling; in the decoder, Upsample operations, skip-connected feature fusion, and DCRB are used for spatial recovery. The intermediate encoder features are labeled as , and the deepest bottleneck feature is labeled as . At the bottleneck stage, BSRA first recalibrates with boundary sensitivity, and SASF then performs adaptive multi-scale feature fusion before decoding.
First, deep semantic features are indispensable for recognizing the target organ and suppressing irrelevant tissues. However, in challenging CT slices, high-level features may still be insufficiently sensitive to ambiguous boundaries and semantically confusable anatomical transitions. Therefore, the first step at the bottleneck is to recalibrate deep feature channels with BSRA, so that the network can explicitly emphasize more discriminative responses before further contextual modeling.
Second, accurate delineation of organ structures depends not only on semantic discrimination, but also on suitable contextual scales. Different organ regions may demand different receptive fields: local branches are more useful for fine boundary recovery, while larger receptive fields are more suitable for maintaining structural completeness and contextual consistency. For this reason, the recalibrated deep feature map is then fed into SASF, where multiple scale branches are constructed and adaptively fused to improve structure-aware context representation.
Third, although stronger feature extraction often benefits segmentation accuracy, the resulting model complexity may become undesirable for practical medical applications. Direct lightweight substitution, however, may also weaken local detail preservation and boundary continuity. Therefore, instead of using standard convolutions throughout the network, this work adopts DCRB as a detail-aware lightweight building block, which aims to reduce redundancy while compensating for the detail loss commonly introduced by lightweight convolutions.
In this way, BSRA, SASF, and DCRB correspond respectively to the three key issues emphasized in this work: boundary ambiguity, heterogeneous scale demand, and efficiency-detail trade-off. Their combination forms a problem-driven framework rather than a generic performance-oriented module stack.
In the following subsections, the feature flow of the proposed network is denoted in a stage-wise manner. Specifically, let denote the intermediate encoder features used for skip connections, and let denote the deepest encoder output at the bottleneck stage. For a feature map in , denotes the bottleneck channel depth, while and denote the spatial height and width of the feature map, respectively. After boundary-sensitive semantic recalibration, the refined bottleneck feature is denoted by . This refined representation is then fed into the structure-adaptive multi-scale fusion module, whose output is denoted by . Based on these definitions, the following subsections describe the proposed BSRA, SASF, and DCRB modules in detail.
3.2 Boundary-sensitive semantic recalibration attention
The proposed Boundary-Sensitive Semantic Recalibration Attention (BSRA) module is applied at the bottleneck stage to recalibrate deep semantic features before multi-scale context modeling. Given the deepest encoder feature , BSRA constructs two complementary descriptors, namely a semantic descriptor and a boundary-sensitive descriptor, and uses their fused representation to generate channel-wise attention weights. The refined feature is denoted by .
Instead of performing generic channel reweighting solely based on global semantic statistics, BSRA introduces an additional boundary-sensitive description process to encode local variation information from the bottleneck feature. In this way, the module jointly considers global semantic responses and structure-sensitive local fluctuations when estimating channel importance. The overall process of BSRA consists of three steps: semantic descriptor extraction, boundary-sensitive descriptor extraction, and descriptor fusion with channel recalibration. Through this design, the bottleneck feature can be refined before subsequent multi-scale fusion, yielding a more discriminative deep representation for later decoding.
3.2.1 Module structure.
The structure of BSRA is shown in Fig 2. Starting from the deep feature map, the module first extracts a semantic descriptor through global average pooling and a boundary-sensitive descriptor through local difference enhancement followed by global average pooling. These two C-dimensional descriptors are concatenated into a 2C-dimensional vector, mapped through 2C → C/r → C fully connected transformations, and normalized with a Sigmoid function to produce C channel weights. The resulting weights are used to recalibrate the original deep feature map channel by channel, yielding the refined output.
The channel-weighting path maps the fused 2C descriptor to C/r and then to C channels, followed by Sigmoid normalization.
3.2.2 Semantic descriptor branch.
The semantic descriptor branch models the global semantic importance of each channel. For the input feature the semantic descriptor is obtained by global average pooling:
or equivalently,
where denotes global average pooling. This descriptor captures the overall response strength of each channel in the deep representation.
3.2.3 Boundary-sensitive descriptor branch.
To encode local structural variation, BSRA further introduces a boundary-sensitive descriptor branch. First, a local average feature is computed by
where denotes local average pooling with kernel size 3 × 3, stride 1, and appropriate padding. Then, a local difference-enhanced feature is constructed as
where denotes element-wise absolute value. This operation emphasizes locally varying responses in the deep feature map. The boundary-sensitive descriptor is then obtained by global average pooling:
or
3.2.4 Descriptor fusion and channel recalibration.
After obtaining the semantic descriptor and the boundary-sensitive descriptor , the two descriptors are concatenated to form a joint representation:
The fused descriptor is passed through two fully connected transformations to produce channel attention scores:
where the learnable parameters are used in the two fully connected transformations and denotes the ReLU activation function. The final channel weights are generated by a sigmoid function:
where denotes the sigmoid function. The refined deep feature is then obtained by channel-wise recalibration:
or equivalently,
where ⊙ denotes channel-wise element-wise multiplication.
The output is subsequently used as the input of the SASF module for adaptive multi-scale context modeling.
3.3 Structure-adaptive scale selection fusion
The proposed Structure-Adaptive Scale Selection Fusion (SASF) module is applied after BSRA at the bottleneck stage. Given the BSRA-refined bottleneck feature , SASF performs parallel multi-scale feature extraction and adaptive scale reweighting before feature fusion. The output of SASF is denoted by .
Unlike conventional multi-scale modules that directly concatenate features from different receptive fields with fixed importance, SASF introduces a structure-adaptive fusion mechanism to regulate the contribution of each scale branch according to the current feature representation. In this way, the module does not treat all scales equally, but instead selects and emphasizes more relevant receptive-field responses for the input structure. The overall process of SASF consists of three steps: multi-scale branch construction, structure descriptor extraction with scale selection, and adaptive branch fusion. Through this design, the bottleneck feature can be enriched by context information from different receptive fields while preserving structural consistency in a residual manner.
