A multi-dimensional lightweight attention-enhanced model for medical image segmentation

Abstract

Medical image segmentation is crucial in medical imaging analysis: based on grayscale, texture, and structural features, it precisely partitions an image into semantic subregions to enable accurate localization and analysis of lesions or anatomical structures. In recent years, deep learning methods—especially convolutional neural networks (CNNs)—have achieved remarkable progress in this field; however, fixed kernel sizes and limited receptive fields hinder adequate modeling of global dependencies. Although introducing multi-branch designs or global perception based on vision Transformers can partially alleviate this issue, they often increase computational complexity and are therefore limited in resource-constrained medical scenarios. To address this, we propose a multi-dimensional, lightweight, attention-enhanced medical image segmentation model. Our model integrates omni-dimensional dynamic convolution and mask attention: it performs adaptive modeling across spatial, channel, and kernel-number dimensions to significantly expand the effective receptive field, and employs binary masks to suppress background interference and focus on critical boundaries, with lower computational cost than global self-attention. Evaluated on three public benchmark datasets, our model achieves segmentation accuracy and inference efficiency that are superior to or comparable with current mainstream methods, demonstrating its potential applicability and practical value in clinical settings.

Introduction

Medical image segmentation is a fundamental task in medical imaging analysis, aiming to partition images into anatomically or pathologically meaningful regions to assist disease diagnosis, treatment planning, and postoperative assessment¹. However, this task continues to face numerous challenges, such as limited sample sizes, complex morphological characteristics of lesions, and interference from artifacts and noise introduced during image acquisition².

In recent years, rapid advances in deep learning have significantly propelled progress in medical image segmentation, with convolutional neural networks (CNNs) playing a dominant role³. The U-Net architecture proposed by Ronneberger et al.⁴, employing skip connections to merge low and high-resolution feature maps, has greatly improved contextual understanding and detail capture. Although U-Net and its enhanced versions (e.g., UNet++, U-Net3+, Attention U-Net, and nnU-Net) have demonstrated effectiveness in feature representation and multi-scale information capture^5–8, their fixed convolutional kernel sizes inherently limit the modeling of long-range dependencies. Consequently, these architectures struggle with multi-scale lesions, complex anatomical structures, and indistinct boundaries.

To overcome the limitations of static convolutions, Transformer architectures with self-attention have been introduced into medical image segmentation to enhance global context modeling⁹. The Vision Transformer (ViT) partitions images into patch sequences and performs self-attention across all patches, effectively capturing global dependencies but suffering from quadratic complexity with respect to sequence length, which limits its use in high-resolution medical imaging¹⁰. The Swin Transformer alleviates this issue via window-based and shifted window self-attention, reducing computational cost while improving multi-scale feature extraction, and Swin-UNet further integrates Swin blocks into a U-Net-like framework for multi-organ and cardiac segmentation^11,12. However, these models remain suboptimal in precise local feature modeling and still incur nontrivial computational overhead. TransUNet combines high-resolution CNN features with Transformer-based global dependency modeling to better balance localization and context, yet its complex multi-module design leads to high computational cost and constrains deployment on resource-limited devices¹³. Overall, although Transformer-based methods significantly improve global information modeling, they still face challenges in efficient multi-scale fusion and fine-grained local representation under realistic computational constraints.

In parallel, recent work has introduced the Kolmogorov–Arnold representation theorem into deep learning models¹⁴. By approximating high-dimensional mappings via compositions of low-dimensional operators, Kolmogorov–Arnold Networks (KANs) offer strong nonlinear expressiveness, interpretability, and parameter efficiency. UKAN, which embeds KAN modules into a U-Net encoder–decoder with skip connections, demonstrates effective feature reconstruction, stable convergence, and favorable performance in small-sample settings and is therefore adopted as the backbone in this study¹⁵. Nonetheless, two core limitations remain: (i) long-range dependencies in high-resolution images are still insufficiently modeled, limiting the capture of global semantics and distant structural relationships; and (ii) multi-scale representation and cross-scale fusion are relatively weak, as most KAN variants emphasize local enhancement, making them less robust for stable detection of small targets and fine-grained boundaries.

To address these limitations, this paper presents an efficient and lightweight medical image segmentation approach that performs joint adaptive modeling across multiple dimensions of the kernel space (kernel number, spatial scale, input/output channels), fundamentally alleviating the receptive-field insufficiency of fixed convolutions and the inadequate characterization of global relations. Meanwhile, a mask mechanism is introduced into standard attention computation, using adaptive binary masks to suppress background noise and highlight critical boundaries and fine structures, thereby enhancing cross-scale information fusion and boundary depiction without significantly increasing computational overhead. The overall design maintains model lightweightness while delivering effective performance gains, providing a more feasible segmentation solution for clinical applications in resource-constrained settings.

The main contributions of this paper are summarized as follows:

1. We propose a strategy that performs joint, adaptive adjustments across multiple dimensions—kernel number, spatial size, and input/output channels—to substantially enhance feature representation and expand the effective receptive field.

2. Using a data-driven binary mask to selectively activate features, the method suppresses non-target regions, highlights boundary details, and—without sacrificing accuracy—significantly reduces computational overhead compared with global self-attention.

3. Across multiple medical image segmentation datasets, the method achieves segmentation accuracy and inference efficiency that are superior to or comparable with current state-of-the-art approaches, validating its practicality and effectiveness.

The remainder of this paper is structured as follows: “Introduction” presents the research background. “Related work” discusses related work in medical image segmentation. “Methods” describes the proposed model in detail. “Experiments” analyzes experimental results across three datasets, summarizing the model’s strengths and weaknesses. “Discussion” concludes the paper.

Related work

With the rapid development of deep learning, convolutional neural networks (CNNs) have become mainstream methods for medical image segmentation, with U-Net and its variants as representative architectures^5–8. Building on this foundation, researchers have further extended segmentation models to more complex clinical scenarios. Methods such as PPAF-net¹⁶ and VIS-Net¹⁷ employ parallel branches, attention-based feature fusion, and multimodal feature modeling to address tasks like radiotherapy target delineation and robot-assisted surgical instrument segmentation. Other networks adopt multi-branch, multi-scale, or lightweight designs to enhance the segmentation of anatomically complex organs and lesions under limited computational budgets. Overall, although these U-Net–based CNNs, equipped with attention mechanisms and multi-branch/multi-fusion strategies, have achieved strong performance across various clinical tasks, they remain fundamentally constrained by static convolutions with fixed receptive fields, making it difficult to fully capture long-range dependencies and complex cross-scale interactions.