3.3.1 Module structure.
The structure of SASF is shown in Fig 3. Starting from the input feature, the module first generates four scale-specific branches with different receptive fields. A structure descriptor is then extracted from the concatenated multi-scale representation and used to predict four scale-selection logits. Softmax is applied only across these four branch logits to obtain the scale-selection weights. These weights are applied to reweight different branches before fusion. The fused feature is further refined and added to the input feature through a residual connection, producing the final bottleneck output.
3.3.2 Multi-scale branch construction.
Given the input feature , SASF first constructs four parallel branches to model structural information under different receptive fields. Denoting the output of the k-th branch by , the branch features are defined as
where the first branch is implemented by a 1×1 convolution, and the other three branches are implemented by 3×3 convolutions with dilation rates of 1, 2, and 4, respectively. These dilation rates provide a compact receptive-field set that covers local boundary responses, medium-range structural transitions, and broader contextual cues within the liver region.
where denotes the channel number of each branch. In the proposed architecture, C is fixed at 512 at the bottleneck, so for each scale branch.
3.3.3 Structure descriptor and scale selection.
After branch construction, the four branch features are concatenated along the channel dimension:
A global structure descriptor is then extracted by global average pooling:
where denotes global average pooling.
The structure descriptor is further transformed into scale-selection logits through two fully connected mappings:
where the two fully connected layers transform the structure descriptor into scale-selection logits, denotes the ReLU activation function, and the output dimension of the second fully connected layer is four, corresponding to the four scale branches. The scale-selection weights are then obtained through a Softmax operation:
where
Each coefficient represents the relative importance of the corresponding receptive-field branch for the current input feature.
3.3.4 Adaptive branch fusion.
After obtaining the scale-selection weights, SASF adaptively reweights each branch as
The reweighted branch features are concatenated and fused by a 1 × 1 convolution:
where denotes the fusion operation composed of 1 × 1 convolution, batch normalization, and ReLU activation.
To preserve the original bottleneck semantics, the fused feature is added to the input feature through a residual formulation:
Finally, a refinement operation is applied to obtain the SASF output:
where denotes the refinement block after residual fusion. In this implementation, the refinement block consists of a 3 × 3 convolution, batch normalization, and ReLU activation, which further smooths and stabilizes the fused bottleneck representation before decoding. The resulting feature is then forwarded to the decoder for subsequent spatial recovery.
3.4 Detail-compensated residual lightweight block
The proposed Detail-Compensated Residual Lightweight Block (DCRB) is used as the basic feature extraction unit throughout both the encoder and decoder. Given an input feature map , DCRB performs lightweight feature extraction while preserving local structural continuity through an additional detail compensation branch. The output feature is denoted by .
Instead of directly replacing standard convolutions with depthwise separable convolutions in a plain manner, DCRB introduces a dual-path formulation composed of a lightweight main branch and a detail compensation branch. The main branch is responsible for efficient feature transformation, while the compensation branch explicitly enhances local variation information that may be weakened during lightweight convolution. The overall process of DCRB consists of three steps: lightweight feature extraction in the main branch, local detail compensation in the auxiliary branch, and residual aggregation with shortcut mapping. Through this design, DCRB reduces computational redundancy while maintaining feature continuity and boundary-sensitive detail representation.
3.4.1 Module structure.
The structure of DCRB is illustrated in Fig 4. Given the input feature X, the module first extracts an efficient representation through a lightweight main branch composed of two consecutive depthwise-separable convolution stages. Each stage explicitly follows the sequence 3 × 3 depthwise convolution → batch normalization → ReLU → 1 × 1 pointwise convolution → batch normalization → ReLU. In parallel, a detail-compensation branch models local variation information from the same input feature, while a shortcut branch preserves identity information and stabilizes feature propagation. The outputs of the three branches are then aggregated and followed by a nonlinear activation to produce the final output.
In the main branch, each depthwise-separable stage follows depthwise 3 × 3 convolution → BN → ReLU → pointwise 1 × 1 convolution → BN → ReLU, and this stage is repeated twice before residual aggregation.
3.4.2 Main lightweight branch.
Given the input feature , the main branch performs lightweight feature extraction by using two consecutive depthwise separable convolutions. Let and denote the first and second depthwise separable convolutional transformations, respectively. The main branch output is defined as
Here, each depthwise-separable convolution stage is implemented in the explicit order of a 3 × 3 depthwise spatial convolution, batch normalization, ReLU activation, a pointwise 1 × 1 convolution, batch normalization, and ReLU activation. Two such stages are applied consecutively in the main branch. Compared with standard convolution, this formulation reduces parameter redundancy and computational cost while retaining spatial filtering and channel mixing.
3.4.3 Detail compensation branch.
To complement the efficient but potentially detail-insensitive transformation of the main lightweight branch, DCRB further introduces a detail compensation branch. Starting from the input feature X, a local average response is first computed as
where denotes local average pooling with kernel size 3 × 3, stride 1, and appropriate padding. A local variation feature is then defined by
where denotes element-wise absolute value. This operation enhances local structural fluctuation information and suppresses spatially smooth responses.
The variation-enhanced feature is then projected into the target channel space through a lightweight transformation , yielding the detail compensation output
where the lightweight transformation is implemented by a 1×1 convolution when channel alignment is required; otherwise, an identity mapping is used when the input and output channel dimensions are already matched. This criterion keeps the compensation branch lightweight and deterministic for reproduction.
3.4.4 Shortcut mapping and residual aggregation.
To preserve identity information and stabilize optimization, DCRB adopts a shortcut branch. If the input and output channel dimensions are identical, the shortcut mapping is defined as an identity function:
Otherwise, a 1 × 1 convolution is used to align the channel dimension:
where denotes the shortcut projection.
After obtaining the three branch outputs, DCRB aggregates them in a residual manner:
Finally, a nonlinear activation function is applied to produce the output feature:
where the activation function denotes ReLU.
This formulation enables DCRB to jointly exploit efficient feature extraction, local detail compensation, and identity-preserving residual propagation within a single lightweight block.
In the proposed framework, DCRB is used as the fundamental convolutional unit in both the encoder and decoder. For encoder stages, DCRB is combined with downsampling operations to progressively enlarge the receptive field while maintaining efficient feature extraction. For decoder stages, DCRB is used after upsampling and skip-connected fusion to refine recovered spatial representations. By employing a unified lightweight block across the whole network, the proposed framework maintains architectural consistency while reducing the computational burden associated with standard convolution-based U-shaped models.
3.5 Loss function and optimization objective
The proposed network is trained in an end-to-end manner using a hybrid objective that combines a pixel-wise binary classification term and a region-overlap term. Specifically, a binary cross-entropy with logits loss is used to supervise pixel-level classification, while a Dice loss is introduced to directly optimize the overlap between the predicted mask and the ground-truth region. The overall training objective is formulated as a weighted combination of these two terms.
3.5.1 Binary cross-entropy loss.
Let denote the predicted logit map, and let denote the corresponding ground-truth mask. After sigmoid activation, the predicted probability map is written as
where denotes the sigmoid function. The binary cross-entropy loss is defined as
This term enforces pixel-wise consistency between the predicted probability and the corresponding ground-truth label.
3.5.2 Dice loss.
Since medical image segmentation often involves strong foreground-background imbalance, especially in organ delineation tasks, overlap-oriented supervision is further introduced through Dice loss. The Dice coefficient between the predicted probability map and the ground-truth mask is defined as
Compared with purely pixel-wise losses, Dice loss directly optimizes region overlap and is therefore more suitable for dense medical segmentation tasks.
3.5.3 Hybrid optimization objective.
The final training loss is formulated as
where and are the weighting coefficients of the two loss terms. In this work, a Dice-oriented weighting strategy is adopted, and the coefficients are set as
so that the optimization process emphasizes region-level overlap while still preserving stable pixel-wise supervision. This setting is consistent with the current training implementation of the model.
3.5.4 Optimization strategy.
Let the overall network parameters be denoted by Θ. The training objective is therefore written as
During optimization, the model parameters are updated by minimizing the hybrid loss over the training set. All models use the Adam optimizer with an initial learning rate of 3 × 10-4 and a weight decay of 3 × 10-5. The maximum training length is 100 epochs. ReduceLROnPlateau monitors validation Dice in maximization mode, with a reduction factor of 0.5 and a patience of 5 epochs. Early stopping is applied with a patience of 14 epochs. A global random seed of 42 is fixed for reproducibility, and gradient clipping is applied with a maximum norm of 1.0.
3.5.5 Inference and evaluation protocol.
At inference time, the network outputs a probability map , which is converted into a binary segmentation mask by thresholding:
where denotes the decision threshold. The threshold is selected exclusively on the internal validation set from {0.35, 0.40, 0.45, 0.50, 0.55, 0.60} according to validation performance and is then applied to the corresponding independent test fold. The largest connected component post-processing strategy is retained to suppress isolated false-positive regions. Final evaluation includes overlap-based metrics, Boundary F1-score, HD95, ASSD, parameter count, FLOPs, and inference time.
4 Experiments and results
To validate the proposed framework in a problem-oriented manner, experiments are designed from multiple aspects, including module effectiveness, comparative performance, attention mechanism selection, validation on an additional CT dataset, qualitative boundary analysis, and efficiency–accuracy trade-off evaluation. Specifically, the ablation study combines the progressive integration path with isolated controls to examine whether BSRA, SASF, and the detail-compensation mechanism contribute to boundary-sensitive representation, structure-adaptive multi-scale modeling, and lightweight detail-preserving feature extraction, respectively. Comparisons with representative segmentation methods, including a boundary-aware baseline, are further conducted to assess the competitiveness of the proposed model. In addition, several commonly used attention mechanisms are evaluated under the same setting to justify the selection of BSRA. Finally, qualitative results and efficiency-related comparisons are presented to provide a more comprehensive understanding of the proposed method beyond overlap-based accuracy metrics.
4.1 Experimental settings
4.1.1 Dataset and data split.
The experiments were conducted on the CT subset of the CHAOS dataset [43], which is a widely used benchmark for abdominal medical image analysis. Liver segmentation was selected as the target task. The original CT volumes and their corresponding binary liver masks were converted into 2D image-label pairs for network training and evaluation. To avoid information leakage, all data partitions were performed at the patient/case level rather than at the slice level, so all slices from the same CT case were assigned to the same subset.
Patient-level five-fold cross-validation was adopted for the 20 CHAOS CT cases. In each fold, four patients were held out as the independent test set. Among the remaining 16 patients, 12 were used for training and four were used as the internal validation set. The training-validation subdivision was performed strictly at the patient level using the fixed random seed of 42, and all slices from the same patient were kept in the same subset. Thus, each fold comprised 12 training cases, four validation cases, and four independent test cases. All input slices were resized to 256 × 256 before being fed into the network.
All compared methods were evaluated under the same data split, preprocessing pipeline, training budget, threshold-selection strategy, and post-processing strategy. This design ensures that performance differences mainly reflect architectural effects rather than differences in data partition or evaluation protocol.
4.1.2 Implementation details.
All models were implemented in PyTorch 2.1.1 and trained and evaluated on a workstation running Windows 10, equipped with an NVIDIA Quadro RTX 6000 GPU with 24 GB memory and an Intel Xeon Gold 6238R CPU at 2.20 GHz. CUDA 11.8 and cuDNN 8.7.0 were used. All compared models followed the same data split, training configuration, threshold-selection procedure, and post-processing pipeline to ensure a fair comparison.
Segmentation metrics and inference time were summarized at the fold level. Within each fold, each metric was first averaged across all test cases assigned to that fold, and the final result is reported as the mean ± standard deviation of the five fold-wise averages. Parameters and FLOPs were recalculated under the same implementation environment; FLOPs were computed using a fixed input size of 1 × 1 × 256 × 256 and the same profiling procedure for all models. For inference-time measurement, 50 forward passes were first performed as GPU warm-up. Timing was then repeated for 100 forward passes with CUDA synchronization around each measured inference, and the average elapsed time was used as the per-image inference time for that fold.