To overcome the above limitations, researchers have introduced dynamic convolution as an adaptive mechanism to capture feature variations more flexibly. Li et al.¹⁸ proposed MAGIC, which models the correlations among spatial, channel, and global dimensions via a dimension reciprocal fusion module, thereby adaptively adjusting convolution kernels and enhancing multi-scale representation capability, while a kernel recalculation module further improves feature aggregation efficiency, leading to strong performance in segmentation tasks. Wang et al.¹⁹ proposed LightAWNet, a lightweight adaptive weighting network that employs an inverted residual structure with spatial attention and a dual-branch design in the encoder to extract and fuse detailed and spatial features, and replaces transposed convolutions with lightweight upsampling while introducing channel attention in the decoder, thereby significantly improving segmentation accuracy while maintaining high inference efficiency. Overall, such dynamic convolution methods effectively increase the flexibility of feature capture and reduce computational cost.

To address the challenge of capturing long-range dependencies, Transformer architectures leveraging self-attention mechanisms for global feature modeling have been introduced in medical image segmentation. Representative methods like TransUNet and Swin-UNet have achieved superior performance across various datasets^12,13. In multimodal brain tumor segmentation, NestedFormer employs a modality-aware Transformer together with a modality-sensitive gating (MSG) mechanism to model intra- and inter-modality relationships among multimodal MRI sequences and to optimize cross-scale feature fusion²⁰. Zhao et al.²¹ proposed an uncertainty-aware multi-dimensional mutual learning framework, in which 2D, 2.5D, and 3D CNNs are jointly trained and an uncertainty-gated mechanism selects reliable soft labels for mutual supervision, thereby improving robustness and cross-dataset generalization. Built upon DINO-v3, SegDINO provides a unified and efficient segmentation framework applicable to both medical and natural images, offering a new direction for general-purpose cross-domain segmentation^22,23. However, these Transformer-based and multi-dimensional mutual learning approaches typically adopt complex architectures with high inference costs and rely heavily on large-scale annotated datasets and substantial computational resources, making deployment in resource-constrained clinical environments challenging.

To balance modeling capacity and efficiency, Gu et al.²⁴ proposed the Mamba architecture based on the Selective State Space Model (SSM), which achieves attention-like computation with approximately linear complexity, significantly reducing computational cost while strengthening long-range dependency modeling. U-Mamba combines the local modeling ability of CNNs with the long-sequence modeling strength of SSM and shows strong performance in medical image segmentation²⁵; SegMamba extends this paradigm to 3D volumetric data²⁶, and Serp-Mamba targets high-resolution retinal vessel segmentation, using SSM to maintain fine-grained vascular representation while preserving inference efficiency²⁷. Overall, these SSM/Mamba-based methods provide a promising technical route toward a more favorable trade-off between long-range modeling and computational efficiency.

In parallel, researchers have explored multilayer perceptron (MLP) and Kolmogorov–Arnold Network (KAN) architectures to further enhance long-range dependency modeling and global feature capture. U-MLP²⁸ and Rolling-Unet²⁹ aggregate features along multiple directions, offering lightweight design and efficient parallel computation, but their nonlinear representation capacity remains limited. KAN, proposed by Liu et al.¹⁴, overcomes the linear constraints of conventional MLPs by introducing learnable nonlinear operators based on the Kolmogorov–Arnold theorem, thereby strengthening complex feature modeling. KAN-based variants such as UKAN, TransUKAN, KM-UNet, and ResU-KAN demonstrate strong feature representation and segmentation accuracy^15,30–32. However, they still face practical challenges, including suboptimal long-range dependency modeling for high-resolution medical images, limited multi-scale information representation, and increased computational complexity from intricate module designs, which hinders deployment on resource-constrained devices.

In summary, existing methods are mainly limited in two respects: static convolutions struggle to effectively model long-range dependencies and multi-scale features, while global self-attention at high resolutions incurs prohibitive computational costs, making it difficult to balance accuracy and efficiency. To address this, we propose ODMA-UKAN as an enhanced framework that explicitly considers both performance and efficiency. In the encoder, we introduce an omni-dimensional dynamic convolution mechanism that jointly and adaptively adjusts kernel number, spatial scale, and input/output channels, thereby expanding the effective receptive field and strengthening cross-scale feature representation. Meanwhile, a data-driven binary mask is integrated into the attention computation to achieve sparse focusing and background suppression, highlighting fine structures and enhancing boundary delineation and multi-scale information aggregation. Under an overall lightweight network design, this synergistic combination of omni-dimensional dynamic convolution and mask-based attention effectively alleviates the limitations of conventional CNN/Transformer pipelines in global dependency modeling and fine-grained feature fusion, and constitutes the core innovation of this work.

Methods

While adopting UKAN as the baseline, we also note two key limitations: (i) insufficient modeling of long-range dependencies, which restricts the capture of global semantics and distant structural relationships in high-resolution and complex morphological scenarios; and (ii) weak multi-scale representation and cross-scale fusion, leading to unstable depiction of small targets and fine-grained boundaries. To overcome these shortcomings, we propose the improved model ODMA-UKAN, which introduces structural enhancements for global modeling and fine-grained feature characterization while maintaining parameter efficiency and small-sample friendliness. The Fig. 1 shows the overall architecture of ODMA-UKAN, which introduces several improvements to the encoder of the UKAN model. Initially, the encoder incorporates an omni-dimensional dynamic convolution (ODConv) module within the first three blocks³³. ODConv dynamically generates convolutional weights in spatial positions, input channels, output channels, and kernel dimensions based on input features . These dynamic weights multiply standard convolution kernels element-wise, adaptively adjusting their importance across each dimension to produce the output feature map. Subsequently, a lightweight mask attention (MaskAttn) module applies spatial masks³⁴, activated through a Sigmoid function, to the feature map . This mechanism efficiently focuses computational resources on key regions, significantly enhancing feature representation. Structurally, input images undergo processing by the ODConv and MaskAttn modules, after which the resulting features are reshaped into a sequence of length with channels and passed through two sequential Tokenized KAN blocks. The stacked Tok-KAN modules effectively capture global dependencies while preserving spatial details. After three such stacked convolutional units, the encoder reduces spatial resolution to , thoroughly extracting high-level semantic features. The decoder mirrors the encoder, integrating the deepest features globally via Tok-KAN blocks, followed by two sequential upsampling operations that progressively restore spatial resolution. Skip connections fuse shallow encoder features at each upsampling stage, with convolutional modules refining features, finally restoring the original resolution through a convolution to yield segmentation outputs. To mitigate feature loss during transmission, ODMA-UKAN employs skip connections between corresponding encoder-decoder layers, integrating fine-grained low-level details with high-level semantic features, thus enriching multi-scale information representation. The model accommodates varying computational demands by flexibly adjusting convolutional channel numbers via hyperparameters C₀, C₁, C₂, C₃ in the convolutional stages (ODConv, MaskAttn, Tok-KAN), while Tok-KAN and decoder convolutional modules’ channels are controlled by D₁ and D₂ respectively.

Fig. 1.

Overall workflow of the ODMA-UKAN model.

Dynamic convolution

Omni-dimensional dynamic convolution

A static convolutional layer employs a single, fixed convolutional kernel for all input samples. However, in medical images, lesions and tissues often exhibit diverse grayscale intensities, textures, and morphological characteristics, making it difficult for a single kernel to comprehensively capture these subtle variations. To address this limitation, dynamic convolutional layers learn an n-way linear combination of multiple kernels and use attention mechanisms to weight them, enabling the convolution operation to adaptively adjust across spatial, channel, and other dimensions based on the input features. In other words, the weights of dynamic convolution kernels are conditioned on the specific texture and structural information of the current image, thereby significantly enhancing the model’s ability to represent medical image features. Mathematically, dynamic convolution can be formulated as:

1

Here, and denote the input and output feature maps with height , width , and channels, respectively. represents the convolution kernel, which consists of filters, where each , , is the attention scalar used to weight , which is computed by an attention function based on the input features. The symbol denotes the convolution operation. For simplicity, the bias term is omitted here.

The proposed omni-dimensional dynamic convolution (ODConv) introduces a multi-dimensional attention mechanism, independently computing attention across kernel, input channel, output filter, and kernel spatial dimensions. This mechanism adopts a parallel strategy that considers every aspect of the convolution operation, enabling the model to capture salient features along different dimensions at each step of the computation. As a result, it significantly enhances the model’s ability to capture richer contextual information and deliver improved performance. According to the notation in Eq. (1), ODConv can be defined as:

2

Here, represents the attention scalar for the convolution kernel , which is consistent with the attention weight defined in Eq. (1). This component is primarily controlled by the number of grouped convolutions. When the number of kernels is large (i.e., when the number of output channels increases), employing group convolution can significantly reduce the computational cost and improve efficiency. Therefore, when the number of groups is greater than one, the convolution kernels within each group are assigned weights in the range of 0 to 1, with the sum of weights within each group equal to 1, achieved via the Softmax function. Here assigns different attention scalars to each spatial position within the convolution kernel for each filter, thereby enabling the kernel to emphasize or suppress different features within its receptive field. allocates individual attention scalars to each input channel of the convolution filter , allowing the kernel to selectively focus on more informative input channels. ∈independently assigns attention scalars to each of the convolution filters, allowing the model to adaptively adjust the importance of different feature detectors based on the input. These attention mechanisms are computed along the spatial, input channel, and output channel dimensions of each convolution kernel and denotes element-wise multiplication across different kernel dimensions. Specifically,, , , and are calculated using multi-head attention modules .

As shown in Fig. 1, the ODConv module details the dynamic weighting process across four attention dimensions within the convolution kernel. Specifically, ODConv first applies global average pooling (GAP) to the input feature , compressing it into a feature vector whose length equals the number of channels. This vector is then passed through a fully connected (FC) layer followed by ReLU activation, producing a lower-dimensional representation. Subsequently, the feature vector is fed into four independent branches, each consisting of an FC layer and a Softmax function, to generate normalized attention weights corresponding to spatial positions, input channels, output channels, and convolution kernels, respectively. These weights are sequentially multiplied and applied to the n convolution kernels, enabling dynamic adjustment across the four dimensions. In this manner, the kernel parameters can be adaptively adjusted according to different spatial positions, channels, filters, and kernels within the input, significantly enhancing the network’s ability to capture multi-dimensional contextual information. Finally, the weighted sum across all dimensions yields the output feature . Overall, ODConv effectively overcomes the limitations of traditional convolutions that rely solely on static weights, greatly enhancing feature representation and modeling capability.

Implementation details

To enhance the reproducibility of the proposed method, this section provides a brief description of the implementation procedure and key details of ODConv within the proposed model.

Let the input feature tensor of a given encoder layer be denoted by , with shape (), where is the batch size, is the number of input channels, and and are the spatial dimensions of the feature map. We first apply global average pooling to along the spatial dimensions to obtain a global descriptor vector with shape (). Then is passed through a shared dimensionality-reduction MLP and projected into an intermediate embedding space, yielding an intermediate representation with shape (). This vector is shared across all subsequent attention branches and serves as the common driver for adaptive modeling along different dimensions. On this basis, we construct four attention branches from the same intermediate representation, respectively acting on the number of kernels, the output channels, the input channels, and the spatial dimensions of the convolutional kernels. Concretely, the kernel branch produces an output of shape , generating a set of weights over candidate kernels for each sample. The output-channel branch produces an output of shape (, used to recalibrate each output channel. The input-channel branch produces an output of shape (), used to reweight the input channels. The spatial branch produces an output of shape (), which is reshaped in practice into a two-dimensional weight map aligned with the spatial size of the kernels.

A set of base convolution kernels is predefined with shape (). During the forward pass, the attention weights from the four branches are broadcast to these base kernels and multiplied element-wise, and then aggregated along the kernel dimension to obtain a customized dynamic kernel for each sample, with shape . The model then applies this dynamic kernel to convolve the corresponding feature map , producing an output feature map with shape . In our implementation, the ODConv module directly replaces the corresponding static convolution layer in the UKAN encoder, while the rest of the network architecture and training settings remain unchanged, ensuring fairness and comparability for ablation studies and future reproduction of our results.