All models were trained with a unified hyperparameter configuration. The input size was fixed at 256 × 256 and the batch size was 8. The global random seed was set to 42. Adam optimization used an initial learning rate of 3 × 10-4 and a weight decay of 3 × 10-5, with a maximum of 100 epochs. ReduceLROnPlateau monitored validation Dice in maximization mode with a reduction factor of 0.5 and a patience of 5 epochs; early stopping used a patience of 14 epochs. Training data augmentation included random horizontal flipping, random vertical flipping, and random rotations by multiples of 90°. Gradient clipping was applied with a maximum norm of 1.0. The weighted BCE-Dice objective used weights of 0.3 and 0.7, respectively.
During validation, the binarization threshold was selected from {0.35, 0.40, 0.45, 0.50, 0.55, 0.60} according to validation performance. The selected threshold was then fixed for evaluation of the corresponding independent test fold. Largest connected component post-processing was used to suppress isolated false-positive regions and improve the stability of organ mask generation.
4.1.3 Evaluation metrics and statistical analysis.
To comprehensively assess the proposed method, the experiments considered both segmentation performance and model efficiency.
For segmentation performance, conventional overlap-based metrics were adopted, including Dice coefficient, Intersection over Union (IoU), Precision, and Recall. These metrics evaluate the overall agreement between the predicted region and the ground-truth mask. Higher Dice, IoU, Precision, and Recall values indicate better segmentation performance.
Since this work particularly focuses on boundary-sensitive representation, Boundary F1-score was introduced to evaluate contour alignment quality. In addition, 95% Hausdorff Distance (HD95) and Average Symmetric Surface Distance (ASSD) were included as boundary-distance metrics. Unlike the overlap-based metrics, lower HD95 and ASSD values indicate better boundary agreement.
For model efficiency, parameter count (Params), floating-point operations (FLOPs), and inference time per image were reported. Segmentation metrics and inference time are shown as mean ± standard deviation. Params and FLOPs are deterministic model-complexity measures calculated under the same input size and profiling tool and are therefore reported without standard deviations.
To strengthen statistical rigor, significance analysis was performed at the patient level rather than on the five fold-specific summary values. For CHAOS, the Wilcoxon signed-rank test used paired out-of-fold predictions from all 20 independent patients; for LiTS, the corresponding test used all 80 out-of-fold patients. A p-value < 0.05 was considered statistically significant. Dice and Boundary F1 p-values for the principal progressive ablation variants versus the Baseline are reported in the corresponding results text. The two isolated-control experiments are summarized using fold-level mean ± standard deviation; no pairwise significance test was performed for these newly introduced controls.
4.1.4 Experimental protocol.
Based on the above setting, the experimental study is organized into several parts. First, the ablation study combines the progressive path with two isolated controls: a SASF-only variant and a depthwise-separable residual block without the detail-compensation branch. Second, the proposed method is compared with representative segmentation models, a boundary-aware method, and recent lightweight methods, including GA-UNet, LCMUNet, and LUCF-Net. Third, different attention mechanisms are compared under the same bottleneck setting to validate the rationale of BSRA. Finally, additional validation on LiTS evaluates whether the same architectural trends can be reproduced on a second liver CT dataset under the same patient-level five-fold protocol.
4.2 Ablation study
4.2.1 Main quantitative comparison of ablation variants.
To disentangle the contributions of the proposed components, the ablation study combines progressive integration with isolated controls. Six variants were considered: the Baseline, Base+BSRA, Base+SASF, Base+BSRA+SASF (the performance-oriented configuration, Ours-Perf), Base+BSRA+SASF with a depthwise-separable residual block without the detail-compensation branch (DSRB), and Base+BSRA+SASF+DCRB (the lightweight configuration, Ours-Lite). Table 1 reports all six variants, whereas Figs 5 and 6 visualize the four core operating points along the progressive path for readability. This design allows the individual contribution of SASF and the specific contribution of the DCRB detail-compensation branch to be assessed directly.
Table 1 provides the quantitative comparison among the six ablation variants. After introducing BSRA, Dice increased from 0.8443 ± 0.0121 to 0.8492 ± 0.0143 and Boundary F1-score increased from 0.7728 ± 0.0114 to 0.8081 ± 0.0152. The paired case-level Wilcoxon test shows that the Dice change is not statistically significant (p = 0.1362), whereas the Boundary F1 improvement is significant (p = 0.0241). This pattern directly supports the intended role of BSRA: its benefit is clearer in boundary-sensitive representation than in global region overlap.
The SASF-only control achieves Dice = 0.8881 ± 0.0195, Boundary F1-score = 0.8247 ± 0.0178, HD95 = 6.9827 ± 0.6210, and ASSD = 2.1458 ± 0.0442, showing that SASF provides a strong independent contribution even without BSRA. When BSRA and SASF are combined, the performance-oriented configuration further improves to Dice = 0.8929 ± 0.0216, Boundary F1-score = 0.8466 ± 0.0195, HD95 = 6.5729 ± 0.6841, and ASSD = 1.8947 ± 0.0482. Compared with SASF alone, the combined configuration improves the mean Dice by 0.0048 and Boundary F1 by 0.0219 while reducing HD95 by 0.4098 and ASSD by 0.2511. These results indicate that SASF accounts for most of the region-level gain, while BSRA provides additional boundary-oriented refinement when the two modules are used together.
To isolate the contribution of the detail-compensation branch, we additionally evaluated a DSRB variant that retains the same depthwise-separable residual main path but removes detail compensation. This control obtains Dice = 0.8598 ± 0.0159, Boundary F1-score = 0.8204 ± 0.0171, HD95 = 7.3917 ± 0.4892, and ASSD = 3.2841 ± 0.0389 with 2.21M parameters and 8.27G FLOPs. Restoring the detail-compensation branch to form the full DCRB increases the mean Dice to 0.8710 ± 0.0165 and Boundary F1-score to 0.8392 ± 0.0178, while reducing HD95 to 7.1391 ± 0.5134 and ASSD to 3.0425 ± 0.0406. The corresponding complexity increases only modestly to 2.47M parameters and 8.44G FLOPs, with inference time changing from 6.1038 ± 0.0298 ms to 6.2147 ± 0.0314 ms. Thus, the isolated control supports the contribution of the detail-compensation mechanism to recovering segmentation and boundary quality after lightweight substitution. Relative to Ours-Perf, Ours-Lite remains an efficiency-oriented operating point rather than the peak-accuracy configuration.