Mask attention

In medical image segmentation tasks, lesion regions typically occupy only a small portion of the input image. Conventional multi-head self-attention (MHSA) mechanisms, which distribute attention weights uniformly across the entire feature map, often result in valuable attention resources being spread over a large amount of irrelevant background. This not only increases computational overhead but also diminishes the model’s ability to focus on critical regions. To effectively address this issue, we introduce a learnable mask matrix into the standard scaled dot-product attention. This mask matrix dynamically modulates the attention allocation between each pair of spatial positions in the input feature map(reshaped as , where is the batch size, denotes the spatial dimensions, and is the number of channels). By element-wise modulating the attention weights, explicitly enhances the model’s focus on potential lesion and other key regions, while effectively suppressing interference from anatomical boundaries, vascular artifacts, and other background noise.

3

Here,,, and represent the query, key, and value matrices, respectively, where , , and are learnable linear projection parameters for the query, key, and value. These three projection matrices map the input featuresinto different subspaces, facilitating the computation of spatial correlations among all positions in the attention module. denotes the total number of spatial positions, and is the dimension of the query and key vectors, used for scaling to prevent gradient explosion and to stabilize training. The introduced mask matrix can be either learned through backpropagation or dynamically generated based on features. By assigning large negative values to the attention scores associated with background regions, the network’s expressive capacity is concentrated on potential lesion areas. Compared with conventional self-attention without masking, Mask Attention not only explicitly filters out structured noise and improves the detection rate of small lesions, but also reduces attention dispersion, enabling the network to more rapidly focus on key regions. Moreover, sparse masks, whether generated offline or dynamically, allow for the pruning of most background attention computations.

After obtaining the attention-weighted output , we further enhance nonlinear modeling through a two-layer feed-forward network (FFN) with GELU activation, while preserving the original information via a residual connection:

4

Here, , and , are the weight matrices, and are the biases of the feed-forward network (FFN), which further refine and enrich the representation of . This produces the final output of the mask attention module, which not only integrates both global context and salient regional features, but also preserves the original information through residual connections. This facilitates more precise target prediction in subsequent segmentation layers.

It is important to note that, while both mask attention and omni-dimensional dynamic convolution introduce attention mechanisms into the model, these two types of attention are not in conflict. The spatial attention in omni-dimensional dynamic convolution is applied within the convolutional kernel space, where the attention computation scope is determined by the kernel size. In contrast, mask attention operates over the entire feature map. As a result, these two complementary attention mechanisms effectively model both local and global features, thereby further enhancing the model’s ability to capture more informative and valuable representations.

In this work, the mask matrix is learned by the network rather than defined by hand-crafted rules or heuristic design. Specifically, the current-layer feature mapis fed into a lightweight mask-generation subnetwork composed of several convolutional layers followed by nonlinear activations (e.g., ReLU or SiLU), producing mask logits of shape . These logits are passed through a sigmoid function to obtain a soft mask, which is then binarized using a fixed threshold (e.g., 0.5) to yield the mask used in attention computation. The binary mask is reshaped or downsampled according to the window size or sampling strategy so that its spatial resolution matches that of the attention weights, and positions corresponding to masked regions are suppressed by adding a large negative bias or directly setting their attention responses to zero. In this way, background activations are effectively attenuated while foreground and boundary structures are explicitly emphasized. The parameters of the mask branch are initialized together with the backbone (convolutional weights with He/Kaiming initialization and biases set to zero) and optimized end-to-end via standard backpropagation, so the mask matrix is fully learned in a data-driven and adaptive manner.

Tok-KAN block

To balance powerful nonlinear fitting capability with structural interpretability, ODMA-UKAN incorporates the Kolmogorov–Arnold Network (KAN) module—based on the Kolmogorov–Arnold representation theorem into the core pathway of both the encoder and decoder. This module has been shown to provide efficient and interpretable feature learning in medical imaging contexts. KAN utilizes learnable spline functions to enhance the representation of nonlinear features, thereby enabling more accurate approximation of nonlinear functions as a replacement for traditional multilayer perceptrons (MLPs). In conventional MLPs, nonlinear functions are approximated by stacking layers of linear mappings and activation functions:

5

Here, denotes the input feature vector, which carries the original spatial structure and contextual information and serves as the foundation for subsequent nonlinear transformations. r represents the linear transformation matrix, and is the nonlinear activation function. In CNNs, feature extraction is performed through multiple layers of convolutional operations, and both CNNs and MLPs employ stacked nonlinear activation functions to approximate complex nonlinear features. However, while these black-box operations are effective, they do not explicitly reveal the specific functionality of each internal layer, thereby reducing the interpretability of the model. Analogous to MLPs, the KAN formulation at the n layer is given by:

6

Each is composed of several B-spline functions ∈R, where:

7

Here, denotes the order B-spline basis function, and are the learnable coefficients. By appropriately choosing the sampling points such as using 5 sampling points, spline order of 3, and , KAN can achieve a favorable balance between flexibility and computational efficiency. Owing to the local support property of B-spline basis functions, KAN is able to capture key features with greater precision when modeling blurred boundaries, noise artifacts, and multi-scale lesion contours in ultrasound images. Moreover, piecewise spline mapping offers more targeted representational power than an MLP with the same number of parameters.

In the implementation of the Tok-KAN block, the feature sequence , flattened to a length of , is first normalized via LayerNorm, then passed through the aforementioned KAN layer, and finally, the original information is preserved via a residual connection:

8

The core advantage of KAN lies in the excellent local support and piecewise linear interpretability provided by B-spline basis functions. Each neuron’s output depends only on a small segment of its input, enabling fine-grained modeling of common features in ultrasound images, such as blurred boundaries, noise artifacts, and multi-scale structures. Compared to traditional MLPs with the same number of parameters, KAN offers more targeted fitting ability and greater capacity for feature decomposition with clearer physical interpretability. Furthermore, the Tok-KAN block, through its combination of normalization, spline-based mapping, and residual fusion, organically unifies global dependency modeling with the preservation of spatial details, thereby significantly improving both the accuracy and robustness of medical image segmentation tasks.