4.2.2 Integrated quantitative interpretation.
While Table 1 reports all six ablation variants, Fig 5 provides a visual summary of the four core progressive operating points to show how the main gains are distributed across segmentation quality, boundary-specific metrics, and model complexity. The isolated SASF-only and DSRB controls are retained in Table 1 because their purpose is to disentangle module-specific effects rather than to define additional deployment operating points. The figure emphasizes Dice, Boundary F1-score, HD95, ASSD, Params, FLOPs, and inference time.
First, Fig 5 confirms that BSRA yields a clearer improvement in Boundary F1-score than in Dice, which is highly consistent with its intended role as a boundary-sensitive semantic recalibration module. This result suggests that BSRA mainly acts on difficult transition regions and blurred contours, where subtle local structural variation is more important than homogeneous interior prediction.
Second, Fig 5 highlights that SASF contributes the largest overall performance gain. Compared with the BSRA-only variant, adding SASF substantially improves Dice, IoU, Precision, Boundary F1-score, HD95, and ASSD, showing that structure-adaptive scale selection is beneficial for handling heterogeneous anatomical structures in CT liver segmentation.
Third, Fig 5 makes the role of the lightweight variant more explicit. Although Ours-Lite does not outperform the performance-oriented BSRA+SASF variant in peak accuracy, it preserves competitive Dice and Boundary F1-score while dramatically reducing parameter count and FLOPs. This supports the deployment-oriented motivation of DCRB, while also revealing a measurable accuracy-efficiency trade-off.
4.2.3 Trade-off analysis of performance and efficiency.
Beyond absolute segmentation accuracy, Fig 6 further illustrates the practical trade-off among segmentation quality, boundary quality, and efficiency. This figure is especially important because the final goal of the proposed framework is not merely to achieve the highest Dice score, but to provide a reasonable balance among segmentation accuracy, boundary fidelity, model compactness, and inference efficiency.
From the accuracy-complexity trade-off view, Ours-Perf occupies the highest region in Dice and Boundary F1-score, indicating that it is the strongest performance-oriented configuration. In contrast, Ours-Lite shifts toward the low-parameter and low-FLOPs region while still maintaining relatively high Dice and Boundary F1-score.
A similar trend is observed in the boundary quality-complexity trade-off analysis. Ours-Perf achieves the best Boundary F1-score and the lowest HD95 and ASSD on CHAOS. Meanwhile, Ours-Lite remains competitive in boundary quality despite its substantially reduced model size. However, its ASSD is not always better than the baseline or lightweight competitors, indicating that the lightweight design improves the overall compactness and Boundary F1-score but may introduce local distance-based trade-offs in some cases.
The speed-related results show that lower FLOPs do not necessarily translate into proportionally lower latency because practical runtime is also affected by GPU kernels, depthwise-convolution implementation, memory access, and software-level optimization. In the current implementation, Ours-Lite reduces FLOPs from 31.81G to 8.44G relative to Ours-Perf and achieves a lower measured latency (6.2147 ± 0.0314 ms versus 12.4774 ± 0.1063 ms). Inference time is therefore treated as a hardware- and implementation-dependent auxiliary measure, while Params and FLOPs provide the primary evidence for model compactness.
4.3 Comparison with representative methods
To further evaluate the proposed framework, we compared it with representative medical image segmentation methods, a representative boundary-aware segmentation model, and recent lightweight segmentation architectures under the same CHAOS CT protocol. The general methods include U-Net [2], Attention U-Net [44], MedNeXt [45], U-Mamba [46], and nnU-Net Revisited [47]. Boundary-Aware Net [10] was included as a direct boundary-focused reference. Three recent lightweight methods, namely GA-UNet [38], LCMUNet [37], and LUCF-Net [40], were also included. All methods were evaluated using the same metrics, including Dice, IoU, Precision, Recall, Boundary F1-score, HD95, ASSD, Params, FLOPs, and inference time. Ours-Perf denotes Base+BSRA+SASF and is used as the performance-oriented configuration, whereas Ours-Lite denotes Base+BSRA+SASF+DCRB and is used as the compact configuration.
As shown in Table 2, Ours-Perf attains the highest mean Dice (0.8929 ± 0.0216) and Boundary F1-score (0.8466 ± 0.0195), together with the lowest mean HD95 (6.5729 ± 0.6841) and ASSD (1.8947 ± 0.0482) in the current comparison. Boundary-Aware Net [10] achieves Dice = 0.8692 ± 0.0148, Boundary F1-score = 0.8256 ± 0.0143, HD95 = 6.9634 ± 0.4027, and ASSD = 1.9352 ± 0.0318, providing a direct boundary-oriented reference. Ours-Lite uses only 2.47M parameters and 8.44G FLOPs while maintaining Dice = 0.8710 ± 0.0165 and Boundary F1-score = 0.8392 ± 0.0178; however, its HD95 and ASSD are not uniformly better than the boundary-aware or other strong baselines. Among the lightweight competitors, LCMUNet retains lower FLOPs (3.49G) and lower measured latency (5.2700 ± 0.0196 ms), while Ours-Lite has the smallest parameter count in the current comparison. Because pairwise significance tests were not performed against the competing architectures, these differences are interpreted descriptively at the level of mean performance. Overall, Ours-Perf provides the stronger segmentation result, whereas Ours-Lite provides a smaller model with competitive boundary performance.
4.4 Comparison of different attention mechanisms
Since attention enhancement is one of the core components of the proposed framework, it is necessary to verify that the effectiveness of BSRA does not simply come from inserting an arbitrary attention module. To this end, different attention mechanisms were compared under the same bottleneck setting, including SE, CBAM, ECA, CoordAtt, and the proposed BSRA.
For a fair comparison, the attention mechanism is inserted at the same bottleneck position in the segmentation backbone, while all other settings, including the data split, training protocol, and evaluation strategy, are kept unchanged. The compared attention mechanisms include SE [48], CBAM [49], ECA [50], and CoordAtt [51], together with the proposed BSRA. Among them, SE is a representative channel-attention method, CBAM combines channel and spatial attention, ECA is a lightweight efficient channel-attention module, and CoordAtt introduces coordinate-aware feature encoding.