Experiments

Datasets

Datasets To comprehensively validate the feasibility and generalizability of the proposed model, we evaluated it on three medical image segmentation datasets with different modalities, structural characteristics, and clinical contexts: the Breast Ultrasound Image Dataset (BUSI)³⁵, the Gland Segmentation Dataset (GlaS)³⁶, and the Colon Polyp Segmentation Dataset (CVC-ClinicDB, hereafter CVC)³⁷. BUSI contains 647 breast ultrasound images from subjects aged 25–75 years, covering three lesion categories (normal tissue, benign masses, and malignant tumors). All images are resized to 256 × 256 pixels, making this dataset suitable for assessing performance on complex lesion boundaries in ultrasound. GlaS comprises 165 H&E-stained histopathological images of colorectal adenocarcinoma, mainly at pathological stages T3 or T4 and collected from multiple laboratories. Images are standardized to 512 × 512 pixels to evaluate the model’s ability to handle structural variability and blurred gland boundaries. CVC-ClinicDB includes 612 images extracted from 29 colonoscopy video sequences and is widely used as a benchmark for colon polyp detection and segmentation. Owing to large variations in polyp morphology, location, and size, CVC is well-suited for testing robustness and generalization in complex gastrointestinal scenes, and all images are uniformly resized to 256 × 256 pixels.

Experimental details

Experimental setup

All experiments were performed in a unified hardware environment using PyTorch on a single NVIDIA GeForce RTX 4070 GPU (16 GB). We trained the model for 400 epochs with the Adam optimizer, an initial learning rate of scheduled by cosine annealing, and a batch size of 8. The loss function is a weighted sum of a region-overlap loss and pixel-wise cross-entropy, with . To improve generalization and reduce overfitting, we applied random horizontal flipping, scaling, and mild translation/cropping, and all input images were intensity-normalized and resized to a fixed resolution. Each dataset was randomly split into 80% training and 20% validation, and all comparison methods used the same data partition and training settings. Every experiment was run multiple times, and we report results as “mean ± standard deviation” to ensure stability and reliability.

Loss function

To optimize segmentation performance, we adopt a hybrid loss that combines Binary Cross Entropy with Logits (BCE) and Dice loss, jointly accounting for pixel-wise classification and region overlap between prediction and ground truth. The BCE term, suitable for binary or single-channel segmentation, is defined as follows:

9

Here, denotes the ground truth label of the pixel, represents the model’s predicted output (prior to Sigmoid activation), and is the Sigmoid function. Given the prevalent class imbalance between foreground and background in medical images, we further incorporate the Dice loss to enhance the model’s sensitivity to small target regions. The Dice loss is defined as follows:

10

Here, is a smoothing term to prevent division by zero. To fully leverage the advantages of both loss functions, we employ the following weighted composite loss as the final optimization objective:

11

Here, is a weighting parameter used to balance the contributions of the two loss components, and is set to in our experiments. This composite strategy has been widely adopted in various medical image segmentation tasks and has proven effective in enhancing both the robustness of the model under complex backgrounds and the accuracy of region segmentation.

Evaluation metrics

To comprehensively and objectively assess the segmentation performance of the proposed model, we employ the following mainstream metrics: the Dice coefficient and Intersection over Union (IoU). The Dice coefficient is defined as:

12

Here, and denote the sets of pixels in the predicted and ground truth segmentation regions, respectively. The IoU is defined as:

13

This metric reflects the overlap ratio between the predicted and ground truth regions. In addition, to more accurately assess boundary quality, we adopt two boundary-related metrics: Hausdorff Distance (HD) and Average Surface Distance (ASD). HD and ASD measure the maximum and mean deviations between predicted and ground-truth contours, and are widely used for medical image segmentation involving irregular or fine structures. In this study, HD95 and ASD are computed in pixel units, where smaller values indicate more precise boundary localization. To evaluate computational cost, we also report floating-point operations (FLOPs) and the number of parameters (Params), which denote the total multiply–add operations for a single forward pass and the total trainable weights, respectively. Combined with the overlap- and boundary-based metrics, these complexity measures provide a comprehensive assessment of both segmentation accuracy and efficiency, offering a quantitative basis for judging the model’s feasibility and scalability in practical deployment.

Comparative experiments

In this section, we compare the performance of the proposed ODMA-UKAN model against several mainstream approaches across multiple benchmark medical image segmentation datasets. The results are presented in Tables 1, 2 and 3. The comparison includes classical CNN-based architectures such as U-Net⁴ and its variants, attention-based models (e.g., Att-U-Net⁷, efficient Transformer frameworks (e.g., U-Mamba²⁵, Serp-Mamba²⁷, and SegDINO²², multilayer perceptron (MLP)-based segmentation networks like U-NeXt³⁸ and Rolling-U-Net²⁹, and Kolmogorov-Arnold Network (KAN)-based models such as U-KAN¹⁵. Since KAN architectures are regarded as effective alternatives to MLP-based models, direct comparison with MLP-driven networks is particularly important. Moreover, U-KAN explicitly replaces the standard MLP blocks with KAN structures, making it a crucial baseline for evaluating our approach. To ensure fair comparison, all models are trained and evaluated under the same hardware environment and experimental settings.

Table 1.

Comparison of the proposed method and mainstream segmentation models on the BUSI dataset. The best results are highlighted in bold, and the second-best results are indicated by underlined italics. All values are reported as “mean ± standard deviation.”

Methods	Busi
	IOU↑	F1↑	HD95↓	ASD↓
U-Net4	57.22 ± 4.74	71.91 ± 3.54	49.74 ± 4.28	16.43 ± 4.14
Att-Unet7	55.18 ± 3.61	70.22 ± 2.88	50.59 ± 1.92	17.44 ± 3.21
U-Net++5	57.41 ± 4.77	72.11 ± 3.90	48.45 ± 2.26	12.54 ± 3.46
U-NeXt38	59.06 ± 1.03	73.08 ± 1.32	47.25 ± 1.33	15.02 ± 1.53
Roll-UNet29	61.00 ± 0.64	74.67 ± 1.24	40.46 ± 0.78	13.63 ± 1.25
U-Mamba25	61.81 ± 3.24	75.55 ± 3.01	37.33 ± 2.29	13.10 ± 3.31
Serp-mamba27	63.80 ± 0.27	75.35 ± 0.31	34.84 ± 0.27	12.66 ± 0.76
SegDINO22	64.24 ± 0.26	74.20 ± 0.33	39.13 ± 0.45	14.11 ± 0.61
U-KAN15	63.38 ± 2.83	76.40 ± 2.90	34.82 ± 1.87	11.25 ± 1.68
Ours	66.78 ± 0.72	79.80 ± 0.49	30.66 ± 0.53	10.92 ± 0.98

Table 2.

Comparison of the proposed method and mainstream segmentation models on the GlaS dataset. The best results are highlighted in bold, and the second-best results are indicated by underlined italics. All values are reported as “mean ± standard deviation.”

Methods	GlaS
	IOU↑	F1↑	HD95↓	ASD↓
U-Net4	86.66 ± 0.91	92.79 ± 0.56	35.40 ± 0.51	6.47 ± 1.24
Att-Unet7	86.84 ± 1.19	92.89 ± 0.65	35.15 ± 0.57	6.72 ± 1.02
U-Net++5	87.07 ± 0.76	92.96 ± 0.44	33.40 ± 0.53	6.53 ± 0.47
U-NeXt38	84.51 ± 0.37	91.55 ± 0.23	41.68 ± 0.45	8.20 ± 0.38
Roll-UNet29	86.42 ± 0.96	92.63 ± 0.62	36.02 ± 1.02	12.17 ± 0.69
U-Mamba25	87.01 ± 0.39	93.02 ± 0.24	25.17 ± 0.33	8.36 ± 0.41
Serp-mamba27	85.75 ± 1.25	92.10 ± 1.87	16.03 ± 1.24	4.27 ± 0.93
SegDINO22	88.04 ± 0.29	93.12 ± 0.75	18.26 ± 0.14	4.57 ± 0.36
U-KAN15	87.64 ± 0.32	93.37 ± 0.16	31.50 ± 0.43	5.83 ± 0.33
Ours	89.31 ± 0.12	94.31 ± 0.07	22.77 ± 0.23	5.03 ± 0.21

Table 3.

Comparison of the proposed method and mainstream segmentation models on the CVC dataset. The best results are highlighted in bold, and the second-best results are indicated by underlined italics. All values are reported as “mean ± standard deviation.”

Methods	CVC
	IOU↑	F1↑	HD95↓	ASD↓
U-Net4	83.79 ± 0.77	91.06 ± 0.47	17.20 ± 0.66	4.59 ± 0.53
Att-Unet7	84.52 ± 0.51	91.46 ± 0.25	16.69 ± 0.43	4.44 ± 0.32
U-Net++5	84.61 ± 1.47	91.53 ± 0.88	16.67 ± 0.76	3.88 ± 0.90
U-NeXt38	74.83 ± 0.24	85.36 ± 0.17	34.58 ± 0.44	8.93 ± 0.23
Roll-UNet29	82.87 ± 1.42	90.48 ± 0.83	17.63 ± 1.25	4.68 ± 0.79
U-Mamba25	84.79 ± 0.58	91.63 ± 0.39	15.78 ± 0.43	4.21 ± 0.34
Serp-mamba27	85.38 ± 1.22	92.12 ± 0.97	12.18 ± 0.76	3.34 ± 0.43
SegDINO22	86.53 ± 0.66	92.42 ± 0.72	9.34 ± 0.38	2.77 ± 0.61
U-KAN15	85.05 ± 0.53	91.88 ± 0.29	15.96 ± 0.25	4.29 ± 0.36
Ours	88.80 ± 0.49	94.04 ± 0.27	12.01 ± 0.45	3.25 ± 0.32

Experimental results on the BUSI, GlaS, and CVC datasets show that ODMA-UKAN consistently outperforms the baseline UKAN and most comparison methods in both segmentation accuracy and boundary quality, with especially notable gains over the strong baseline SegDINO. On BUSI, ODMA-UKAN improves IoU by about 3.39% and 2.54% over UKAN and SegDINO, respectively, and raises F1 by 3.37% and 5.90%, while reducing HD95 by 4.16 and 8.74 and ASD by 0.37 and 3.19. Compared with U-Net, its attention variants, Serp-Mamba, and SegDINO, it better captures multi-scale lesion morphology and maintains high contour consistency even for blurred boundaries. On GlaS, ODMA-UKAN improves IoU over UKAN and SegDINO by roughly 1.58% and 1.27%, and F1 by 0.89% and 1.19%, with HD95 substantially lower than UKAN (≈ 8.73) and slightly higher than SegDINO. It more accurately separates thin gland branches and densely clustered regions, benefiting from omni-dimensional dynamic convolution with joint spatial–channel adaptive weighting. On CVC, ODMA-UKAN increases IoU by about 3.50% vs. UKAN and 2.33% vs. SegDINO, and improves F1 by 2.02% and 1.62%; HD95 is reduced by 3.95 relative to UKAN but slightly higher than SegDINO, while ASD is 1.04 lower than UKAN and about 0.84 higher than SegDINO. Given the large variation in polyp shape, size, and illumination, conventional CNN and pure MLP models struggle to balance receptive field coverage and detail preservation, whereas ODMA-UKAN expands the effective receptive field via omni-dimensional dynamic convolution and leverages mask attention to enhance key polyp regions, yielding higher overall accuracy and better boundary quality, particularly for low-contrast, small polyps.

The comprehensive experimental results show that integrating omni-dimensional dynamic convolution (ODConv) and a mask attention mechanism into the encoder substantially improves medical image segmentation performance. ODConv adaptively weights spatial and channel dimensions, enhancing feature representation and expanding the receptive field to better model heterogeneous lesions and complex structures. The mask attention mechanism further directs the network toward semantically critical regions, boosting accuracy while limiting computational overhead. Together, these components allow ODMA-UKAN to achieve a favorable balance between lightweight design, representational power, and robustness, confirming its effectiveness and practical potential for medical image segmentation.