As shown in Table 3, BSRA is not necessarily the best module for maximizing global Dice alone; CBAM obtains a slightly higher Dice score (0.8511 ± 0.0113 versus 0.8492 ± 0.0143). However, BSRA achieves the best Boundary F1-score (0.8081 ± 0.0152), the lowest HD95 (7.5518 ± 0.4273), and the lowest ASSD (2.9624 ± 0.0315) among the compared attention modules. This pattern is more aligned with the boundary-sensitive motivation of the proposed module than a marginal gain in global overlap.
Therefore, the attention comparison provides direct evidence that the proposed boundary-sensitive descriptor contributes to boundary-aware feature calibration. The gain in boundary-related metrics is more relevant to the central claim of this paper than a marginal difference in Dice.
4.5 Additional validation on the LiTS dataset
To evaluate whether the architectural trends observed on CHAOS can be reproduced on another liver CT dataset, additional experiments were conducted on LiTS [52]. Because the network was retrained and evaluated on LiTS using patient-level five-fold cross-validation rather than being trained on CHAOS and directly transferred to LiTS, this experiment is interpreted as validation on an additional dataset rather than cross-dataset generalization. The purpose is therefore to examine whether the relative effects of BSRA, SASF, and DCRB remain consistent under a second liver CT data distribution.
For the LiTS liver-segmentation evaluation, 80 patient cases were included, yielding 4,520 valid two-dimensional CT slices after preprocessing. Patient-level five-fold cross-validation was adopted. In each fold, 16 patients were held out as the independent test set. The remaining 64 patients were further divided at the patient level into 51 training cases and 13 internal validation cases using the same fixed random seed of 42. All slices from a given patient were kept within a single subset to prevent information leakage. Across the five folds, every patient served as an out-of-fold test case exactly once. The comparison focused on the Baseline and the three proposed ablation variants to examine whether BSRA, SASF, and DCRB maintain similar effectiveness trends across datasets.
As shown in Table 4, compared with the Baseline, the BSRA variant increases Dice from 0.8223 ± 0.0137 to 0.8301 ± 0.0152 and Boundary F1-score from 0.7398 ± 0.0132 to 0.7709 ± 0.0164. The Dice change is not significant (p = 0.1517), whereas the Boundary F1 improvement is significant (p = 0.0286). The BSRA+SASF variant achieves the best mean performance on LiTS, with Dice = 0.8751 ± 0.0231 and Boundary F1-score = 0.8245 ± 0.0207, and both improvements are statistically significant versus Baseline (p = 0.0008 and p = 0.0004, respectively). The lightweight variant retains Dice = 0.8578 ± 0.0179 and Boundary F1-score = 0.8109 ± 0.0189 while reducing FLOPs to 8.44G; its Dice and Boundary F1 improvements versus Baseline are also significant (p = 0.0085 and p = 0.0053). These results reproduce the same overall module trend on an additional liver CT dataset, without implying direct cross-dataset transfer or zero-shot generalization.
4.6 Qualitative analysis
To further compare the visual behaviors of different variants in anatomically challenging regions, two representative independent test slices are selected for qualitative analysis, with emphasis on predicted masks, boundary overlays, and error-difference maps.
As shown in Fig 7, the qualitative comparison focuses on the three proposed ablation configurations shown in the figure. The BSRA-only variant captures the main liver region but still exhibits local contour deviations along curved peripheral boundaries and thin transition areas. These localized differences are consistent with the fact that BSRA is primarily intended to improve boundary-sensitive recalibration rather than to maximize region-overlap performance by itself.
The BSRA+SASF configuration shows the strongest contour consistency in the representative cases, with its predicted boundary generally following the ground-truth contour more closely. This visual pattern is consistent with the fold-level quantitative results, where Ours-Perf obtains the highest mean Dice and Boundary F1-score and the lowest mean HD95 and ASSD among the core ablation configurations.
The BSRA+SASF+DCRB configuration remains visually close to the ground truth in most regions but can show localized boundary deviations relative to Ours-Perf. This is consistent with its intended role as Ours-Lite: a compact efficiency-oriented operating point that gives up some peak segmentation accuracy in exchange for a much smaller model and lower computational cost. The error-difference maps are interpreted only as relative changes in error locations between variants, rather than as independent evidence of statistical superiority.
Overall, the qualitative observations complement the quantitative results: SASF provides the clearest visual improvement after BSRA, whereas DCRB preserves the global liver structure under a substantially lighter configuration while allowing limited local boundary trade-offs. The figure is therefore used as qualitative support for the performance–efficiency interpretation rather than as a substitute for the fold-level and patient-level statistical analyses.
5 Conclusion
This paper addressed a practical limitation in medical CT liver segmentation: a model may achieve acceptable region-overlap accuracy while still failing to delineate reliable anatomical boundaries, especially in weak-contrast and structurally complex regions. To address this issue, we proposed a boundary-sensitive and structure-adaptive lightweight segmentation framework that integrates BSRA, SASF, and DCRB in a U-shaped architecture.
Experimental results demonstrated that the proposed framework improves both region-level segmentation and boundary-level structural fidelity. In particular, the gain introduced by BSRA is clearer in Boundary F1-score than in Dice; on CHAOS, the Boundary F1 improvement is statistically significant (p = 0.0241) whereas the Dice change is not (p = 0.1362). The SASF-only control confirms that SASF provides a strong independent contribution, while the combined BSRA+SASF configuration gives the best mean overall performance. The DSRB control further shows that restoring the detail-compensation branch improves mean Dice, Boundary F1, HD95, and ASSD with only a modest increase in Params, FLOPs, and measured latency. DCRB therefore serves primarily as an efficiency-oriented design that substantially reduces model size and computation relative to Ours-Perf while preserving competitive segmentation quality, rather than as the peak-accuracy configuration.