In addition, we systematically evaluate the efficiency of the proposed method from two perspectives—parameter count and computational cost (measured in GFLOPs)—across different datasets. GFLOPs denote the order of 10⁹ floating-point operations required for a single forward inference of the model; a lower value implies lower computational cost during inference and thus better suitability for deployment on resource-constrained devices. As shown in Table 4, we report the number of parameters (in millions, M) and the floating-point operations during inference (in GFLOPs) for each model on multiple datasets, where the best results are highlighted in bold and the second-best results are indicated by underlined italics. The results show that although SegDINO, U-NeXt, and U-KAN are more compact in terms of parameter count and have certain advantages in model compression, their segmentation accuracy is still clearly inferior to that of the proposed ODMA-UKAN. Considering both parameter complexity and segmentation performance, ODMA-UKAN attains superior segmentation results while maintaining a relatively low computational cost, demonstrating more prominent advantages in terms of practical applicability and overall performance.

Table 4.

Overall comparison with state-of-the-art segmentation models in terms of efficiency and accuracy metrics.

Methods	Average Seg.		Efficiency
	IOU↑	F1↑	Gflops↓	Params (M) ↓
U-Net4	75.89 ± 2.14	85.25 ± 1.52	524.2	34.53
Att-Unet7	75.51 ± 1.77	84.85 ± 1.26	533.1	34.9
U-Net++5	76.36 ± 2.33	85.53 ± 1.74	1109	36.6
U-NeXt38	72.80 ± 0.54	83.33 ± 0.57	4.58	1.47
Roll-UNet29	76.76 ± 1.01	85.92 ± 0.89	16.82	1.78
U-Mamba25	77.87 ± 1.47	86.73 ± 1.25	2087	86.3
Serp-mamba27	78.07 ± 0.91	86.52 ± 1.05	234.07	25.94
SegDINO22	79.60 ± 0.40	86.65 ± 0.60	4.81	23.83
U-KAN15	78.69 ± 1.27	87.22 ± 1.15	10.56	25.36
Ours	81.51 ± 0.40	89.31 ± 0.59	21.88	27.49

Figure 2 provides a qualitative visual comparison across all three datasets, intuitively illustrating the differences between the segmentation results of various models and the corresponding ground truth annotations. The example images are sampled from three publicly available medical image segmentation datasets: BUSI, CVC-ClinicDB, and GlaS. The BUSI dataset comprises ultrasound images of breast lesions; the CVC-ClinicDB dataset focuses on polyp segmentation in colonoscopy images; and the GlaS dataset contains histopathological images targeting glandular structure segmentation. In the BUSI dataset, the proposed ODMA-UKAN model demonstrates a remarkable ability to accurately extract lesion boundaries, showcasing superior contour-preserving performance. In the CVC dataset, where extensive background similarity poses a challenge, many baseline models can roughly localize the target area but tend to be misled by irrelevant regions. In contrast, our model effectively suppresses background noise and achieves more precise boundary delineation. The GlaS dataset presents a more complex challenge due to the intricate and diverse gland boundaries. Despite this, our model maintains excellent detail reconstruction, accurately separates adjacent structures, and minimizes missegmentation while preserving contour continuity. Overall, ODMA-UKAN consistently delivers superior performance across diverse imaging modalities and segmentation complexities. Its predictions demonstrate notable advantages over existing mainstream methods in terms of boundary accuracy, detail preservation, and robustness.

Fig. 2.

The comparison chart between the original image and the real label, and the segmentation results of different network models is displayed.

Ablation experiments

To thoroughly evaluate the effectiveness and robustness of the ODMA-UKAN model across diverse scenarios, we conducted multiple ablation experiments using three heterogeneous medical image segmentation datasets. These experiments involved several distinct configurations: employing only the omni-dimensional dynamic convolution (ODConv) module, solely integrating the mask attention mechanism, and combining both modules. The objective was to examine both individual and synergistic contributions of these modules to the overall performance enhancement. Initially, dynamic convolution was introduced into medical image segmentation to overcome the limitations of static convolutions. The core concept involves replacing single fixed convolutional kernels with adaptive linear combinations of multiple kernels, dynamically weighted based on input feature conditions, thus enhancing representational capability across different image regions. Representative works include CondConv proposed by Yang et al.³⁹ and DyConv by Chen et al.⁴⁰, both significantly increasing feature modeling flexibility while maintaining computational efficiency. However, these methods predominantly focus on dynamically adjusting the number of convolution kernels, neglecting dynamic adaptations across multiple dimensions such as spatial size, input channels, and output channels, resulting in suboptimal performance when handling complex lesions. Although Li et al.⁴¹ introduced DCD dynamic convolution to mitigate this limitation via dynamic channel fusion, their approach still lacks comprehensive multi-dimensional adaptability. Our ablation experiments systematically compared various dynamic convolution modules within the same network architecture, as presented in Table 5. Although DyConv demonstrated slightly better segmentation accuracy compared to our proposed ODConv and DCD dynamic convolution, it significantly increased parameter complexity, rendering it impractical for deployment in resource-constrained scenarios. Considering both performance and complexity, we selected ODConv as our dynamic convolution module, achieving superior segmentation accuracy with minimized parameter overhead.

Table 5.

Performance comparison of different dynamic convolution methods.

Methods	Busi		GlaS		CVC		Efficiency
	IOU	F1	IOU	F1	IOU	F1	Params(M)
DCD43	60.82	75.41	86.02	92.40	75.40	85.60	35.406
DyConv42	65.64	78.99	89.76	94.56	89.10	94.21	35.406
ODconv35	64.41	78.03	88.28	93.77	86.83	92.84	27.418

Next, we evaluated the individual effectiveness of ODConv by systematically replacing various encoding layers within the baseline U-KAN architecture with ODConv. The experimental outcomes across the three datasets are reported in Table 6. Results revealed that employing ODConv in four layers achieved optimal performance on the BUSI and GlaS datasets, whereas on the CVC dataset, ODConv performed best when integrated into five layers. This variance likely originates from dataset-specific characteristics: polyp boundaries in the CVC dataset are relatively clear and smooth, facilitating easier target delineation; conversely, BUSI and GlaS datasets contain more complex structures and ambiguous boundaries, necessitating more refined feature representations.

Table 6.

Performance comparison of omni-dimensional dynamic convolution with different numbers of layers.