Overall, the proposed framework provides two complementary operating points: Ours-Perf (BSRA+SASF) emphasizes segmentation and boundary quality, whereas Ours-Lite (BSRA+SASF+DCRB) emphasizes compactness while retaining competitive performance. The LiTS experiment reproduces the same overall architectural trend on an additional liver CT dataset, but it should be interpreted as additional-dataset validation rather than direct cross-dataset generalization. Several limitations should be acknowledged. First, CHAOS contains only 20 patients, so statistical conclusions should remain cautious despite patient-level five-fold cross-validation and paired case-level testing. Second, the current study focuses on liver segmentation, and the observed benefits should not be assumed to transfer automatically to other organs or imaging modalities. Third, measured inference time is hardware- and implementation-dependent; Params and FLOPs indicate compactness but do not establish universal runtime superiority. Future work will evaluate larger cohorts, additional organs and modalities, direct cross-dataset transfer, and further inference optimization.
References
- 1. Xia Q, Zheng H, Zou H, Luo D, Tang H, et al. A comprehensive review of deep learning for medical image segmentation. Neurocomputing. 2025;613.
- 2.
Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Medical Image Computing and Computer-Assisted Intervention, 2015. 234–41.
- 3. Okada T, Linguraru MG, Hori M, Summers RM, Tomiyama N, Sato Y. Abdominal multi-organ segmentation from CT images using conditional shape-location and unsupervised intensity priors. Med Image Anal. 2015;26(1):1–18. pmid:26277022
- 4. Gibson E, Giganti F, Hu Y, Bonmati E, Bandula S, Gurusamy K, et al. Automatic Multi-Organ Segmentation on Abdominal CT With Dense V-Networks. IEEE Trans Med Imaging. 2018;37(8):1822–34. pmid:29994628
- 5. Li X, Chen H, Qi X, Dou Q, Fu C-W, Heng P-A. H-DenseUNet: Hybrid Densely Connected UNet for Liver and Tumor Segmentation From CT Volumes. IEEE Trans Med Imaging. 2018;37(12):2663–74. pmid:29994201
- 6. Jin Q, Meng Z, Sun C, Cui H, Su R. RA-UNet: A Hybrid Deep Attention-Aware Network to Extract Liver and Tumor in CT Scans. Front Bioeng Biotechnol. 2020;8:605132. pmid:33425871
- 7. Tagnamas J, Ramadan H, Yahyaouy A, Tairi H. CS-Net: combined ConvNeXt-Swin-Unet for accurate medical image segmentation. J Supercomput. 2026;82(5).
- 8. Xia Q, Ren Y, Xiao H, Liu S, Wang Y, Bian G-B. ET-TransUNet: Integrating EMs-RstTrans with triple attention for robust noise suppression and feature enhancement in medical image segmentation. Digital Signal Processing. 2026;176:106064.
- 9. Sajjanar R, Dixit UD. Enhanced Segmentation of High-Grade and Low-Grade Brain Tumors Using Advanced 3D U-Net++ with Hybrid Lion-Spider Monkey Optimization. IJE. 2025;38(12):2819–33.
- 10. Hatamizadeh A, Terzopoulos D, Myronenko A. End-to-End Boundary Aware Networks for Medical Image Segmentation. Lecture Notes in Computer Science. Springer International Publishing. 2019. p. 187–94.
- 11.
Milletari F, Navab N, Ahmadi S-A. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV), 2016. 565–71. https://doi.org/10.1109/3dv.2016.79
- 12. Alom MZ, Yakopcic C, Hasan M, Taha TM, Asari VK. Recurrent residual U-Net for medical image segmentation. J Med Imaging (Bellingham). 2019;6(1):014006. pmid:30944843
- 13.
Jha D, Riegler MA, Johansen D, Halvorsen P, Johansen HD. DoubleU-Net: A Deep Convolutional Neural Network for Medical Image Segmentation. In: 2020 IEEE 33rd International Symposium on Computer-Based Medical Systems (CBMS), 2020. 558–64. https://doi.org/10.1109/cbms49503.2020.00111
- 14.
Jha D, Smedsrud PH, Riegler MA, Johansen D, Lange TD, Halvorsen P, et al. ResUNet++: An Advanced Architecture for Medical Image Segmentation. In: 2019 IEEE International Symposium on Multimedia (ISM), 2019. 225–2255. https://doi.org/10.1109/ism46123.2019.00049
- 15.
Cao H, Wang Y, Chen J, Jiang D, Zhang X, Tian Q, et al. Swin-Unet: Unet-Like Pure Transformer for Medical Image Segmentation. COMPUTER VISION - ECCV 2022 WORKSHOPS. 2022. 205–18.
- 16. Wang H, Cao P, Wang J, Zaiane OR. UCTransNet: Rethinking the Skip Connections in U-Net from a Channel-Wise Perspective with Transformer. AAAI. 2022;36(3):2441–9.
- 17.
Gao Y, Zhou M, Metaxas DN. UTNet: A Hybrid Transformer Architecture for Medical Image Segmentation. MEDICAL IMAGE COMPUTING AND COMPUTER ASSISTED INTERVENTION - MICCAI 2021, PT III 12903. 2021. 61–71.
- 18. Zhou H-Y, Guo J, Zhang Y, Han X, Yu L, Wang L, et al. nnFormer: Volumetric Medical Image Segmentation via a 3D Transformer. IEEE Trans Image Process. 2023;32:4036–45. pmid:37440404
- 19. Huang X, Deng Z, Li D, Yuan X, Fu Y. MISSFormer: An Effective Transformer for 2D Medical Image Segmentation. IEEE Trans Med Imaging. 2023;42(5):1484–94. pmid:37015444
- 20.
Wang W, Chen C, Ding M, Yu H, Zha S. TransBTS: Multimodal Brain Tumor Segmentation Using Transformer. In: Medical Image Computing and Computer Assisted Intervention - MICCAI 2021, 2021. 109–19.
- 21. Huang Z, Xie F, Qing W, Wang M, Liu M, Sun D. MGF-net: Multi-channel group fusion enhancing boundary attention for polyp segmentation. Med Phys. 2024;51(1):407–18. pmid:37403578
- 22. Rezvani S, Fateh M, Khosravi H. ABANet: Attention boundary‐aware network for image segmentation. Expert Systems. 2024;41(9).