Methods	Busi		GlaS		CVC
	IOU	F1	IOU	F1	IOU	F1
2 layer	64.94	78.46	87.84	93.52	82.48	90.13
3 layer	66.17	79.43	88.18	93.71	86.93	92.98
4 layer	66.32	79.27	88.31	93.78	86.73	92.82
5 layer	64.41	78.03	88.28	93.77	86.83	92.84

Subsequently, to assess the impact of different attention mechanisms on segmentation performance and efficiency, we replace the attention block in U-KAN with SCSA⁴², DCA⁴³, and mask attention (Mask-Attn), and report results on the BUSI, GlaS, and CVC datasets as summarized in Table 7. The experiments show that Mask-Attn is superior to or on par with the compared methods in overall performance and inference efficiency, and is particularly robust in low-contrast, noisy, and small-target—dominant scenarios; SCSA has advantages when fine spatial—channel recalibration is emphasized; and DCA offers potential benefits when tasks rely more heavily on long-range cross-scale dependencies. Considering computation and memory overhead, Mask-Attn achieves sparse focusing with binary masks, providing a better performance—efficiency trade-off and suiting resource-constrained or real-time applications, whereas SCSA and DCA can serve as complementary options under specific conditions.

Table 7.

Performance comparison of different attention mechanisms.

Methods	Busi		GlaS		CVC
	IOU	F1	IOU	F1	IOU	F1
SCSA44	63.46	77.39	88.64	94.05	85.18	91.94
DCA45	61.23	75.54	88.60	93.92	84.92	91.80
Mask-Attn36	64.76	78.33	87.71	93.45	85.90	92.36

Table 8 reports the ablation results under different values of the loss weighting coefficient. On the BUSI dataset, we selected five distinct settings to conduct a small-scale comparative study, aiming to evaluate how the trade-off between the region-overlap loss and the pixel-wise cross-entropy loss affects model performance, and to verify the rationality of using equal weighting. The results show that when , all segmentation metrics reach their optimum, indicating that an equal-weight combination of the two losses achieves a good balance between global region consistency and local pixel-wise discriminability. Therefore, adopting as the final equal-weight setting in this work is a reasonable and effective design choice.

Table 8.

Ablation results under different loss function configurations. The best results are highlighted in bold, and the second-best results are indicated by underlined italics.

	IOU	FI
0.2	65.95	77.14
0.4	63.48	75.05
0.5	66.78	79.80
0.6	65.61	77.02
0.8	65.13	76.09

Finally, we systematically evaluated the synergistic integration of omni-dimensional dynamic convolution and mask attention in ODMA-UKAN using multiple configurations with different convolutional depths (Table 9). Models combining both modules consistently outperformed those using either module alone, revealing clear complementarity. In terms of complexity, the mask-attention-only configuration yields the smallest number of parameters, while extending ODConv from four to five layers under the joint setting introduces only a slight parameter increase and remains lightweight. Notably, although four-layer ODConv performs best when used alone on the GlaS and BUSI datasets, the five-layer ODConv variant becomes optimal once combined with mask attention, indicating that attention reshapes the internal information flow and shifts the optimal structural configuration of dynamic convolution. Overall, these ablation results confirm the effectiveness of each module and highlight that global optimal performance relies on structural synergy and joint tuning within the integrated framework.

Table 9.

Results of individual and combined model usage.

Methods	Busi		GlaS		CVC		Efficiency
	IOU	F1	IOU	F1	IOU	F1	Params(M) ↓
+ODConv +Mask-Attn +ODConv + Mask-Attn	66.32	79.27	88.31	93.78	86.73	92.82	27.42
	64.75	78.33	87.71	93.45	85.90	92.36	25.43
	66.77	79.77	89.22	94.26	88.55	93.90	27.49

Discussion

Although experiments on the public BUSI, GlaS, and CVC datasets show that ODMA-UKAN has clear advantages in both segmentation performance and efficiency, the current evaluation is still confined to single-domain or mildly shifted scenarios. In real clinical practice, multi-center, multi-device, and multi-protocol acquisitions introduce substantial domain shifts in imaging parameters, populations, and annotation styles, often causing performance degradation in cross-center deployment. Enhancing the cross-domain generalization of ODMA-UKAN, especially when source-domain data are unavailable, is therefore an important direction for future work.

In this context, Source-Free Active Domain Adaptation (SFADA) is a promising solution: it assumes only a trained source model and relies on a small number of labeled and many unlabeled target samples, combined with active sampling and adaptive learning, to mitigate domain shift. SFADA has shown effectiveness in multi-center tumor and vessel segmentation and multi-rater medical image segmentation, and recent VLM-based variants further improve the balance between fairness and accuracy^44–47. Building on the strong single-domain backbone of ODMA-UKAN—enabled by omni-dimensional dynamic convolution and mask attention—we plan to integrate SFADA with uncertainty- and structure-aware active sampling, incorporate self-supervised learning on unlabeled multi-center images, and introduce boundary-aware consistency constraints to support robust deployment of ODMA-UKAN in realistic multi-center clinical environments.

Conclusion

We propose ODMA-UKAN, a multi-dimensional, lightweight, attention-enhanced model for medical image segmentation. The encoder integrates omni-dimensional dynamic convolution with a mask-based attention mechanism, effectively improving multi-scale feature representation and long-range dependency modeling. Experimental results on three public medical image segmentation benchmark datasets show that ODMA-UKAN achieves higher IoU and F1 scores than current state-of-the-art methods, validating its advantages in both accuracy and inference efficiency.

It should be noted that the generalization ability of ODMA-UKAN on highly heterogeneous data, as well as its capability to delineate very small targets and blurred boundaries, still has room for improvement. In particular, the fusion between local details and global context remains insufficient in cases involving complex lesion structures. Future work will therefore focus on enhancing boundary sensitivity and improving the quality of multi-scale feature fusion, with the goal of further strengthening the applicability and robustness of the model in more challenging clinical scenarios.

Author contributions

M.S—Methodology, software, validation, writing—original draft. P.K.K.—Resources, supervision, writing—review and editing.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

All datasets used in this study are publicly available and can be obtained from their official repositories or from the references cited in this paper. The implementation code for our method is available from the first author, Mei Shang, upon request.

Declarations

Competing interests

The authors declare no competing interests.