- 23. Yu L, Min W, Wang S. Boundary-Aware Gradient Operator Network for Medical Image Segmentation. IEEE J Biomed Health Inform. 2024;28(8):4711–23. pmid:38776204
- 24. He Y, Yi Y, Zheng C, Kong J. BGF-Net: Boundary guided filter network for medical image segmentation. Comput Biol Med. 2024;171:108184. pmid:38417386
- 25. Xu R, Xu C, Li Z, Zheng T, Yu W, Yang C. Boundary guidance network for medical image segmentation. Sci Rep. 2024;14(1):17345. pmid:39069513
- 26. Wang H, Qi Y, Liu W, Guo K, Lv W, Liang Z. DPGNet: A Boundary-Aware Medical Image Segmentation Framework via Uncertainty Perception. IEEE J Biomed Health Inform. 2026;30(10):8302–14. pmid:40844947
- 27. Lei Y, Yan S, Zhang J, Li X, Wang P, Gao X, et al. BAG-Net: A Novel Architecture for Enhanced Medical Image Segmentation with Global Context Attention and Boundary Self-Attention. Symmetry. 2025;17(4):531.
- 28. Schlemper J, Oktay O, Schaap M, Heinrich M, Kainz B, Glocker B, et al. Attention gated networks: Learning to leverage salient regions in medical images. Med Image Anal. 2019;53:197–207. pmid:30802813
- 29. Chen L-C, Zhu Y, Papandreou G, Schroff F, Adam H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. Lecture Notes in Computer Science. Springer International Publishing. 2018. p. 833–51.
- 30.
Yang M, Yu K, Zhang C, Li Z, Yang K. DenseASPP for Semantic Segmentation in Street Scenes. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3684–92.
- 31.
Lin T-Y, Dollar P, Girshick R, He K, Hariharan B, Belongie S. Feature Pyramid Networks for Object Detection. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 936–44. https://doi.org/10.1109/cvpr.2017.106
- 32. Valanarasu JMJ, Sindagi VA, Hacihaliloglu I, Patel VM. KiU-Net: Overcomplete Convolutional Architectures for Biomedical Image and Volumetric Segmentation. IEEE Trans Med Imaging. 2022;41(4):965–76. pmid:34813472
- 33. Feng S, Zhao H, Shi F, Cheng X, Wang M, Ma Y, et al. CPFNet: Context Pyramid Fusion Network for Medical Image Segmentation. IEEE Trans Med Imaging. 2020;39(10):3008–18. pmid:32224453
- 34. Yuan Y, Cheng Y. Medical image segmentation with UNet-based multi-scale context fusion. Sci Rep. 2024;14(1):15687. pmid:39468067
- 35. Cui J, Wang L, Jiang S. A Multi-Scale Cross-Fusion Medical Image Segmentation Network Based on Dual-Attention Mechanism Transformer. Applied Sciences. 2023;13(19):10881.
- 36.
Valanarasu JMJ, Patel VM. UNeXt: MLP-based rapid medical image segmentation network. In: Medical Image Computing and Computer Assisted Intervention, 2022. 23–33.
- 37. Zhang S, Niu Y. LcmUNet: A Lightweight Network Combining CNN and MLP for Real-Time Medical Image Segmentation. Bioengineering (Basel). 2023;10(6):712. pmid:37370643
- 38. Pang B, Chen L, Tao Q, Wang E, Yu Y. GA-UNet: A Lightweight Ghost and Attention U-Net for Medical Image Segmentation. J Imaging Inform Med. 2024;37(4):1874–88. pmid:38478188
- 39.
Jiang J, Wang M, Tian H, Cheng L, Liu Y. LV-UNet: A Lightweight and Vanilla Model for Medical Image Segmentation. In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2024. 4240–6. https://doi.org/10.1109/bibm62325.2024.10822465
- 40. She Q, Sun S, Ma Y, Li R, Zhang Y. LUCF-Net: Lightweight U-Shaped Cascade Fusion Network for Medical Image Segmentation. IEEE J Biomed Health Inform. 2025;29(3):2088–99. pmid:40030349
- 41. Li Y, Hu B, Wang X, Liu K. Lightweight Unet with depthwise separable convolution for skin lesion segmentation. Sci Rep. 2025;15(1):34327. pmid:41038851
- 42. Shahid GES, Ahmad J, Warraich CAR, Ksibi A, Alsenan S, Arshad A, et al. LIU-NET: lightweight Inception U-Net for efficient brain tumor segmentation from multimodal 3D MRI images. PeerJ Comput Sci. 2025;11:e2787. pmid:40567779
- 43. Kavur AE, Gezer NS, Barış M, Aslan S, Conze P-H, Groza V, et al. CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation. Med Image Anal. 2021;69:101950. pmid:33421920
- 44.
Oktay O, Schlemper J, Le Folgoc L, Lee M, Heinrich M, Misawa K. Attention U-Net: Learning Where to Look for the Pancreas. In: 2018. https://doi.org/10.48550/arXiv.1804.03999
- 45.
Roy S, Koehler G, Ulrich C, Baumgartner M, Petersen J. MedNeXt: Transformer-Driven Scaling of ConvNets for Medical Image Segmentation. In: Medical Image Computing and Computer Assisted Intervention, 2023. 405–15.
- 46.
Ma J, Li F, Wang B. U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation. In: 2024. https://doi.org/arXiv:2401.04722
- 47.
Isensee F, Wald T, Ulrich C, Baumgartner M, Roy S, et al. nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation. In: Medical Image Computing and Computer Assisted Intervention - MICCAI 2024, 2024. 488–98.
- 48.
Hu J, Shen L, Sun G. Squeeze-and-Excitation Networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7132–41.
- 49. Woo S, Park J, Lee J-Y, Kweon IS. CBAM: Convolutional Block Attention Module. Lecture Notes in Computer Science. Springer International Publishing. 2018. p. 3–19.
- 50.
Wang Q, Wu B, Zhu P, Li P, Zuo W, Hu Q. ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 11534–42.
- 51.
Hou Q, Zhou D, Feng J. Coordinate Attention for Efficient Mobile Network Design. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 13713–22.
- 52. Bilic P, Christ P, Li HB, Vorontsov E, Ben-Cohen A, Kaissis G, et al. The Liver Tumor Segmentation Benchmark (LiTS). Med Image Anal. 2023;84:102680. pmid:36481607

관련 뉴스